|J Pathol Inform 2018,
Pathology Informatics Summit 2018
|Date of Web Publication||31-Dec-2018|
Source of Support: None, Conflict of Interest: None
|How to cite this article:|
. Pathology Informatics Summit 2018. J Pathol Inform 2018;9:50
| Clearing Histology with MultiPhoton Microscopy: Primary Diagnosis of Prostate Biopsies in Un-Embedded Specimens|| |
Richard Torres1, Eben Olson1, Preston Sprenkle1, Darryl Martin1, Michael Levene2, Peter Humphrey1
1Laboratory Medicine, Yale School of Medicine, New Haven, 2Applikate Technologies, Weston, Connecticut, USA. E-mail: [email protected]
Content: Multiphoton microscopy (MPM) is capable of producing optical sections of resolution that can match or surpass that of wide-field microscopy with significant advantages over slide-based approaches. Avoiding the need for wax-embedding, physical cutting, and separate staining, simplifies workflow and reduces costs and labor. Digital images may become available in a shorter time frame. And the imaging is non-destructive and non-consuming, preserving tissue for ancillary studies. Optical sectioning also avoids challenges of slide imaging such as focus problems and cutting artifacts. Perceived drawbacks of MPM include long sample processing times, long data-acquisition times, high cost, and image degradation with depth. We have sought to systematically address these limitations for the practical implementation of MPM in clinical diagnostics, a technique we call Clearing Histology with MultiPhoton microscopy (CHiMP), and present preliminary data on its application to primary diagnosis of human prostate biopsies. Technology: Samples were prepared using previously described methods of tissue processing and clearing, employing the use of benzyl alcohol/benzyl benzoate (BABB) for refractive index matching. A custom multiphoton microscope based on a polygonal (spinning) mirror was designed and built for high speed and high resolution image capture at depth. Custom software tools were developed in python for image processing and visualization of multi-level image stacks. Whole slide imaging (WSI) of traditional slides was performed on an Aperio ScanScope. Design: Single prostate core biopsy specimens were obtained from consented individuals. Processing goals were making images available same day of biopsy. Specimens were subsequently submitted for traditional physical slides and scanning on a WSI system. Pathologists evaluated and compared image features of pseudo-colored CHiMP microscopy and WSI for primary diagnosis. Results: All samples were successfully processed and imaged with CHiMP on the same day of biopsy. Multiphoton sections demonstrated equal or superior overall quality with only minor differences in image features. Large file sizes affected visualization performance for CHiMP images, while focus issues were noted with some WSI samples. Conclusions: Fast image collection in a polygon-based CHiMP microscope produces timely optical sections that are amenable to primary diagnosis and present multiple advantages over traditional physical slide sections with WSI.
| Evaluation and Development of Standard Operation Protocol of Micro-Computed Tomography for Pathology Practice|| |
Kazuhiro Tabata1,2, Alexei Teplov1, Xiujun Fu1, Naohiro Uraoka1, Michael Roehrl1, Peter Ntiamoah1, Qing Chen3, John L. Humm3, Sahussapont J. Sirintrapun1, Melissa Murray1, Jinru Shia1, William D. Travis1, Meera Hameed1, Yukako Yagi1
Departments of 1Pathology and 3Medical Physics, Memorial Sloan Kettering Cancer Center, New York, USA, 2Department of Pathology, Nagasaki University Hospital, Nagasaki, Japan. E-mail: [email protected]
Content: Micro-CT is an emerging technology within the biomedical field and holds great promise for imaging pathology specimens for reconstruction, modeling, and analysis in 3D. Micro-CT’s capacity to create high resolution 3D models of ex vivo tissue therefore creates a unique opportunity to correlate micro-CT image data with other imaging modalities and bring additional information for diagnosis in future Pathology. In this study, one of the objectives is to develop of SOP of micro-CT specific to the tissue specimens, and another is to evaluate micro-CT image in tissue specimens in pathology. Technology: We visualized micro-CT images scanned by custom-built micro-CT scanner (Nikon Metrology, MI, US) with volume rendering and analyzing software, VG Studio MAX 2.2 (Volume Graphics GmbH, Heidelberg, DE). After visualization, we tried to do multimodality analysis using whole slide image digitized by whole slide imaging scanner, Aperio AT2 (Leica, IL, US). Design: All samples which are fresh tissue and FFPE blocks in our institute were CT scanned with variety of scanning condition to decide the SOP of micro-CT images. About fresh tissue, we investigated how to produce the high quality image with limited condition such as 5 minutes scanning for the purpose of using it as intraoperative frozen section. About SOP of FFPE block scanning, we aimed to be as first step to evaluate how details we can observe with micro-CT. As preliminary checking, radiation damage to tissue samples by micro-CT was also evaluated using immunohistochemistry for some antibodies. Results: We have established setting position on scanning and basic scanning parameter. The highest spatial resolution of scanning reached about 5um/pixel, and image quality of FFPE blocks was equal to microscope in low magnification to understand structure of tissue. Although evaluation of image quality and SOP were still under investigation, it was suggested that we must improve optimization with image processing including noise reduction, colorization, and analysis of Hounsfield units. So far we could not see any damages to tissue sample by the radiation of micro-CT using immunohistochemistry. Conclusions: Micro-CT imaging of tissue specimens can be of clinical utility in fresh and FFPE blocks. We have established SOP, but evaluation of image quality and SOP for FFPE blocks are still under investigation. We also found no damage by micro-CT in tissue samples, so future clinical application of micro-CT may be expected.
| Faster, Better, More Reliable than Deep Features: A Projection-Based, Pathologist-Centric Approach to Identification of Histopathology Images|| |
Hamid R. Tizhoosh1,2
1Huron Digital Pathology, St. Jacobs, ON, 2Engineering, KIMIA Lab, University of Waterloo, Waterloo, ON, Canada. E-mail: [email protected]
Content: This talk is about retrieval of histopathology images. A novel approach to represent tissue types will be described that uses Radon projections in local neighborhoods to generate barcodes for tagging images. Technology: The research on content-based image retrieval (CBIR) is about two decades old. CBIR deals with search in large archives of digital images when the query is an image. CBIR systems generate features or descriptors for each image to tag it. Subsequently, similar images can be retrieved either through direct comparisons or through classification. Design: We use a well-established technique, Radon transform, to develop a novel image descriptor. Given an image, we detect unique projections in local neighborhoods and quantify their variability to assemble a descriptor. Applying specific binarization methods can then produce barcodes to tag the image. This method allows us to create searchable image archives. Our solution enables the pathologist to select a region of interest on a new scan and run an archive-wide search for similar images. Our descriptor and its binary version, the barcode, are compact and hence can provide faster search results than deep features. Results: We used 24 scans and extracted approximately 30,000 patches of size 1000x1000 pixels. The images are publicly available under the label “Kimia Path24” (http://kimia.uwaterloo.ca). We performed experiments to validate the correct patch classification. We measured the patch-to-scan accuracy for our projection based approach and 5 different sets of deep solutions. The following [Table 1] shows the result. Conclusions: Digital pathology can benefit from intelligent image search and retrieval when it is designed around the pathologist’s needs. We demonstrate that handcrafted descriptors can be more expressive than deep features for digital pathology. The proposed projection-based descriptor and barcode are faster, more accurate and more reliable than deep features on the Kimia Path24 dataset. Besides, in contrast to deep networks, handcrafted descriptors cannot be fooled into wrong decisions through “adversarial attacks.”
| Enhanced-Depth Imaging Optical Coherence Tomography and Dynamic Focus for the Detection of Tumor-Like Features in Prostate Tissue Phantoms|| |
Mark D. Zarella1, Gautham Nandakumar2, Shantel Maharaj1, David E. Breen3, Fernando U. Garcia4
1Department of Pathology, College of Medicine, Drexel University, 2School of Biomedical Engineering, Drexel University, 3Computer Science, College of Computing and Informatics, Drexel University, Philadelphia, Pennsylvania, 4Department of Pathology, Cancer Treatment Centers of America, Boca Raton, Florida, USA. E-mail: [email protected]
Content: In vivo microscopy offers the potential to detect tumors at high resolution without the need for excision. However, the usefulness of this approach is inherently limited by the effective optical penetration of the technique. We examined the relationship between penetration depth and tumor visibility using a tissue phantom model designed to mimic prostate carcinoma. We present a novel method for the detection of deep structures and emphasize the benefits of image processing on enhanced depth imaging. Technology: We developed tissue phantoms with optical properties designed to mimic prostate tissue and which exhibited fine spatial structures similar to those observed histologically. We used enhanced-depth imaging optical coherence tomography (EDI-OCT), a recently developed method increasingly being used in the retina, to image prostate phantoms with and without image enhancement. By applying only minor modifications to a commercially available OCT system, we developed a novel dynamic focusing procedure that, when combined with B-spline interpolation, revealed deep structures previously unresolvable by conventional methods. Design: We independently modulated refractive index and transmittance in prostate phantoms with values representative of prostate carcinoma and benign tissue. Additionally, we created a configuration that emulated “ice ball” formation to demonstrate an application for intra-surgical monitoring of the freezing wavefront during cryoablation therapy. We acquired images using EDI-OCT and compared the impact of conventional image enhancement methods and dynamic focus-based methods. Results: We found that EDI-OCT enabled the visualization of structures and wavefronts at depths greater than 1 mm. We found that conventional image enhancement methods, including coregistration and wavelet denoising [Figure 1]a, provided complementary enhancements in signal quality and detection. Dynamic focus [Figure 1]b further improved the ability to resolve deep structures approaching 2 mm without sacrificing the quality of superficial layers [Figure 1]c, bottom vs. top], thereby achieving enhanced depth capabilities with fewer acquired frames. Conclusions: EDI-OCT achieves increased penetration depth in a model of prostate carcinoma. Image processing produces complementary methods of enhancement, potentially aiding in the interpretability of OCT images. These results demonstrate the synergistic role of improved optical methods and computational image processing, and expand the possible applications of in vivo microscopy to deeper tissues to support cancer diagnostics and surgical monitoring.
|Figure 1: (a) Wavelet denoising + averaging; (b) Dynamic focus; (c) Higher power view of regions marked in a (top) and b (bot)|
Click here to view
| Clinical and Histopathological Parameters of the Patients with Breast Cancer from North West Pakistani Population|| |
Asif Ali1, Mushtaq Ahmad2, Nabila Javeed3, Mah Muneer Khan4, Rauf Khattak3
1Pathology, Khyber Medical University, 2Department of Surgery, MTI Hayatabad Medical Complex, 3Department of Oncology, IRNUM Hospital, 4Department of Surgery, MTI Khyber Teaching Hospital, Peshawar, Pakistan. E-mail: [email protected]
Content: The aim of the study was to analyze the clinical and histopathologic parameters of patients with breast cancer from the NorthWest population of Pakistan. Technology: Pakistan is a resource limited country; therefore, we used a dropbox based database to document clinical and pathological data for the current project. Design: Demographic, clinical and histopathological data was extracted from patient files using proformas. The following parameters were assessed: age, family history, marital status, side and type of surgery, resection margins, tumor markers, foci, tumor grade, TNM stage and lymph node, vascular and lymphatic invasion. Data was analyzed for descriptive statistics. Logistic regression was performed by stratifying patients according to the disease stage as early stage (ES) (stage I and II) and late stage (LS) (stage III and IV) to get odds ratios (ORs) and P-values. Results: Clinical and histopathological data of 362 patients with breast cancer was profiled. From the available data 82 (33%) patients were early stage breast cancer, while 167 (67%) were late stage breast cancer. The mean age of patients in the ES breast cancer (45.8 years) was not statistically different from LS breast cancer (45.8 years) (p=0.99). ER+ cases were 62%, PR+ cases were 47% and HER2 positive cases were 49%. Lymph node invasion (p<0.0001), vascular invasion (p=0.05) and lymphatic invasion (0.009) were statistically significantly associated with LS disease. Lymph node invasion was predictive of LS breast cancer (OR=17.1, p<0.0001). In addition, lymphatic invasion was predictive of LS breast cancer (OR=3.2, p=0.01). Conclusion: The clinical and histopathologic patterns in ES and LS breast cancer are different which may require different management approaches. Majority of the patients present with late stage disease. Tumor markers positivity pattern differs from western population. Lymph node invasion is a better predictor of late stage disease.
| CD8(+) and PD-L1(+) Cell Densities in Biopsies May Predict Response to Durvalumab in Nonsmall-Cell-Lung-Cancer Patients|| |
Sonja Althammer1, T. H. Tan1, A. Spitzmüller1, L. Rognoni1, T. Wiestler1, T. Herz1, M. Widmaier1, M. Rebelatto2, S. A. Hammond2, M. C. Dieu-Nosjean3,4,5, K. Ranade2, G. Schmidt1, B. W. Higgs2, K. Steele2
1Definiens, Munich, Germany, 2MedImmune, Gaithersburg, MD, USA, 3Cordeliers Research Center, UMRS 1138, INSERM, 4Cordeliers Research Center, UMRS 1138, Paris Descartes University, Sorbonne Paris Cite, 5Cordeliers Research Center, UMRS 1138, Pierre and Marie Curie University, Sorbonne University, Paris, France. E-mail: [email protected]
Content: Predicting response to immunotherapies is an active area of cancer research. While the manually determined PD-L1 tumor cell status enriches for patients responding to anti-PD1/PD-L1 treatments, a more accurate predictive signature is needed. We previously reported that CD8 combined with PD-L1 provided an improved means of predicting response to anti-PD-L1 therapy in Non-Small-Cell-Lung-Cancer (NSCLC) patients than PD-L1 alone. Here, we extend this line of study by adjusting for potential confounders in the statistical analysis. Technology: Digital slides were fully automatically scored using the product of CD8(+) and PD-L1(+) cell densities with the Definiens’ Developer XD Software. Design: Baseline tumor biopsies were analyzed for PD-L1 (Ventana SP263) and CD8 (Ventana SP239) by immunohistochemistry. Samples were taken from patients with advanced NSCLC enrolled in Study NCT01693562, a nonrandomized phase 1/2 trial evaluating Durvalumab, an α-PD-L1 monoclonal antibody, (N=163, split into train and test set). In additional to these Durvalumab treated patients we also analyzed samples from NSCLC patients (T2+T3) who received chemotherapy (N=134). Multi-variate Cox analysis was deployed to adjust for potential confounders while comparing signature candidates. Results: The product of CD8(+) and PD-L1(+) cell densities provided greater overall survival stratification (p=0.0001 on test set) than high cell densities of CD8(+) or PD-L1(+), and of the manually determined PD-L1 tumor cell status for NSCLC patients who received α-PD-L1 [Figure 1]a. This signature was not prognostic for survival of patients treated with chemotherapy (p=0.6) [Figure 1]b. Conclusions: Automated digital scoring of tumor biopsy immunohistochemistry can be a useful part of the clinical management of NSCLC patients who might receive immunotherapy. Combining CD8 and PD-L1 data into a single immunohistochemistry signature provides greater predictive value than the prevailing manual PD-L1 assay alone. The lack of prognostic significance of this signature further supports its predictive utility.
|Figure 1: (a) patients receiving a-PD-L1 treatment; (b) patients treated with chemotherapy|
Click here to view
| Biological Insights into Spatial Distribution Patterns of CD8+ Cells–Distinct Cancer Phenotypes Defined by Image Analysis|| |
Stefan Bentink1, Andreas Spitzmüller1, Tze Heng Tan1, Lorenz Rogoni1, Ruben Cardenes1, Tobias Wiestler1, Hadassah Sade1, Brandon W. Higgs2, Keith E. Steele2, Sonja Althammer1
1Definiens AG, Munich, Germany, 2Medimmune, Gaithersburg, MD, USA. E-mail: [email protected]
Content: The overall density of CD8+ lymphocytes is important for characterizing the level of immune infiltrates in the tumor microenvironment. Beyond the densities (quantity) of CD8+ T cells, both location and spatial patterns of these cells in the tumor microenvironment may have relevance. Here we evaluated samples from 520 tumor resections encompassing eleven tumor types, and correlate location and spatial patterns of CD8+ T cells to pathway activity. Technology: We integrated image analysis results from digitized immunohistochemistry slides to derive spatial cell distribution statistics with gene expression data from a targeted Ion Torrent Panel. We deployed gene set enrichment analysis to identify activated pathways associated to the abundance or spatial distribution of CD8+ cells. Design: The spatial distribution and abundance of CD8+ T cells were assessed on digitized immuno-histochemistry slides from each tumor sample. Image analysis was used to compute overall density of CD8+ cells and the exact position of each individual CD8+ T cell on slides. A variant of Ripley’s K score was used as continuous spatial distribution metric to identify specimens with a dispersed CD8+ T cell distribution and to distinguish them from cases with a less uniform distribution of CD8+ T cells. Within a subset of sections from 33 bladder tumors, 43 head and neck carcinomas, and 43 lung adenocarcinomas the continuous dispersion metric was correlated with biological pathways using targeted mRNA sequencing and gene set enrichment analysis. Results: CD8+ T cell distribution patterns differed significantly between tumor types. Breast and pancreatic indications present a less uniform pattern, while lung tumors exhibited a dispersed distribution of CD8+ T cells. Transcriptional profiling data generated on a subset of three cancer indications revealed differences between both image analysis phenotypes. CD8+ density associated most with the prevalence of immune cells, while spatial distribution patterns suggested differences in motility, migration and activation status of tumor infiltrating T cells. Conclusions: We demonstrated the value of jointly analyzing transcriptional profiles and spatial image analysis data to identify biologically meaningful phenotypes. Our data suggests that high density and high dispersion of CD8+ T cells describe phenotypes with unique underlying biology.
| Multidimensional, Quantitative Assessment of PD-1/PD-L1 Expression in Patients with Merkel Cell Carcinoma and Association with Response to Pembrolizumab|| |
Nicolas A. Giraldo1, Peter Nguyen1, Elizabeth A. Engle1, Genevieve J. Kaunitz1, Tricia R. Cottrell1, Sneha Berry1, Benjamin Green1, Abha Soni1, Jonathan D. Cuda1, Julie E. Stein1, Joel C. Sunshine1, Farah Succaria1, Haiying Xu1, Aleksandra Ogurtsova1, Ludmila Danilova1, Candice D. Church1, Natalie J. Miller1, Steve Fling1, Lisa Lundgren1, Nirasha Ramchurren1, Jennifer H. Yearley1, Evan J. Lipson1, Mac Cheever1, Robert A. Anders1, Paul T. Nghiem1, Suzanne L. Topalian1, Janis, M. Taube1
1Pathology, Johns Hopkins Hospital, Baltimore, Maryland, USA. E-mail: [email protected]
Content: We recently reported a 56% objective response rate in patients with advanced Merkel cell carcinoma (MCC) receiving pembrolizumab (anti-PD-1 monoclonal antibody). However, a biomarker predicting clinical response was not identified. In this study, we determined the associations of the density and distribution of CD8+, PD-1+ and PD-L1+ cell populations in the MCC tumor microenvironment (TME) with anti-PD-1 response. Technology: Digital pathology. Automatic quantification of IHC/IF. Multispectral immunofluorescence. Design: Pretreatment FFPE tumor specimens (n=26, from eight academic institutions participating in the clinical trial) were stained for CD8, PD-L1, and PD-1 by immunohistochemistry/immunofluorescence (IHC/IF), and the density of positive cells was quantified with automatic digital pathology software. In addition, the digital images were z-stacked to analyze the geographic interaction between these immune cells and their correlation with response to anti-PD1 treatment. Finally, seven-color multispectral IF was used to test a separate cohort of MCC archival specimens (n=16), to identify cell types expressing PD-1. Results: Tumors from patients who responded to anti-PD-1 showed higher densities of PD-1+ and PD-L1+ cells when compared to non-responders (median cells/mm2, 70.7 vs. 6.7, p=0.03; and 855.4 vs. 245.0, p=0.02, respectively). There was no significant association of CD8+ cell density with clinical response. Quantification of PD-1+ cells located within 20 µm of a PD-L1+ cell showed that PD-1/PD-L1 proximity was associated with clinical response (p=0.03), but CD8/PD-L1 proximity was not. Interestingly, CD4+ and CD8+ cells in the TME expressed similar amounts of PD-1. Conclusions: While the binomial presence or absence of PD-L1 expression in the TME was not sufficient to predict response to anti-PD-1 in patients with MCC, we show that quantitative assessments of PD-1+ and PD-L1+ cell densities, as well as the geographic interactions between these two cell populations, correlate with clinical response. Cell types expressing PD-1 in the TME include CD8+ T-cells, CD4+ T-cells, Tregs, and CD20+ B-cells, supporting the notion that multiple cell types may potentiate tumor regression following PD-1 blockade. Our findings indicate that the next generation of IHC/IF biomarkers for immunotherapy will benefit from precise quantitation, spatial resolution, and studies of co-expression that will go beyond the simple qualitative interpretation of individual markers as positive or negative.
| Automated H-Score Analysis of Carcinoembryonic Antigen-Expressing Tumor Cells in Wholeslide Immunohistochemistry Brightfield Imaging|| |
Auranuch Lorsakul1, Emilia Andersson2, Oliver Grimm2
1Roche Tissue Diagnostics/Ventana Medical Systems, Inc., Imaging and Algorithms, Mountain View, CA, USA, 2Roche Innovation Center Munich, Roche Pharmaceutical Research and Early Development, Penzberg, Germany. E-mail: auranuch. [email protected]
Content: Carcinoembryonic antigen (CEA) is a common tumor marker currently evaluated in preclinical and clinical studies as a target for immunotherapy. We proposed to increase opportunities to develop a reliable companion diagnostic and overcome the limitations of semi-quantitative manual immunohistochemical scoring. This work presents: 1) a fully-quantitative automated image-analysis solution to report H-Score’s CEA-expressing tumors in wholeslide immunohistochemical brightfield imaging, and 2) the verification of quantitative cell-by-cell CEA expression in comparison against pathologists’ ground truth and gene-copy number from sequencing data. Technology: Forty-one tissue slides (2.5-μm FFPE) of human carcinoma tissues were included: 2×primary colorectal cancer, 16×metastatic colorectal cancer, 16×primary pancreas cancer, and 7×metastatic pancreas cancer. Slides were stained and detected for CEA [Clone: CEA31,VENTANA#7604594,Lot-independent] using ultraView DAB-detection systems, digitized on iScanHT scanner, and processed using automated image analysis: Stains were unmixed to separate DAB and hematoxylin staining presence in images. Computer-vision and supervised machine-learning methods detected and individually identified viable tumor cells with rejecting non-tumor cells and non-target CEA signals, e.g., necrosis. A total of 123 fields-of-view were included in algorithm development. Detected tumor cells were automatically classified based on their intensity of CEA-related-cell-adhesion-molecule5(CEACAM5) expression into: 1) CEACAM5+_high, 2) CEACAM5+_medium, 3) CEACAM5+_low, and 4) CEACAM5-negative. The different intensities’ discriminatory thresholds were determined by ROC studies on annotated ground-truth regions. H-scores from wholeslide readouts were automatically reported based on percentage of cells stained at different intensities, together with each tumor cell’s location, heatmaps, and intensity histograms. Results: On 41 fields-of-view, the algorithm’s cell counts correlated well with pathologists’ ground truth: R2=0.94, Lin’s concordance-correlation-coefficient=0.95 (95%CI [0.917;0.971]), with 16,353 cells counted by algorithm and 18,300 cells by ground truth [Figure 1]a. For gene-copy verification, two datasets of 16 tumor cell lines on tissue microarray and wholeslide images with known CEA-copy and mRNA numbers were included. The scatter-corrected DAB intensity measured by algorithm correlated to gene-copy numbers (R2=0.75,p<0.001) with intensity measurement’s saturation observed at copy numbers > 62,000 [Figure 1]b.For reproducibility results, Kolmogorov–Smirnov tests showed good intensity measurement reproducibility using 6-paired cell-line slides from different pellet blocks stained under the same conditions(p<0.005). Conclusions: Automated wholeslide image analysis with fully quantitative measures of CEA-expressing tumor cells is feasible and produces robust and reproducible results that correlate to copy numbers. This method offers support for finding relevant cut-offs in companion diagnostic developments.
|Figure 1: (a) The verification plot between the algorithm and pathologist counts, using 41 fields-of-view with R2 = 0.94; Lin’s concordance-correlation coefficient = 0.95 (95% confidence interval [0.917; 0.971]). A total of 16,353 cells counted by algorithm and 18,300 cells manually were included; and (b) The comparison of carcinoembryonic antigen intensities to the cell lines with known copy numbers for the gene-copy verification|
Click here to view
| Machine Learning in the Clinic: Dynamic Evolution of a Predictive Model for Somatic Variant Reporting|| |
Lev Lipkin1, Ryan Schmidt1, Enrique Dominguez Meneses1, Maciej Pacula1, Allison Macleay1, A. John Iafrate1, Jochen K. Lennerz1, Long P. Le1
1Department of Pathology, Center for Integrated Diagnostics, Massachusetts General Hospital, Boston, MA, USA. E-mail: [email protected]
Content: NGS-based tumor genotyping assays are rapidly becoming a standard tool in the personalized treatment of cancers. However, the interpretation of genotyping results remaining challenging. At the Center for Integrated Diagnostics at the Massachusetts General Hospital (MGH CID), our SNaPshot cancer genotyping assay typically detects about 2000 variants per case, with only about 0.1% ultimately being reported following interpretation by a molecular pathologist. The process is labor-intensive and additionally carries a risk of missing clinically relevant mutations. To address those challenges at MGH CID, we implemented a novel machine learning model which automatically scores variants for reporting. Here we describe our experience deploying the model into clinical practice, and its evolution into to a powerful interpretation tool applied to over 8 million genomic alterations a year. Technology: To address challenge of strong imbalance of reported vs. detected variants, we used meta-learning ensemble of two learners. We trained Random Forest and Logistic Regression classifiers on pathologist-annotated variants from over 2000 molecular cases as ground truth with initial filter deciding the learner to apply. We continuously evaluated the achievable accuracy via cross-validation to decide when to deploy. Design: The classifier works as a two-step ensemble [Figure 1], where the first learner is tuned for high specificity and the second learner for sensitivity for novel or otherwise promising variants. Our experience shows distinct performance characteristics of the two predictors since launch over time. Results: The model achieves an AUC of 98.8%. Since launch, most accuracy improvements came from the reduction of false positives [Figure 2]. The first classifier detects 97.6% of all positives, while the second rescues an additional 1.7%. The model identified several novel or unusual tumor mutations not reported before. The analysis of false negatives revealed borderline variants without consensus and reflected evolving understanding of cancer genetics. Conclusions: Our target statistics concentrated on helping pathologists with the signout process. We focused on minimizing false negatives, while keeping false positives manageable. Splitting variants into different cohorts for each predictor helped achieve accuracy goals despite extreme class imbalance.
| Prediction of Tumor Mutation Burden to Guide Immunotherapy in Lung Adenocarcinoma Using a 130 Gene Panel|| |
Rohan P. Joshi1, Soo-Ryum Yang1, Henning Stehr1
1Department of Pathology, Stanford School of Medicine, Palo Alto, CA, USA. E-mail: [email protected]
Content: Tumor mutation burden is a measure of the total number of tumor somatic mutations by whole exome sequencing and is a predictor of response to programmed cell death protein 1 immunotherapy. We investigated prediction of tumor mutation burden using our 130 gene targeted sequencing Stanford Actionable Mutation Panel. Technology: We created (1) a logistic regression model using the number of targeted sequencing mutations as a predictor and (2) an L2-penalized logistic regression model that additionally includes the identity of individual mutated genes in our panel. We used nested cross-validation for hyperparameter selection. We applied Cox regression to associate prediction probability scores with progression free survival in response to pembrolizumab. Design: We used previously published lung adenocarcinoma whole exome sequencing data from The Cancer Genome Atlas More Details (n=501), Broad Institute (n=159), and Memorial Sloan-Kettering (n=34) as training, test, and clinical outcome datasets, respectively. We applied a cutoff of 272 whole exome mutations, corresponding to a level that predicts response to nivolumab in a recent clinical trial, to define low and high tumor mutation burden. We filtered whole exome sequencing data for genes and positions present in our gene panel to create in silico targeted sequencing data. Results: In cross-validation, model 1 demonstrated a mean accuracy of 0.736 and area under curve of 0.815, while model 2 demonstrated statistically significant increases in mean accuracy of 0.780 and area under curve of 0.863. On the test set, model 1 and model 2 demonstrated areas under curve, sensitivities, and specificities of 0.843, 0.671, and 0.838 and 0.897, 0.823, and 0.825, respectively. Raw probability scores from both models statistically correlated with progression free survival in response to pembrolizumab treatment, with concordance indices of 0.682 and 0.707 for model 1 and 2, respectively. Conclusion: We developed a model that predicts clinically actionable tumor mutation burden levels using only 130 genes (about 0.2 megabases). The predictive ability of a model including the identity of individual mutated genes was superior to that of a model using the number of panel mutations alone. Future directions include validation of this model using paired whole exome and targeted sequencing.
| Computerized Histomorphometric Features Relating to Nuclear Shape and Architecture Correlate with EGFR and KRAS Mutations in Early-Stage Nonsmall Cell Lung Cancer|| |
Priya D. Velu1, Kaustav Bera2, Cheng Lu2, Jennifer J. D. Morrissette1, David B. Roth1, Michael D. Feldman1, Anant Madabhushi2
1Pathology and Laboratory Medicine, Hospital of the University of Pennsylvania, Philadelphia, Pennsylvania, 2Biomedical Engineering, Case Western Reserve University, Cleveland, Ohio, USA. E-mail: [email protected]
Content: Currently, lung tumor samples are not screened for genetic testing as pathologists cannot determine mutational status by visual inspection of tumor alone. Screening at histologic diagnosis may facilitate rush genetic analysis of cases with mutations that can be treated with targeted therapies. Here we use computational approaches to explore whether sub-visual features of nuclear morphology from digitized lung cancer pathology images are correlated with pathogenic mutations in EGFR and KRAS that are frequently found in non-small cell lung cancer. Technology: Hematoxylin and eosin-stained histologic guide slides for samples submitted for targeted next-generation DNA sequencing on the MiSeq platform (Illumina, San Diego, California, USA) were scanned using Intellisite Digital Pathology Solution (Philips, Amsterdam, Netherlands). Custom analysis scripts were written in Matlab R2017a (Mathworks, Natick, Massachusetts, USA). Design: Cases that were determined by next-generation sequencing to be EGFR+/KRAS-(n=50), EGFR-/KRAS+ (n=50), and EGFR-/KRAS-(n=50) were used for three classifications. Median allele frequencies of EGFR and KRAS mutations were 21% (range: 2-80%) and 19% (range: 2-98%) respectively. Tumor areas were annotated by a pathologist and tumor nuclei were segmented using a watershed approach. 1800 morphologic, local graph, cell graph, cluster graph, local co-occurrence of cell morphology, and cell run-length features were extracted. 100 iterations of four-fold cross validation were performed using a random forest classifier that used the top three features as determined through minimal redundancy maximal relevance feature selection. Results: Features of tumor nuclei shape and architecture are associated with EGFR and KRAS mutations. Classification yielded area under the receiver operator curve of 0.68 for EGFR+/KRAS-vs. EGFR-/KRAS+ using global nuclei arrangement features; 0.60 for EGFR+/KRAS-vs. EGFR-/KRAS-using disorder of edge length between nuclei and average size and number of nuclei in local nuclei arrangements; and 0.57 for EGFR-/KRAS+ vs. EGFR-/KRAS-using local cellular diversity in terms of nuclear shape and intensity of region immediately surrounding nuclei.Conclusion: Histomorphometric features of tumor nuclei shape and architecture were found to be associated with EGFR and KRAS mutations in non-small cell lung cancer. Further validation needs to be performed using multi-site data. To improve model performance, ongoing work includes additional feature analysis to find more robust signals.
| A Novel Network and Gene Expression Pipeline Identifies Important Signaling Subnetworks in Eml4-Alk Translocated Lung Cancer|| |
Drew F. K. Williamson1, Andriy Marusyk2, Jacob G. Scott1
1Department of Translational Hematology and Oncology Research, Cleveland Clinic Foundation, Cleveland, OH, 2Department of Cancer Imaging and Metabolism, Moffitt Cancer Center, Tampa, FL, USA. E-mail: [email protected]
Content: The evolution of resistance represents a critical problem for cancer patients. Pathologists require new informatics techniques to more effectively manage disease progression, especially given the gap of knowledge about non-genetic mechanisms of resistance. Our novel computational pipeline to analyze gene expression data in the context of protein-interaction networks illuminates the signaling systems that underlie those resistant phenotypes, aiding design of second-line therapeutic strategies. Technology: Our bespoke pipeline computes the Gibbs free energy of each node in a protein interaction network with superimposed RNA-seq data, combining information about the connections of each node with the expression levels of those nodes to which it is connected. Gibbs free energy is high when all nodes have equal expression levels and is low when nodes are over-or underexpressed compared to those around them. To reduce noise, we prune the network to contain only those nodes with the most negative Gibbs free energy. Using a metric known as Betti number, we quantify how important a node is to the resiliency of the signaling network. Nodes with high Betti numbers are important to the phenotype in two orthogonal ways and constitute key vulnerabilities. Design: We analyzed normalized gene expression data of EML4-ALK translocated cell lines that have evolved resistance to four targeted therapies: alectinib, ceritinib, crizotinib, and lorlatinib and compared to a DMSO-treated control line. Our threshold of filtration was four standard deviations above the mean Gibbs free energy. Results: [Figure 1] displays the final networks with node coloring correspond to increasing Betti number. Note that in three cell lines, the resultant networks contain EGFR after pruning and in each it is the single most important node by Betti number. This agrees well with experimental results that these cancers often use EGFR as an escape mechanism. Additionally, the final networks contain a variety of novel therapeutic targets important to the cells’ phenotypes. Conclusion: This pipeline demonstrates how cancer cells have evolved resistance to targeted therapies and exposes their weak points for second-line attacks. Using techniques such as this will allow pathologists to better provide clinicians with evolutionarily-rational targets when patients relapse.
|Figure 1: Final networks after filatration for each ALK tyrosine kinase inhibitor. The greater the color intensity of the node, the more important the node is to the structure of the network and, therefore, the better target that node would be for therapy|
Click here to view
| Meta-Analysis Emphasizes Role of IGF-1 Signaling in the Pathogenesis of Psoriatic Arthritis|| |
Laraib Z. Safeer1, Jihad Aljabban2, Maryam Panahiazar3, Isaac Neuhaus4, Dexter Hadley3
1Baylor College of Medicine, Houston, Texas, 2Ohio State University, College of Medicine, Columbus, Ohio, 3Institute for Computational Health Sciences, University of California, 4Department of Dermatology, University of California, San Francisco, California, USA. E-mail: [email protected]
Content: Employ a crowd-sourced meta-analysis platform to define the pathogenesis of Psoriatic arthritis. Technology: The National Center for Biotechnology Information Gene Expression Omnibus (GEO) is an open database of more than 2 million samples of functional genomics experiments. Our Search Tag Analyze Resource for GEO (STARGEO) platform allows for meta-analysis of genomic signatures of disease and tissue through tagging of individual samples across different studies. Design: We analyzed 87 blood samples from psoriasis patients against healthy controls using STARGEO. The signature was then analyzed in Ingenuity Pathway Analysis to identify candidate biomarkers and disease processes within the context of biological conditions. Results: Blood analysis showed IGF-1 signaling and prolactin signaling as top canonical pathways. Top upstream regulators include COL18A1 [Figure 1] and fibronectin 1. We noted upregulation of angiopoietin 1 and aromatic hydrocarbon receptor, and noted decreased biosynthesis of histamine. Conclusion: Our analysis emphasized the role of IGF-1 signaling in the pathogenesis of psoriatic arthritis. IGF-1 signaling is known to be involved in rheumatoid arthritis and promotes survival of neutrophils and T lymphocytes. Prolactin signaling has also been described in the synovial tissue of both psoriatic and rheumatoid arthritis, and plays a role in macrophage activation.Studies have shown increased angiogenesis in psoriasis vulgaris, and upregulation of angiopoietin 1 in our study is consistent with increased vascular development and angiogenesis. Although systemic angiogenesis is upregulated in psoriasis, studies have shown increased expression of the antiangiogenic protein COL18A1 in the synovium of psoriasis patients. This may indicate limited angiogenesis in psoriatic arthritis.Plasma levels of fibronectin 1 indicate connective-tissue repair and are a marker of psoriatic arthritis. Polymorphisms of the aromatic hydrocarbon receptor and its repressor gene are associated with susceptibility to rheumatoid arthritis. Accumulating evidence from recent studies has shown that these play crucial roles in several immune diseases and may potentially play a role in psoriatic arthritis. Plasmacytoid dendritic cells in psoriatic lesions highly express the histamine H4 receptor and downregulate immune response in the skin. Interestingly, blood analysis demonstrated a decreased biosynthesis of histamine. Further study may clarify if this finding is associated with disease severity.
| Toward the Automated Scoring of Fluorescence in situ Hybridization Using a Confocal Whole Slide Image Scanner|| |
Naohiro Uraoka1, Xiujun Fu1, Paul Matises1, Lu Wang1, Mamta Rao1, Yanming Zhang1, Meera Hameed1, Yukako Yagi1
1Department of Pathology, Memorial Sloan Kettering Cancer Center, New York, USA. E-mail: [email protected]
Content: The standard, manual scoring for fluorescence in situ hybridization (FISH) is labor-intensive and time-consuming. Confocal imaging which eliminates out-of-focus noise offers higher resolution and more spatial information than conventional widefield imaging. The purpose of this study was to establish the automated FISH scoring method using a confocal whole slide image (WSI) scanner and an image analysis software commercially available. Technology: FISH slides were digitized by a Pannoramic Confocal WSI scanner (3DHistech Ltd., Budapest, Hungary) with a 40× water immersion objective (0.1625 micrometer/pixel). Imaris (Bitplane, Zurich, Switzerland) was used for image analysis including nuclear segmentation, spot quantification and detection of spot co-localization inside the segmented nuclei. Design: Three archival break-apart FISH slides (two negative slides and one positive slide clinically) were prepared for the current study. Several regions of interest (ROIs) were selected for confocal scanning according to the corresponding H&E slides. We scanned seven layers at 0.6 μm intervals based on the optimization results of the previous study. The images were viewed in CaseViewer (3DHistech), and an adequate number of ROIs were defined so that at least 200 nuclei with interpretable signals were obtained by the subsequent analysis in Imaris. For the nuclear segmentation in Imaris, only the sphericity parameter was manually decided to exclude overlapping nuclei, while the other parameters were decided automatically. Spot detection was done by fully automated algorithms. The accuracy of nuclear segmentation and spot detection was assessed by a pathologist. Results: Confocal scanning of FISH slides provided sharp images with spatial information of spot signals. By our semi-automated algorithms in Imaris, nuclear segmentation and spot detection were successfully performed. The number of each signal patterns were correlated with the clinical result [Figure 1]. Conclusions: We established the semi-automated method of FISH scoring combined with confocal scanning. This method was helpful to obtain accurate, plentiful information of FISH more efficiently than conventional manual methods. The next step is to enroll more clinical cases for evaluation and validation. Furthermore, according to the protocol and the data of this study, we are currently developing an in-house software for the fully automated FISH scoring system also considering deep learning.
| Adult and Pediatric Autopsy Web-Based Templating App|| |
Rufei Lu1, Ngoc Tran1, Christopher William1
1Department of Pathology, University of Oklahoma Health Science Center, Oklahoma City, Oklahoma, USA. E-mail: [email protected]
Content: Generating a standardized preliminary autopsy report in a timely manner often provides an immense amount of comfort for the deceased’s family and physicians. To streamline the workflow, improve efficiency and reduce turnaround time, we created a web-based application (app) that will generate a complete preliminary macroscopic examination report upon the completion of autopsy procedure for both pediatric and adult populations. Technology: Web-based application with the use of HTML 5.0, enhanced real-time interactive features with Knockout JS (v3.5.2), Bootstrap (v4.0.0), and jQuery (v3.2.1). Design: The app encompasses both adult and pediatric reporting templates. The dynamic drop-down tab on the left 40% of screen allows navigation through different portions of the report with ease. A real-time report is visible and editable on the right. Age and gender specific reference ranges for organ measurement were based on the Autopsy Pathology: A Manual and Atlas, 3rd edition by Conolly et al. App can generate a simple diagnosis by comparing the patient’s values with reference. For example, cardiomegaly will be auto-filled when an increased heart weight is put in, accomplished by comparing to reference range based on patient’s gender and age. Results: Utilizing the new dynamic feature of Bootstrap, we are also able to create a web-based, device-responsive, standardized, and user-friendly app that can be accessed via intranet on any devices [Figure 1]. This app structures the input in organ based fashion, however, dynamic does also customize sequence of input. This streamlining of input is of great educational purpose for pathology residents, who are not yet familiar with autopsy by providing a systematic approach to dissection and examination. Additionally, the app has a copy-paste and save as a word documentation features, which also help user to save the working process to prevent data loss and case logging purposes. Conclusion: The implementation of this app not only will streamline a standardized workflow with a universal format, but also aid the education of pathology resident and familiarize them with a systematic approach to autopsy. By automatically generating the same data to the preliminary report, the automatic reference filling features helps minimize clerical error from manual repetition.
|Figure 1: Screenshot of the layout of the web-based app on autopsy templates|
Click here to view
| Implementation of Electronic Order Entry for Anatomic Pathology|| |
1Department of Pathology, Anatomy and Cell Biology Jefferson University Hospital, Philadelphia, Pennsylvania, USA. E-mail: [email protected]
Content: Historically, electronic order entry has been primarily used for clinical laboratory testing. Here, we examine our experience with the implementation of electronic order entry for anatomic pathology. Technology: Epic (Epic, Verona, Wisconsin, United States), an integrated electronic medical record system. Beaker, the laboratory module for anatomic and clinical pathology. OpTime, the operating room management system. Design: Prior to Epic, the workflow involved the use of hand-written pathology orders on paper requisition forms and specimen labels. Specimens were delivered to the lab and accessioned. Orders were manually transcribed and cases were created. With the implementation of Epic, this manual process was changed to a paperless workflow in which pathology orders are placed electronically. To accomplish this, joint meetings were held between the pathology and clinical staff to define workflow requirements and reach a consensus regarding how specimens would be ordered. In the current process, once a specimen order is created, a barcoded label is generated for each specimen. Specimens are then delivered to the lab where electronic orders are automatically linked to the pathology case upon scanning the specimen label. Results: Numerous technical and operational challenges were encountered. A significant operational issue has been compliance of clinical staff to follow the proper workflows. Another significant challenge is related to the integration of OpTime and Beaker, where specimen designations are not synced between the two modules. Specimens that are sequentially numbered in OpTime may not necessarily follow the same sequence Beaker. To address these problems, numerous work-arounds have had to be developed. Benefits to implementing electronic orders: decreased accessioning time, reduced handwriting and miscommunication errors, and the ability to audit a case from order creation to the final report. Conclusions: The implementation of electronic ordering for anatomic pathology presents numerous challenges. The complexities associated with the ordering of pathology tests can difficult for clinical staff. Many of these issues originate from upstream workflows outside of the lab’s control. Successful implementation requires close collaboration with all stakeholders both within and outside of the laboratory and adherence to established workflows. When these are not followed, there are numerous downstream effects which may negatively impact the laboratory.
| Application of Modern Full-Text Search Technology and Information Visualization for Improving Decision Support in Pathology|| |
Wade Schulz1, Peter Gershkovich2
Departments of 1Laboratory Medicine and 2Pathology, Yale School of Medicine, New Haven, CT, USA. E-mail: [email protected]
Content: Laboratory information systems can improve the quality of pathological diagnoses by making context sensitive pertinent information available to the pathologist at the right time. Modern full-text search methods and service-based integration technologies can provide real-time, dynamic access to pertinent information. Augmented with dynamic visualization methods, these technologies can provide rapid access to information, but more importantly, can also provide context and relevance without the need to sift through superfluous text. Technology: Pathology reports where extracted from CoPathPlus Laboratory Information System (Cerner, Kansas City, MO) using our existing Repetitive Tasks Scheduling service – a Java based custom software. The texts of pathology reports were parsed and then indexed using Elasticsearch (Elastic, Mountain View, CA). These data were then interfaced with a web-based application to provide a contextual search engine to pathologists and trainees. Design: Pathology Portal is a web-based application developed at Yale that aggregates and displays information from a number of disparate systems that include a range of LIS systems used by different departments, the diagnostic imaging system, and the electronic health record. To provide contextual search, a full-text search engine with additional visualization tools was added to the Pathology Portal. Visualization components were built using d3.js framework. Results: A full-text search module in Pathology Portal improved the ability of pathologists and residents to review prior reports and other historical data. It also helped to identify and review similar cases, compare cases on a timescale, and assess the relevance of information through information visualization. Conclusion: Pathology Portal augmented with full-text indexing and contextual relevance scores helps improve pathologists’ access to pertinent information. Full-text search can be used to rapidly find relevant information without time-consuming review of documents in their entirety. By highlighting matching and relevant text, the interface can provide quick access to desired information with the ability to review additional details or where necessary.
| Virtual Slide Labels as a Solution for Image Identification Challenges in Digital Pathology|| |
Nicholas C. Jones1, Markus D. Herrmann2, Jochen K. Lennerz1, Veronica Klepeis1
1Department of Pathology, Harvard Medical School, Massachusetts General Hospital, Boston, Massachusetts, USA, 2MGH and BWH Center for Clinical Data Science. E-mail: [email protected]
Content: Organized, structured, and standardized identification of anatomic pathology materials is necessary for efficient, safe, and interoperable digital pathology processes within and between clinical, research, and educational contexts, while protecting patient privacy. Implementation of digital pathology highlights the need for use of images in different contexts and reassessment of traditional pathology workflows. Challenges and suggested resolutions are described. Technology: Anatomic Pathology Laboratory Information Systems, pathology imaging data structures, viewers, and systems, slide label printers, slide file systems, static and dynamic imaging modalities, and workflows for asset production in pathology practice. Design: Analysis of current practice and technology for asset identification within anatomic pathology highlight several technical and practical problems which prevent and/or hamper use and re-use of pathology images. Results: Challenges in the de-identification of clinically identified whole slide images for use in educational or research contexts, limits the use of scans and/or mandates costly manual relabeling and rescanning of slides. Relabeling and a concept called “virtual labels” are proposed for de-identification, multi-identification, pseudonymization, and re-identification. Virtual labels, the coupling of an alternate slide label image to a whole slide image, would allow for software-level creation of supplemental and/or alternative slide label images for education, research, and clinical consultation contexts. If adopted and standardized, these could allow for reuse of tiled image data in multiple contexts without need for rescanning or risk of eliminating original identifiers from clinical records. These are in line with precedents set in radiology through DICOM supplements 55 and 142. For other static and dynamic images in anatomic pathology, standardized and structured identification practices would allow for improved practice and interoperability with burgeoning digital pathology systems. Conclusions: Adoption and incorporation of virtual labels within digital pathology along with improvements to identification processes in other areas of pathology imaging would increase the value of whole slide images and decrease costs associated with reuse while protecting patient confidentiality. Challenges that remain include software development and process redesign to handle these new features.
| Deep Learning for Breast Tumor Segmentation at Pixel-Level in Whole Slide Images without Manual Annotation|| |
Fangyao Hu1, Jennifer Giltnane2, Zineb Mounir3, Cleopatra Kozlowski1
Departments of 1Safety Assessment, 2Research Pathology and 3Oncology Biomarker Development, Genentech Inc., San Francisco, CA, USA. E-mail: [email protected]
Content: Accurate semantic segmentation in scanned hematoxylin and eosin (H&E) stained pathology slides of tumor samples may allow deep and detailed biologically relevant analyses. However, currently the standard practice in training a semantic segmentation deep learning network requires hundreds or more pixel-level labeled images created by pathologists, which is highly labor-intensive and inaccurate at the single cell level. We developed an immunohistochemistry (IHC) serial section based method to generate labeled images for tumor detection on H&E stained slides, which required minimal input from pathologists while achieving single cell level tumor segmentation. Technology: Image processing was performed with Python, Definiens Developer (Definiens Inc, Munich, Germany) and Slidematch (MICRO DIMENSIONS, Munich, Germany). A fully convolutional neural network from Visual Geometric Group (FCN-VGG) was trained with Caffe framework. Design: We obtained serial sections of strong HER2-positive invasive breast carcinoma resections. One slide had been stained with H&E, and the other with HER2, that we used as a tumor marker. The H&E and HER2 slides were imaged in a slide scanner at magnification 20x. Of these 20 slide pairs were selected for training, and 3 for testing. The serial section slide images were aligned using Slidematch. Traditional image analyses techniques (image thresholding and hole-filling algorithms) in Definiens were used to create pixel-level binary mask labels from the HER2 images. The slide images (~40k x ~40k pixels) were then tiled at full resolution to create multiple 512x512 pixel images per slide. The HER2 labels were used to train a FCN-VGG network on corresponding serial H&E images. Results: A pixel-level accuracy of over 93% was achieved for FCN-VGG in a testing dataset (N=3). [Figure 1] shows the representative H&E images, gold standard tumor masks generated from the HER2 slides and the corresponding tumor area segmented by our deep learning models. Scale bar is 0.25 mm. Conclusions: Our approach not only achieved tumor segmentation with high accuracy on H&E slides, but also required minimal time from pathologists. Our approach can also be applied to other tumor indications and markers of interests.
| Deep Learning-Based PD-L1 Tumor Cell Scoring of Resected Nonsmall-Cell-Lung-Cancer|| |
Nicolas Brieu1, Armin Meier1, Ansh Kapil1, Aleksandra Zuraw1, Craig Barker2, Marietta L. Scott2, Tobias Wiestler1, Moritz Widmaier1, Keith Steele3, Marlon C. Rebelatto3, Günter Schmidt1
1Definiens, Munich, Germany, 2AstraZeneca, Cambridge, UK, 3MedImmune, Gaithersburg, MD, USA. E-mail: [email protected]
Content: PD-L1 expression measured by immunohistochemistry helps identify Non-Small-Cell-Lung-Cancer (NSCLC) patients that may respond to anti-PD-1/PD-L1 immunotherapies. We present a novel deep learning solution for the automated scoring of PD-L1+ tumor cells (TC) in whole slide images of resected NSCLC. To our knowledge, this is the first deep learning-based approach for the analysis of PD-L1 stained images. Technology: We use a convolutional neural network (CNN) for the fine grained classification of tissue regions into three classes: (1) regions of membrane-positive epithelial tumor cells TC(+), (2) regions of membrane-negative epithelial tumor cells TC(-) and (3) other regions that could wrongly influence scoring, i.e. macrophages, positive and negative lymphocytes, stroma and/or necrosis. We calculate the TC score as the ratio of the area of the classified TC(+) region to the sum of the areas of the classified TC(-) and TC(+) regions. Design: Two sets of ~225k training and ~30k testing patches (128x128 pixels) were created from manual partial annotations from (N=22) train and (N=12) test slides (Ventana-SP263). Training a modified inception network yielded maximum accuracy of 0.89 on test patches. The trained network was applied on (N=433) unseen confirmation slides and the TC score calculated for each slide based on the classified TC(+) and TC(-) regions. A non-linear gamma mapping to the manual TC scores by a trained pathologist was then estimated to maximize Overall Percent Agreement (OPA) at ≥25% criterion using two-fold cross-validation. Results: Evaluation against pathologist scoring on the confirmation slides [Figure 1]a and [Figure 1]b yielded higher Overall (OPA), Negative (NPA) and Positive (PPA) Percent Agreement values at ≥25% criterion, higher Lin’s correlation and lower mean absolute error than a baseline approach relying on a heuristic detection of individual epithelium cells (ESMO-2017-103P). Scoring by a second pathologist was available on a subset (N=170) of the confirmation slides. Average and standard deviation results on this subset [Figure 1]c confirm the above observations and suggest that our approach is getting close to inter-pathologist variability. Conclusions: Using deep learning to identify PD-L1 positive and negative tumor cell regions enables the automated scoring of PD-L1 TCs at the ≥25% expression level in resected NSCLC. These findings should be confirmed with additional tumor sets.
|Figure 1: (a) Evaluation of the deep learning-based PD-L1 tumor cell proportion scoring against pathologists on the unseen resectates NSCLC slides; (b) Pairwise scatter plots between the first reference visual scores by pathologists and the automatic scores, generated either using the baseline heuristic approach (left) or the proposed deep learning approach (right); (c) Quantitative concordance measures between the first visual scores by pathologists and the two automatic scoring approaches; On the subset of samples with two visual scores by pathologists available, mean and standard deviation of the aforementioned measures for the two approaches over these two visual scores, and concordance between these two visual scores|
Click here to view
| Community Crowdsourcing Tool to Expedite Annotations for Deep Learning in Pathology|| |
Liron Pantanowitz1, Erastus Allen2, Keith Callenberg2, Adit Sanghvi2, Sara E. Monaco1, Juan Xing1, Brian Kolowitz2
1Department of Pathology, University of Pittsburgh Medical Center, 2Division of Enterprises, University of Pittsburgh Medical Center, Pittsburgh, Pennsylvania, USA. E-mail: [email protected]
Content: A major bottleneck for developing deep learning algorithms is getting pathologists to perform timely annotations on digital images. Crowdsourcing is one mechanism that can alleviate this barrier. However, there are limited commercial tools for crowdsourcing in computational pathology. Our aim was to build such a tool and validate its feasibility for deep learning. Technology: A crowdsourcing tool, using Meteor JS framework (https://www.meteor.com/), was designed to maximize cross platform access. Whole slide images were converted to JPEG files and regions of interest then presented to users via this tool to perform annotation tasks. Design: Digital patches containing cells derived from Aperio whole slide images were selected for annotation. Each task (e.g. determination of cell type) was optimized for maximum efficiency and minimal time spent performing the task. Users had the option to perform or skip tasks. A ruler was included along with each image patch as a reference for cell size. 8037 annotations were recorded for 3 expert pathologists over 16 days. Results: The crowdsourcing tool functioned well. End user feedback was positive. For this 3-class annotation task the reviewers exhibited 64% concordance, and majority consensus (≥2 of the raters in agreement) was available for 96% of images. Concordance between pairs of raters was 76%, 75% and 71% for Raters 1 and 3, 1 and 2, and 2 and 3, respectively. Annotation tasks were performed with median and minimum completion times of 1.4 and 0.3 seconds, respectively. The mean task completion time for tasks where all 3 raters agreed was 4.0 seconds, in contrast to 5.0 seconds when 2 raters were in agreement or 7.7 seconds when all raters disagreed. Conclusion: Our novel crowdsourcing tool for Pathology facilitated quick and easy annotations of thousands of cells by several expert pathologists. Performing the same tasks directly on entire whole slide images would take much longer and be more laborious. Reducing the time burden and cognitive load with this tool allows the end user to focus and gives the deep learning development team the freedom to customize the user experience and collect expedited annotation data.
| Automatic Cancer and High-Grade Cancer Detection and Localization on Whole-Mount Digital Histopathology Images of Mid-Gland Radical Prostatectomy Specimens|| |
Wenchao Han1,2,3, C. Johnson1, M. Gaed4, J. A. Gomez4, M. Moussa4, J. L. Chin4,5, S. E. Pautler4,5, G. Bauman2,4, A. D. Ward1,2,3,4
1Baines Imaging Research Laboratory, London Regional Cancer Program, Departments of 2Medical Biophysics, 3Pathology and Laboratory Medicine, 4Surgery and 5Oncology, Western University, 6Lawson Health Research Institute, London, Ontario, Canada. E-mail: [email protected]
Content: There is an unmet need for quantitative and graphical pathology reporting for prognosis and adjuvant therapy treatment planning post radical-prostatectomy. Our goal is to develop a software system which detects and localizes cancerous and high-grade cancerous foci on mid-gland whole-slide-images (WSIs) of radical prostatectomy specimens. Technology: We obtained 199 WSIs of mid-gland hematoxylin and eosin-stained sections scanned at 20X (0.5 µm/pixel) from 49 radical prostatectomy specimens. Tumors were manually contoured and graded by a genitourinary pathologist at full resolution. Design: Computations were conducted independently on 480 µm χ 480 µm sub-images completely covering each WSI. 33 WSIs from 8 patients were used for system tuning and a separate set of 166 WSIs from 41 patients comprising 703,745 480 µm χ 480 µm sub-images was used for validation. The system: (1) created tissue component maps labeling each pixel as nucleus, lumen, and stroma/other using color deconvolution and our novel adaptive thresholding algorithm to compensate for staining variability; (2) extracted first- and second-order statistical features from the tissue component maps, selecting 13 via backward feature selection; (3) classified sub-images as (high-grade) cancer or non-cancer using supervised machine learning with fisher and logistic classifiers (PRtools 5.0, Delft Pattern Recognition Research, The Netherlands), and a support-vector-machine (OpenCV 3.1); (4) performed leave-one-out and 2-fold cross-validation using expert-drawn contours on the validation dataset, with sample grouped on per-patient basis. Results: Our system generates graphical whole-slide cancer and high-grade cancer maps as shown in [Figure 1]. The best performing SVM classifier yielded an AUC of 0.95 ± 0.04 with error, false negative, and false positive rates of 10.6% ± 4.7%, 12.3% ± 11.4, 10.5% ± 4.7%, respectively in leave-one-patient-out cross-validation for cancer detection. The fisher classifier yielded an AUC of 0.93 ± 0.07 with error, false negative, and false positive rates of 12.6% ± 9.2%, 17.3% ± 29.8%, 12.5% ± 9.2%, respectively for high-grade cancer detection.Processing time for an un-optimized single-threaded Matlab 2017a (The Mathworks, Natick, MA) implementation is approximately 1 hour/WSI of size ~48 billion pixels. Conclusion: Our system demonstrated accurate performance for cancer and high-grade cancer detection on mid-gland prostate WSIs, despite staining variations. System performance is stable with respect to different training sample sizes for cancer detection, indicating that the system may be ready for multi-center validation.
|Figure 1: (a) Cancer and high-grade cancer map. Red label: System predicted high-grade cancer foci. White label: System predicted low-grade cancer foci. Grey label: System predicted noncancer region. (b) Mid-gland hematoxylin and eosin stained digital histology image with pathologist-drawn cancer contours overlaid, with different colors indicating different grades as in the legend. (d) Region of interest zoomed in from the region indicated in (b) with an arrow|
Click here to view
| A Robust Machine Learning Algorithm for Better Detection of Thyroid Follicular Neoplasm|| |
Keluo Yao1, Xin Jing1, Amer Heider1, Judy C. Pang1, Robertson Davenport1, Madelyn Lew1
1Department of Pathology, University of Michigan, Ann Arbor, Michigan, USA. E-mail: [email protected]
|Figure 1: The segmentation and feature extraction of follicular cells start with the original image (a), followed by the background subtraction (b), conversion to 8-bit grey scale image (green channel) through color deconvolution (c), automatic threshold segmentation, and nuclear feature extraction (d). The 8-bit grey scale image (c) can also be processed with gaussian blur (e) followed by threshold segmentation to extract architectural information represented as particles (f)|
Click here to view
| Real-Time Sharing of Digital Intraoperative Telepathology Consultations with Surgeons|| |
Jennifer Picarsic1, Ishtiaque Ahmed2, Douglas J. Hartman1, Clayton Wiley1, Jeffrey Fine1, Anthony Piccoli2, Jon Duboy2, Matthew O’Leary2, Liron Pantanowitz1
1Department of Pathology, School of Medicine, UPMC, University of Pittsburgh, 2Division of Information Services, Enterprise Pathology, UPMC, Pittsburgh, Pennsylvania, USA. E-mail: [email protected]
Content: Efficient communication between subspeciality pathologists and surgeons during an intraoperative consultation (IOC) is critical for specialized surgeries. Telepathology for IOC permits evaluation by subspecialty pathologists, who are often not in the same physical location as the surgery. A platform to improve telepathology dialogue among clinicians for enhanced patient care and quality assurance is needed. To address this need, we present a web-based solution that allows real-time sharing of digital frozen sections within the operating room. Technology: Aperio LV1 (Leica Biosystems, Vista, CA, USA) hybrid robotic and whole slide image scanner with console (version 220.127.116.1112) located at Children’s Hospital of Pittsburgh/UPMC (CHP). GoToAssist (LogMeIn, Boston, MA, USA) remote desktop solution. Spacedesk (datronicsoft, Metro Manila, Philippines) freeware to expand Windows desktop screen to other networked computers. Design: When a CHP neuropathology frozen section is requested, a trained pathology assistant prepares and loads glass slides into the CHP-LV1 and contacts an off-site pediatric neuropathologist (NP). The NP accesses the CHP-LV1 workstation via GoToAssist from their remote workstation to launch the console software. The CHP-LV1 also hosts the spacedesk DRIVER application which broadcasts the desktop to other dedicated computer monitors on the network. When the pediatric neurosurgery team (CHP-OR) launches Spacedesk VIEWER on the operating room workstation this immediately connects them to the CHP-LV1 desktop, allowing simultaneous viewing of the same digital slide as the remote NP user, in real-time. Results: Despite three locations [Figure 1], with multi-monitor broadcast capability enabled by Spacedesk software, the NP teleconsultant was able to both control the CHP-LV1 desktop for intraoperative evaluation and concurrently share their findings with the CHP-OR team. All parties used mobile phones for verbal communication. Conclusions: Employing multi-monitor desktop sharing and broadcast software permits real-time visualization of digital slides by the NP during IOC while simultaneously allowing the pediatric neurosurgeon to also visualize what is being reviewed. This platform improves telepathology dialogue among clinicians for enhanced patient care and quality assurance.
|Figure 1: Multi-location platform for pediatric neurosurgery intraoperative consultation telepathology service. (a) Neuropathologist user controlling remote Children’s Hospital of Pittsburgh-LV1 system; (b) Children’s Hospital of Pittsburgh-LV1 scanner and workstation, and (c) Children’s Hospital of Pittsburgh-OR surgeon viewing real-time broadcast desktop|
Click here to view
| Real-Time Streaming Telepathology between Networks: A New Approach with Scalers, Encoders and Dedicated Network Pathways|| |
Andrew Quinn1, Jyoti Balani1, Kyle Molberg1
1Department of Pathology, University of Texas Southwestern Medical Center, Dallas, Texas, USA. E-mail: [email protected]
Content: Real-time streaming of microscope camera feeds, such as for rapid on-site evaluations, intra-operative consultations and educational conferences, is essential to pathology enterprises in distributed care delivery environments. Traditional technologies were unable to consistently meet our pathologists’ demands owing to undesirable lags and choppiness. We present an approach that addresses these limitations in the setting of educational conferences. Technology: Our solution employs two Microcast HD HDXS cameras (Optronics, Goleta, CA, USA) with digital video image output to an EXT-MFP (multi-format processor) scaler (Gefen, Petaluma, CA, USA) at one hospital and a DM-RMC-SCALER-C (digital media) scaler (Crestron Electronics, Rockleigh, NJ, USA) at another hospital, each with high-definition multimedia interface outputs to an SX10 encoder (Cisco, San Jose, CA, USA) with ethernet output to an inter-institutional link, which amounts to a protocol with dedicated inter-network pathways that prevent interfering transmissions and exposure to the Internet. Design: We deployed encoders in each of two hospital conference rooms with distinct networks. Each encoder can only initiate voice-over-internet-protocol sessions with the other (point-to-point connection) via the inter-institutional link. The video output of each encoder is set to the input it receives from the microscope camera via the scaler with one only conference room sharing video feed at one time. The encoders also handle audio during sessions. Continued on next page. Results: The above design [Figure 1] allows for true real-time streaming of microscope camera feeds with no measurable lag. At rest, cellular and subcellular details faithfully resemble those viewed via microscope objectives. When moving slides across microscope stages in the usual fashion, streamed images exhibit no measurable choppiness, but suffer from pixelation, precluding cellular and subcellular resolution at magnifications less than 100x and 200x, respectively. Said pixelation resolves near-instantaneously when slide movement ceases. Architecture is always resolvable. Pathologist satisfaction is high. Conclusions: In an educational conference setting, we have developed a viable, real-time solution to streaming microscope camera feeds between hospitals with distinct networks. Adoption of such technologies has been slow in our hands, so a step-wise augmentation is planned, including validation for use in clinical case consultation, the addition of contact points and the conversion to session-initiation-protocol sessions.
| A Reader Study on a 14-Head Microscope|| |
Brandon D. Gallas1, Jamal Benhamida2, Qi Gong1, Matthew G. Hanna2, Partha P. Mitra3, S. Joseph Sirintrapun2, Kazuhiro Tabata2, Yukako Yagi2
1FDA/CDRH/OSEK/DIDSR, Silver Spring, MD, 2Memorial Sloan Kettering Cancer Center, 3Neuroscience, Cold Spring Harbor Laboratory, Cold Spring Harbor, NY, USA. E-mail: [email protected]
Content: In this work, we conducted two feature studies on detecting mitotic figures (MFs) with whole slide images (WSI) and a microscope. Technology: Supervised image analysis algorithms are only as good as the ground-truth on which they are trained and tested. The most practical ground-truth is a pathologist’s assessment with WSI. These are limited as the pathologist is unable to focus on nearby planes of a section (as can be done on a microscope). Another limitation arises from inter-pathologist variability. To overcome these limitations, we propose collecting ground truth from multiple pathologists using a microscope. Design: We used a custom hardware and software evaluation environment for digital and analog pathology that allows us to automatically present the same regions of interest (ROIs) to a pathologist on a microscope or WSI. In Study 1 we collected MF counts and locations in 40 ROIs from 4 H&E slides of canine oral melanoma (five pathologists, institutional guidelines regarding animal experimentation were followed). The ROIs were 200 um x 200 um (800 x 800 pixels at 0.25 um/pixel; Aperio AT2). Study 2 was conducted on a 14-head microscope (four original + six new pathologists, working independently). We collected MF counts and locations on the same 40 ROIs. In Study 2 we also asked the pathologists to quantify their confidence that a candidate was an MF. Results: In Study 1, the pathologists identified 94 “candidate” mitotic figures, and they identified more with the microscope than with the WSI [Table 1]. We call them candidate MFs because only 18 of 94 were unanimously identified. In Study 2, the pathologists identified 170 candidates. More pathologists lead to more candidates. Lastly, we did not find noteworthy differences in the between-reader variability in count differences across the modalities studied [Table 1]. More results will be presnted at the conference. Conclusion: Detecting and quantifying mitoses is an important pathology task when evaluating tumors of various subtypes; it is also challenging and burdensome to pathologists, subject to significant pathologist variability. Future studies are underway, leveraging the results of these two studies, to train or test an automated mitosis detection algorithm.
| Automated Imaging and Scoring of Histological Specimens during Routine Pathology Workflow|| |
Gabe Siegel1, Edmund. Sabo2
1Augmentiqs, Misgav, 2Lady Davis Carmel Medical Center, Haifa, Israel. E-mail: [email protected]
Content: Large scale collection of pathology images will play a key role in developing deep-learning artificial intelligence algorithms for pathology diagnostics. Image capture from the existing pathology workflow of pathologist and microscope, without active participation on the part of the pathologist, could potentially result in gathering relevant regions of interest, in a fast and cost-efficient manner. These collected regions of interest can be reexamined by the pathologist for assuring accuracy of the diagnosis and research purposes. Technology: A novel system based on augmented reality was added to the pathologist’s existing microscope. The system integrates within the microscope’s optical plane, adding a digital overlay of the tissue and embedded image sensor of the field of view – thereby enhancing the existing microscope with digital capabilities. Without altering the optical view, the system enables multiple digital pathology applications and automated image capture within the existing workflow. The system also allows morphometric measurements, annotations and other tools when needed. Design: The pathologist reviewed 10 histological specimens of colonic tumors (premalignant, malignant and normal controls) immunohistochemical stained, each case containing 4 slides. The image sensor was commanded to automatically capture images according to the methodology of the pathologist during the review, and a score would be given based on preset parameters. The image score was set to rise according to length of time pathologist stopped movement of the stage, changing of the magnification and use of available morphometric tools. Results: The review of the slides lasted under 30 minutes and resulted in 348 images. Images varied in score, such that a majority had a low score indicating a brief stop of stage movement, while the regions of tissue containing relevant clinical data had a higher score indicating a longer stopping of the microscope stage. The pathologist reported no slowdown of workflow. Conclusion: It is feasible to automatically capture images of tissue during the routine pathology workflow, and to provide scores to images where regions of interest are likely to be found based on the methodology of the pathology review. Captured images can be combined with specific case data and other metadata from the pathology workflow to further enhance image value.
| Use of Telepathology for Pathology Collaboration and Peer Review in Multinational Studies|| |
G. Siegel1, D. Regelman1, R. Maronpot2, M. Rosenstock3, A. Nyska4
1Augmentiqs, Misgav, Israel, 2Maronpot Consulting LLC, Raleigh, NC, 3LEA, Nonclinical Safety Consultancy, Talmei Elazar, Israel, 4Toxicologic Pathology and Tel Aviv University, Timrat, Israel, USA. E-mail: [email protected]
Content: Toxicologic pathology is a highly collaborative science that relies on both real time consultation and “peer review”. The consultation stage of study ensures both quality and timeliness, and has a direct impact on the drug development cycle. The use of real-time telepathology can enhance collaboration and expedite the peer review process, while reducing the need for travel and increasing efficiency. Technology: Using an existing microscope with a novel telepathology system (AugmentiqsTM) that operates by sharing high resolution live views of slides on the microscope stage with remote parties. Design: Following completion of the pathology evaluation of a preclinical toxicity study at a laboratory in the United States, histopathology slides were shipped to the peer review pathologist (PRP) located in Israel. The PRP reviewed the slides, and then using the telepathology technology and his microscope, conducted a live telepathology session with the study pathologist (SP) sharing the slides for which there were questions or disagreements in the initial diagnoses. Results: The PRP was able to simultaneously share the actual histopathology slides with the SP and obtain a consensus diagnosis for lesions in question. The SP was able to see the slides on his personal computer screen in high resolution, and discuss these lesions with the PRP, while microscopic fields were instantly photographed in publication quality high resolution and saved by the participating pathologists. Following this live telepathology session and achieved consensus, the SP issued a revised report expressing the agreed upon consensus diagnoses. Following completion of the entire peer review process and updating of the study pathology report, a formal GLP-compliant Peer Review Statement was signed by both the SP and PRP for regulatory submission. Conclusion: Telepathology running live and directly off the microscope is a highly cost and time efficient method for conducting peer review with documented images. Based upon our experience, telepathology running off the microscope can be used for general collaboration, peer review and other GLP-compliant review applications.
| A Web Application Based “Cockpit” for Protein Electrophoresis|| |
Keluo Yao1, Christopher L. Williams2, Ulysses G. J. Balis1, David S. McClintock1
1Pathology Informatics, University of Michigan, Ann Arbor, MI, 2Pathology, University of Oklahoma, Oklahoma City, OK, USA. E-mail: [email protected]
Content: The Sebia Capillarys (Sebia, Issy-les-Moulineaux, France) middleware application provides a limited user interface, functionally, and portability to adequately manage the protein electrophoresis (PEP) assay results produced by the instruments. Routine workflow requires manual entry of laboratory data and physical onsite interpretation. This study investigates the feasibility of building a “cockpit” for interpretation of protein electrophoresis using application programming interface (API) of electronic health record (EHR) and advances in web-application technology. Technology: Node.js v8.0, D3.js v4.11.0, Web pack v4.0.0, Riot.js v3.0.7, PostgreSQL v10.3, Representational state transfer (REST), Radis v4.0.8. Design: Our goals include: 1) provide an automated data request from the EHR; 2) create a secure, flexible, and interactive user interface unifying relevant clinical, laboratory, and PEP data; and 3) streamline the pathologist interpretation workflow to save both time and resources. Using the recently available REST based web API, which enables reverse federated architecture, and server-side scripting through the node. js ecosystem, we are developing a custom web a p p l i c a t i o n integrating data from our EHR (Epic) and our protein electrophoresis middleware application. Results: [Figure 1] shows the overall node.js based design of the PEP web application. We will have read-only interfaces with the EHR (Epic) and Sebia PostgreSQL database, each connected via a dedicated data API server. We will create a dedicated database (PEP cockpit server) to manage extracted data from EHR and the Sebia PostgreSQL database and store user data as well as application settings. A central data manager server will handle all data streams and relay only pertinent data to the user interface server, and the data flow will be cached by redis and the PEP cockpit database. The user interface server uses express.js for web application framework, Riot.js for custom reusable HTML tags and React like functionalities, Webpack for on-demand code generation, and D3.js for PEP diagrams. We have built a working demo that proves the feasibility this web application and a wireframe that demonstrates end-user experience. Conclusion: The use of a modern web-based architecture to integrate multiple sources of disparate clinical data has great potential to revolutionize PEP workflows by consolidating multiple manual, labor-intensive processes into a single streamlined application.
| EHR-Based Assessment of the Current Practice of Screening for Primary Aldosteronism|| |
Xiruo Ding1, Daniel Herman1
1Department of Pathology and Laboratory Medicine, University of Pennsylvania, Philadelphia, PA, USA. E-mail: [email protected]
Content: Primary aldosteronism (PA) comprises approximately 5% of hypertension and is treatable with targeted medications or adrenalectomy. Clinical guidelines recommend screening for PA in specific subsets of hypertension patients by quantifying blood aldosterone and renin. However, it is unclear how well we identify these patients. Notably, amongst known PA patients, it is not diagnosed until after a median of ~10 years of hypertension. To assess the need for improved screening, we sought to describe the practices across our institution. Technology: We extracted clinical encounter, laboratory, diagnosis code, and note data from the University of Pennsylvania clinical data repository and EPIC Clarity by SQL queries via python, processed results into encounter-level information using python and R, and stored intermediate results in a SQLite database. Design: We selected patients with at least 5 encounters with blood pressure reported or at least 3 such encounters in distinct years between 1/1/2007 and 12/31/2017. Patients were considered to have documented hypertension if they had ≥2 encounters with hypertension diagnosis codes and considered to have PA if they had ≥2 PA diagnosis codes, met strict laboratory criteria (Aldosterone ≥15 ng/dL, plasma renin activity ≤0.5 ng/mL/hr, Aldosterone: renin ≥30 (ng/dL)/(ng/mL/hr)), or underwent adrenal vein sampling. Results: We surveyed 2,208,984 office visits for 207,172 patients seen at 24 practice locations over 8.6±4.2 (mean ± sd) years. The prevalence of documented hypertension was 24.6% ± 6.8% across sites. Among hypertension patients, only 0.5% ± 0.4% were documented to have PA. The distribution of PA frequencies across sites was strongly associated with the frequency of laboratory PA screening (3.3% ± 2.9%; r=0.9; p=3x10-9) and referral to renal specialists (5.9% ± 7.0%; r=0.8; p=1x10-5). Conclusions: The substantial gap between the frequency of PA we observed and the expectation from population studies, suggests that we are dramatically underdiagnosing and thus undertreating PA patients. The strong correlation, among sites, between the PA frequency and both ordering PA screening tests and specialty care referrals, implies that increased screening should lead to increased PA diagnoses. To this end, we are developing automated tools to identify patients likely to have PA for laboratory screening.
| Immediate Bed Side Estimation of Trabecular Bone Fraction in Freshly Obtained Bone Marrow Core Biopsies|| |
Srikanth Ragothaman1, Rajan Dewar1, Riccardo Valdez1
1Pathology, University of Michigan, Michigan, USA. E-mail: [email protected]
Content: Bone marrow core biopsies have a significant proportion of trabecular bone. Routine histopathological examination enables identification of trabecular bone pathology, such as thickened bony trabeculae in osteopetrosis, end stage primary myelofibrosis (osteosclerosis) or thin trabeculae in osteopenia or renal osteodystrophy. Trabecular bone and the region immediately adjoining it (paratrabecular space or osteoblastic niche) is also affected in various pathologies including preferentially involved by lymphoma, such as follicular lymphoma, and stem cell development processes (instance PGE2 stimulated marrow). Technology: We have developed an automated image analytic method for accurately quantifying trabecular bone fraction (TBF) from freshly biopsied bone marrow cores, by x-ray analysis. We propose that this bedside method could gain clinical utility in patients with BM pathology. Design: 10 consecutively collected bone marrow specimen were utilized in this study. Soon after the core biopsy was obtained, the specimen was placed in a cabinet model x-ray analysis instrument, that is routinely used for breast specimen radiography (Faxitron, Tucson, AZ, USA). Bone marrow images were obtained at a high resolution. An image analytic algorithm using ImageJ software was developed that could segment the bony trabeculae based on density and calculate the trabecular bone fraction (TBF) from the input dimensions of the marrow (constant width) as shown in [Figure 1]. Histological correlates were performed by routine microscopy and computation of trabecular bone fraction was done automatically using ImageJ software, as shown in [Figure 1]. Results: Results of TBF is given as a percentage fraction of bone volume to the total biopsied area. The estimates in the 10 specimens range from 12-38% and maximum of 25% deviation from microscopy images (H&E). This discrepancy is thought to be due to the three dimensional images obtained by X-ray techniques, compared to 2D images of routine microscopy. Conclusion: The clinical utility of measuring bone marrow trabecular fraction in routine hematopathology is not yet explored. We feel that the x-ray based image analysis enables immediate results, 3D reconstruction of bone marrow biopsy and accurate estimation of trabecular bone fraction and fibrosis than conventional H&E/trichrome and Reticulin stains. Further refinement of these techniques and correlation with a larger clinical cohort of myelofibrotic patients and other patients with bone marrow pathology, is warranted.
|Figure 1: X-RAY image of the bone marrow core biopsy(right). Green areas in the left indicates automatic trabecular bone region quantification|
Click here to view
|Figure 2: Bone marrow core biopsy (right). Blue areas in the left indicates automatic trabecular bone region quantification|
Click here to view
| Strong Correlation of Sample Collection Date and Patient Admission Date in Microbiological Testing Complicates Sharing of Phylodynamic Metadata Sharing|| |
Alex Greninger1, Ryan Shean1
1Pathology, University of Washington, Washington, USA. E-mail: [email protected]
Content: Infectious pathogens are known for their rapid evolutionary rates with new mutations arising over days to weeks. The ability to rapidly recover whole genome sequences and analyze the spread and evolution of viral pathogens using metagenomics and sample collection dates has lead to interest in real-time tracking of infectious transmission and outbreaks. However, the level of temporal resolution afforded by these analyses may conflict with definitions of what constitutes protected health information and privacy requirements for de-identification for publication and sharing of such data. In the United States, dates and locations associated with patient care that provide greater resolution than year or the first three digits of the zip code are generally considered patient identifiers; admission and discharge dates are specifically named as identifiers in Department of Health and Human Services guidance. Technology: Retrospective laboratory information system review with candidate chart review. Design: To understand the degree to which one can impute admission dates from specimen collection dates, we examined sample collection dates and patient admission dates associated with more than 270,000 unique microbiological results from the University of Washington Laboratory Medicine Department between 2010 and 2017. Cumulative distribution curves were plotted and compared using two-sample Kolmogorov-Smirnov tests. Results: Across all positive microbiological tests, the sample collection date exactly matched the patient admission date in 68.8% of tests. Collection dates and admission dates were identical from emergency department and outpatient testing 86.7% and 96.5% of the time, respectively, with more than 99% of tests collected within one day from the patient admission date. Samples from female patients were significantly more likely to be collected closer to admission date that those from male patients. Conclusions: We show that protected health information-associated dates such as admission date can confidently be imputed from deposited collection date. We suggest that publicly depositing microbiological collection dates at greater resolution than the year may not meet routine Safe Harbor-based requirements of patient de-identification. We recommend the use of Expert Determination to determine protected health information for a given study and/or direct patient consent if clinical laboratories or phylodynamic practitioners desire to make these data available.
| Approaching Preanalytical Automation in Next Generation Sequencing|| |
Andrew Quinn1, Brandi Cantarel2, James Malter1, Dwight Oliver1, Benjamin Wakeland3
Departments of 1Pathology, 2Bioinformatics and 3Immunology, University of Texas Southwestern Medical Center, Dallas, Texas, USA. E-mail: [email protected]
Content: Next generation sequencing (NGS), owing to molecular testing roots and the (potential) need for tumor and non-tumor tissues for results, is ill-suited to traditional provider order entry. We developed a nearly fully automated pre-analytical workflow with focused manual interventions. Technology: The workflow includes an electronic health record (Epic, Verona, WI, USA), interface engine (Cloverleaf, Infor, New York, NY, USA), three laboratory information systems (Sunquest CoPathPlus and Laboratory, Tucson, AZ, USA; Clarity, Illumina, San Diego, CA, USA) and custom code (Python, Wilmington, DE, USA). Design: Via ask-at-order-entry questions, providers identify tumor and non-tumor tissues. The interface engine translates question responses into HL7 requests for formalin-fixed, paraffin-embedded tissue (CoPathPlus) and/or fresh blood/marrow/saliva (Sunquest Laboratory). Pathology staff checks CoPathPlus for NGS orders daily, creating tissue designation cases with rush priority and assigning them to pathologists (pathologists receive daily automated email reminders to finalize). HL7 messages to CoPathPlus pre-populate all NGS details except tissue blocks for processing. Once finalized, pathology staff orders NGS block sectioning protocols. In Sunquest Laboratory, pathology staff receives and accessions blood/marrow/saliva specimens and affixes labels (specimens collected by clinics; Epic labels affixed initially). Finalization of tissue designation cases and accessioning of fresh specimens trigger HL7 messages to the interface engine. The interface engine digests all inbound HL7 messages, transmitting contents via XML to a network directory. Custom code collects XML files there and uploads contents to Clarity via its application programming interface. NGS staff monitors the entire process in Clarity. Results: Pre-analytical turnaround times are incalculable because providers can place orders prior to collection. When providers identify archival, formalin-fixed, paraffin-embedded tissues, slides arrive in the NGS lab within two days (fresh specimens collected/transported to NGS lab within 24 hours via usual channels). No orders have been lost in translation. The below design [Figure 1] was developed via two pre-live and one post-live iterations over six months. Conclusion: Pre-analytical NGS automation is achievable, requiring significant clinical, information technology and laboratory resources. The above does not include billing and consent confirmation prior to testing (NGS staff handles in Epic). Additionally, fresh specimen tracking is captured via Sunquest Laboratory batches.
| Whole-Tissue Phenotyping via Three-Dimensional Reconstruction of Human Gastrointestinal Tissues|| |
Navid Farahani1, Liron Pantanowitz2, Douglas J. Hartman2
13Scan Inc., San Francisco, CA, 2Department of Pathology, University of Pittsburgh Medical Center, Pittsburgh, PA, USA. E-mail: [email protected]
Background: Over the past decades, pathologists have relied on thin (< 10-μm-thick) 2D slices and light microscopy in order to render a pathologic diagnosis. As our understanding of disease has expanded, it has become clear that studying 3D structures such as tumours and other pathologic specimens in 2D results in a significant loss of information between the recorded data and the state of the original tissue. A recent study using fluorescence has identified the generation of tertiary lymphoid organs in the mesentery of specimens from patients with Crohn’s disease. Herein, we evaluate the usefulness of automated serial-section microscopy for whole-tissue phenotypic evaluations of clinical tissue. Technology: Following deparaffinization, blocks were whole-mount stained with H&E, re-embedded and scanned using a 4th gen. Knife-Edge Scanning Microscopy (KESM) platform with a voxel size of 0.7 x 0.7 x μm. Digitized samples were stored using Amazon’s S3 web service (Seattle, WA). Design: Two deidentified formalin-fixed, paraffin-embedded blocks of waste small intestinal tissue (1 control, 1 test) were shipped to 3Scan (San Francisco, CA). Custom-built software was then used for annotation, 3D reconstruction, and visualization of tissue microarchitecture. Results: Clinical FFPE tissue, which has undergone traditional 2D histopathologic analysis, can be easily repurposed for 3D analysis. An average of 1270 serial sections were required for complete tissue exhaustion within each sample. The average accumulated file size was 8.5 TB and the sample digitization process required an average of 8 hours. 3D analysis revealed several striking findings within the tissue microarchitecture including numerous vascular abnormalities within the test sample [Figure 1]. Summary: KESM and other emerging imaging methods can yield novel insights into the underlying nature of normal and/or diseased tissues which are comprised of innumerable 3D structures. Given the relative infancy of this field, more work needs to be performed characterizing normal and abnormal 3D tissue structures. Here we leverage 3D analysis to generate a whole-tissue phenotype in a patient with Crohn’s disease [Figure 2], which revealed dilated and congested blood vessels extending from the submucosa through the muscularis propria.
|Figure 2: Two separate screenshots of Crohn’s sample at different levels|
Click here to view
- Tanaka N, Kanatani S, Tomer R, Sahlgren C, Kronqvist P, Kaczynska D, et al. Whole-tissue biopsy phenotyping of three-dimensional tumours reveals patterns of cancer heterogeneity. Nat Biomed Eng 2017;1:796-806.
- Randolph GJ, Bala S, Rahier JF, Johnson MW, Wang PL, Nalbantoglu I, et al. Lymphoid aggregates remodel lymphatic collecting vessels that serve mesenteric lymph nodes in Crohn disease. Am J Pathol 2016;186:3066-73.
| Meta-Analysis Reveals Multiple Drivers of Crohn’s Disease Pathogenesis|| |
Jihad Aljabban1, Kamal Khorfan2, Laraib Z. Safeer3, Hisham Aljabban4, Maryam Panahiazar5, Dexter Hadley5
1Ohio State University College of Medicine, Columbus, Ohio, 2Henry Ford Health System, Detroit, Michigan, 3Baylor College of Medicine, Houston, Texas, 4Regis University, Denver, Colorado, 5Institute for Computational Health Sciences, University of California, San Francisco, California, USA. E-mail: [email protected]
Content: Demonstrate the utility of our meta-analysis platform to elucidate pathogenesis of Crohn’s disease (CD). Technology: The National Center for Biotechnology Information Gene Expression Omnibus (GEO) is an open database of more than 2 million samples of functional genomics experiments. Our Search Tag Analyze Resource for GEO (STARGEO) platform allows for meta-analysis of genomic signatures of disease and tissue through tagging of individual samples across different studies. Design: We analyzed 441 intestinal biopsies from CD patients against an equal number of healthy intestinal samples as a control using STARGEO, and then we analyzed the meta-data using Ingenuity Pathway Analysis. Results: Intestinal sample analysis revealed agranulocyte/granulocyte adhesion and diapedesis, atherosclerosis signaling, role of interleukin-17α, and hepatic fibrosis as top canonical pathways. We also observed activation of colorectal cancer metastasis and the Triggering Receptors Expressed on Myeloid cells (TREM1) signaling. Interferon-γ [Figure 1], STAT3, interleukin-1α, and TNF were top upstream regulators. We noted upregulation of a list of cytokines, genes implicated in shaping the intestinal microflora and in extra-intestinal disease, and downregulation of genes involved in DNA stability and metabolism. Conclusion: Our analysis builds off of known characteristics of CD such as the role of immune mediators, including interleukin-1α and 17α, interferon-γ, TNF, and TREM1, in driving pathogenesis. Upregulation of hydroxycarboxylic acid receptor 3 (HCAR3, a regulator of macrophage reactivity to gut luminal contents) suggests abberant reactivity to gut microbes. Additionally, our analysis links CD to extra-intestinal pathologies associated with inflammatory bowel disease such as atherosclerosis and hepatic fibrosis. A dysbiotic gut microbiome leads to maladaptive immune development and pathogen regulation. Our analysis showed marked upregulation of anti-bacterial proteins such as dual oxidase 2 (produces reactive oxygen species) and lipocalin 2 (sequesters iron). Overactivity of these proteins could impair mucosal barrier defense and increase microflora antigenicity. We also found decreased genetic stability and metabolic function, which is consistent with activation of the colorectal cancer metastasis pathway in our analysis. We found downregulation of MSH5 (DNA mismatch repair), MT-ND3 (subunit of NADH dehydrogenase), and SLX1A/B (regulates genetic stability). Lastly, we illustrate how interferon-γ drives the disease processes described above through regulation of key proteins [Figure 1].
|Figure 1: Interferon-g regulation on molecules implicated in metabolic dysregulation and intestinal immune response. Prediction legend illustrates relation of molecules|
Click here to view
| Bridging the Collaboration Gap: Real-Time Identification of Clinical Specimens for Biomedical Research|| |
Thomas Durant1,2, Wade Schulz1,2
1Department of Laboratory Medicine, Yale School of Medicine, 2Center for Outcomes Research and Evaluation, Yale New Haven Health, New Haven, CT, USA. E-mail: [email protected]
Content: Biomedical and translational research often rely on the identification and testing of patient specimens that meet specific laboratory or diagnostic criteria. Historically, the process of identifying samples which meet inclusion and exclusion criteria is a manual process, and requires a research assistant position that exists outside the clinical workflow, which can be cost prohibitive. Emerging data management technologies offer novel approaches to enhance specimen identification practices. To this end, we present a component of a data science platform developed at our institution – Baikal – adapted to stream processing, identification, and notification of clinical specimens for translational research projects. Technology: Components include the Hortonworks Data Platform (version 2.4.2; Hortonworks, Santa Clara, CA, USA) and Hortonworks Data Flow (HDF) version 1.2 (Hortonworks, Santa Clara, CA, USA). Custom Python (version 2.7) scripts are executed within NiFi to identify specimens based on streaming laboratory results from the Cloverleaf integration engine (Infor, NY, USA). Data from HL7 ORU messages are stored and visualized with the ELK stack (version 6.2.2; Elastic, Mountain View, CA, USA). Notification of specimen availability is pushed to users with Watcher (version 6.2.2; Elastic, Mountain View, CA, USA). Design: A custom emissary service was deployed to receive a stream of HL7 ORU messages from Cloverleaf, which were validated and transformed into JSON documents for storage and stream processing. A streaming workflow, including a custom Python script, was used to identify relevant specimens and target them for notification. Meta-data from filtered specimens were routed to Elasticsearch for visualization and user notification. Results: Specimen Flagging gives our institution the ability to maintain real-time dashboards displaying relevant specimen information for translational research studies. End-users are able to view dashboards which provide specimen IDs and location to facilitate sample acquisition by laboratory personnel. In addition, end-users can subscribe to email push notifications to alert them when a sample that meets their predefined criteria has been processed by the lab. Conclusion: This work demonstrates that adoption of emerging data management technologies can offer extended capabilities for translational research in the clinical laboratory with minimal overhead. Future research will seek to evaluate implementation efforts to determine the benefit of real-time specimen flagging.
| Development of a Laboratory-Focused Data Warehouse Using Open Source Software|| |
Patrick Mathias1, Niklas Krumm2, Noah Hoffman2
Departments of 1Laboratory Informatics and 2Laboratory Medicine, University of Washington, Washington, USA. E-mail: [email protected]
Content: Ready access to data from the laboratory information system (LIS) is essential for effective quality improvement in the laboratory, especially when integrated with data from other information systems. Commercial solutions may not easily integrate data from other information systems, or may impose rigid data models insufficient for monitoring complex workflows. To address our institution’s needs, we designed a data warehouse and developed a supporting application to build the database using open source software on existing departmental infrastructure. Technology: A Python (version 3.6) application to process data was developed to implement the data warehouse using PostgreSQL (version 9.6). To facilitate development, the database schema is defined using SQL statements constructed using templating language (Jinja2) that facilitates conditional statements and variable substitution. A command line interface is used to clean and l oa d data and perform maintenance tasks. Data is loaded into the database from files transferred from information systems via scheduled tasks. Design: The data warehouse schema was designed by eliciting stakeholder needs and identifying key data elements for core business processes. A star schema was adopted to develop the first set of tables in the data warehouse that detail laboratory test orders. Views represent transactional data in a format more convenient for common queries. While the application is LIS-agnostic, our schema was developed based on the data structure of Sunquest Laboratory (version 7.2, Tucson, AZ). Results: The initial phase of the application includes tables that represent laboratory test orders and associated transactional order events. To demonstrate its functionality, we populated the database with 3 months of laboratory orders and order events, consistingof 1.7 and 7 million rows, respectively. Using our initial schema, we replicated the functionality of an existing custom program to generate tallies of orders, and added the capability to filter on characteristics such as priority and cancellation codes, and partition by features such as performing lab or patient location. Conclusions: The ability to efficiently analyze large volumes of laboratory data is a key competency for computational pathology. We demonstrate a laboratory-focused data warehouse built with open-source tools to enable our laboratory to answer clinical and operational questions.
A Searchable Electronic Archive of Biobank Specimens from Paper Pathology Reports: 1-Year Pilot, Overmapping Functional Ontologies (Systematized Nomenclature of Medicine Reference Code Clinical Terms, International Classification of Diseases-O-3, Medical Subject Headigns)
Cullen R. Vos1, Paraic A. Kenny1, Grzegorz T. Gurda1,2
1Kabara Cancer Research Institute, 2Department of Pathology, Gundersen Medical Health Systems, La Crosse, WI, USA. E-mail: [email protected]
Content: Manual data entry to convert paper pathology records to searchable, electronic biobank database can be labor intensive and fraught with error. Additionally, biospecimen depositories house complex data that span multiple knowledge realms and do not fit into a single ontology. Here, we describe a pilot to automate electronic biobank record creation, with emphasis on assignment of bioontology classifications and how it affects searchability. Technology: 16,000+ pathology reports, corresponding to decommissioned slides/blocks from year 1997 were scanned with BookEye V4 (Image Access, Wuppertal, Germany). Corresponding LogicalDOC repository files were parsed via Python v2.7 scripts to identify 2223 potential malignancies, and thru negation logic/manual review 1825 (~82%) were added into Gundersen Medical Foundation biobank (GMF-BB). Cases were assigned ontology codes via online ontology browsers and R Bioconductor packages (R-project.org). Design: Pathology diagnosis text, microscopic description key words and Systematized Nomenclature of Medicine reference codes (SNOMED RT) were used to generate SNOMED-Clinical Terms (SNOMED CT). Batched identification of International Classification of Diseases for Oncology (ICD-O-3) codes and Medical Subject Headigns (MeSH) terms was also performed, with manual review vs. SNOMED CT and pathology reports. Searches of uncommon malignancies (prevalence 0.2%-2%) were used to test interoperability and search yield. Results: Automated data extraction is less labor intensive and (likely) less prone to operator error/variance. Updating legacy ontology (SNOMED RT) to SNOMED CT and overmapping additional ontologies was relatively quick (<15 hrs), amendable to batch processing and thus more standardized. In searches, SNOMED CT and ICD-O-3 morphology codes showed good overlap (~80%, range 67-100%), increasing sensitivity of the database search by up to 25%, without loss of specificity. Union vs. intersection of ICD-O-3 morphology and topography codes can pinpoint metastatic vs. primary disease, or subcategorize ontologically difficult to separate entities, i.e. cervical vs. vulvar SIL. Overmapping MeSH terms was more difficult, but MeSH may be useful as metadata, such as receptor status in breast cancer, or molecular alterations in lung cancer. Conclusions: The pilot project showed utility in processing decommissioned pathology specimen in automated fashion to generate a searchable, data-rich electronic database and thus increasing utility of GMF-BB for future research.
| Using Clinical Laboratory Data to Forecast Clinical Decompensation|| |
Danielle E. Kurant1, Jason Baron1
1Department of Pathology, Harvard Medical School, Massachusetts General Hospital, Boston, MA, USA. E-mail: [email protected]
Content: Advance warning of impending clinical decompensation may provide opportunities for early clinical interventions to improve outcomes. We identified patients admitted to non-intensive care inpatient medicine units and used transfer to the intensive care unit (ICU) as a surrogate for clinical decompensation. By comparing patterns of laboratory test results between patients who were and were not transferred to the ICU, we sought to build a laboratory-based model to forecast clinical decompensation. Technology: We trained and tested a logistic regression model using open source packages within the R statistical scripting language. Design: We compiled patient and laboratory data from our hospital’s laboratory information system for patients admitted to selected hospital locations over a one-year period via a laboratory datamart. Using a case control design, we compared medicine inpatients who were transferred to the ICU during admission (“cases”) to patients who were not transferred (“controls”). Defining “event time” for cases as time of ICU transfer, and for controls as a random time selected from a distribution intended to match the cases, we used test results preceding event time by 12 hours or more as predictors. By applying generalizable inclusion criteria and down sampling the controls, our dataset included 119 cases and 239 controls which were randomly divided into training and test partitions in a 2:1 ratio. We trained and tested a logistic regression model using this data. Results: On the test partition, the logistic regression model demonstrated some discriminative power to distinguish patients who were transferred from those who were not with an area under the curve (AUC) of 0.70. At a sensitivity (proportion of cases correctly classified) of 80%, the model provided a specificity of 61%. We are continuing to enhance our framework and use more sophisticated models to improve performance. Conclusions: These data serve as proof of concept that laboratory data may be useful in forecasting clinical deterioration. We hope to refine our model to achieve sufficient sensitivity and specificity to provide a foundation for laboratory-based clinical decision support that can alert to impending clinical decompensation.
| Whole Slide Imaging Equivalency and Efficiency: A Large Academic Center Experience|| |
Matthew G. Hanna1, Victor E. Reuter1, Meera R. Hameed1, Lee K. Tan1, Sarah Chiang1, Carlie Sigel1, Travis Hollmann1, Dilip Giri1, Jennifer Samboy1, Carlos Moradel1, Andrea Rosado1, John R. Otilano1, Christine England1, Lorraine Corsale1, Evangelos Stamelos1, Yukako Yagi2, Peter Schueffler2, Thomas Fuchs2, David S. Klimstra1, Sahussapont Joseph Sirintrapun1
1Department of Pathology, Memorial Sloan Kettering, New York, USA, 2Warren Alpert Center for Digital and Computational Pathology. E-mail: [email protected]
Content: Whole slide imaging (WSI) has been approved for primary diagnosis in the US, however few US pathology departments have fully implemented a digital pathology system. WSI has significant potential to disrupt pathology clinical practice, however studies pertaining to the efficiency of a true digital pathology workload are lacking. The aim of this study was to replicate clinical workflow in comparing a true digital pathology signout compared to optical, and compare equivalency and efficiency of glass slide and WSI reporting of true clinical workload. Technology: Glass slides were scanned on Leica Aperio AT2 at 40x (0.25um/pixel). The Leica eSlide manager database interfaces with the laboratory information system (LIS), Cerner CoPathPlus, integrating WSI for each accessioned case. WSI viewing was performed using an internally developed agnostic WSI viewer, MSKCC Universal Slide Viewer, which includes slide navigation and annotation tools (i.e. thumbnails, ruler). Pathologists utilized a dual monitor (1920x1200) and standard institution computer configuration (8GB RAM). Design: Subspecialized pathologists reported optical signout using routine clinical workflow. Glass slides were deidentified, scanned, and reaccessioned in the LIS test environment. After a washout period of 3 months, reported the same clinical workload using WSI integrated within the LIS. Intraobserver equivalency metrics included top line diagnosis, margin status, lymphovascular and/or perineural invasion, pathology stage, and ancillary testing (i.e. recuts, IHC). Turnaround time (TAT) efficiency evaluation was defined by starting each case when opened in the LIS and completing the case for that day (case sent to signout queue or pending ancillary studies). Results: Eight pathologists participated (subspecialties: bone and soft tissue, genitourinary, gastrointestinal, breast, gynecologic, and dermatopathology). 2091 glass slides, from 204 cases were reported. Digital signout of 199 cases, 2073 WSI, was performed. Median file size, 1.54 GB; scan time/slide, 6 min 24 sec; scan area 18.52x32.1mm. [Table 1] shows equivalency concordance and TAT efficiency. No difference by subspecialty or specimen type was identified. Conclusions: Our experience is the most comprehensive study, and shows high intraobserver digital to glass slide equivalence in reporting of true clinical workflow and workload. Digital pathology showed a median overall 19% decrease in efficiency per case. This may be ameliorated by better input devices and computational pathology.
|Table 1: Intraobserver concordance and turnaround time of glass and whole slide imaging reporting|
Click here to view
| Intelligently Auto Filled Peripheral Blood Smear and Bone Marrow Templates|| |
Rufei Lu1, Ngoc Tran1, Christopher William1
1Department of Pathology, University of Oklahoma Health Science Center, Oklahoma City, Oklahoma, USA. E-mail: [email protected]
Content: The microscopic interpretation reports of peripheral blood smears (PBSM) and bone marrow (BM) often lack standardization, despite a well-defined diagnostic algorithm. Furthermore, the recipient physicians prefer the reporting format they are familiar with. Thus, creating a universal and versatile reporting templates can not only improve efficiency, reduce turnaround time, minimize errors, but also aid in the education of pathology residents as a report practicing composer as well as non-pathology recipient residents. Technology: The template application (app) is largely web-based using HTML 5.0. Real-time interactive features of the app are achieved by using Knockout JS (v3.5.2), Bootstrap (v4.0.0), QuillJS (v1.3.5), and jQuery (v3.2.1). Design: The app encompasses both PBSM and BM reporting templates. The dynamic drop-down tab on the left 40% of screen allows navigation through different portions of report with ease. A real-time report is visible and editable on the right, in case of patient-specific comments. Institution and age specific reference ranges for CBC can be entered for auto-fill and differential diagnoses generation portion of the app. Although we lack capability of integrating LIS into this web-based app, backend of the app can be made to handle automatic reporting of CBC values into the app using the KnockoutJS and jQuery. Results: Utilizing the new dynamic feature of Bootstrap, we are also able to create a web-based, device-responsive, versatile, and user-friendly app can be accessed via intranet on any device [Figure 1]. This app structures the input in hematopoietic lineage based fashion, however, dynamic does also customize sequence of input. This streamlining of input is of great educational purpose for pathology residents, especially who are early on their learning curve. Additionally, the app has a built-in simplified differential diagnoses generator based on entered CBC values. It also contains auto-filling of most common ICD-10 codes, resident/pathologist signature, and current date. Conclusion: The implementation of this app not only will streamline a standardized workflow with a personalized reporting format, but also provides the clinicians with a format they are most comfortable with. The synoptic-like format can also potentially facilitate the structured data extraction for both clinical and research purposes.
|Figure 1: Screenshots of the layout of the one-paged web-based app (a) and some example of collapsible tab on the left for the ease of control and filling (b)|
Click here to view
| Implementing the DICOM Standard for Digital Pathology|| |
Markus D. Herrmann1, Sean W. Doyle1, Veronica Klepeis2, George L. Mutter3, David Milstone3, Nicholas Jones2, Gopal K. Kotecha1, Long P. Le2, John Gilbertson2, David H. Hwang3, Mark Michalski1,4,5, Katherine Andriole1,4, A. John Iafrate2, James Brink5, Keith Dreyer1,5, Jeffrey Golden3, David Louis2, Jochen K. Lennerz2
1MGH and BWH Center for Clinical Data Science, Departments of 2Pathology and 5Radiology, Harvard Medical School, Massachusetts General Hospital, Departments of 3Pathology and 4Radiology, Harvard Medical School, Brigham and Women’s Hospital, Boston, Massachusetts, USA. E-mail: [email protected]
Content: Applications of digital pathology go beyond interactive viewing and require structured machine-readable data. In particular machine learning demands integration of image data with clinical metadata and efficient programmatic data access. A common data format – akin to DICOM in radiology– has not been adopted in pathology practice. Here, we implemented the proposed DICOM standard for digital pathology and assessed its practicality and performance. Technology: DICOM standard version 3.0. Design: We selected attributes from the “Visible Light Whole Slide Microscopy Image” Information Object Definition (supplements 122/145). We implemented an algorithm to generate DICOM data sets, building on existing DICOM software libraries. We prospectively tracked our software engineering-and project management efforts, verified completeness and integrity of generated data sets using existing DICOM archives and validation tools, and benchmarked data storage and access efficiency for different compression methods. Performance metrics included conversion times, relative data size, and access rates. Results: Taking into account our clinical case review specifications, we selected a set of n=124 DICOM attributes (n=103 ‘required’; n=21 ‘optional’). Our algorithm includes the following subroutines: (a) extract pixel data and pixel-related metadata from vendor-specific file formats, (b) obtain relevant patient and specimen-related metadata from the laboratory information system, (c) encode DICOM data elements. We encoded frames in the Pixel Data element using different compression methods. We created DICOM series for 20 slides by combining image data from 5 image file formats with clinical metadata from different sites and sources. Prospective tracking (120h documentation review, 153h algorithm development, 120h software implementation, and 48h consultation with content experts) equates to 0.66 full-time equivalents (FTE) for a 90-day period or ~59.4 days full-time effort. Our program requires several minutes to encode a slide (JPEG: ~12 min, JPEG-LS: ~25 min, JPEG2000: ~71 min) and resulting data sets increase in size (JPEG: ~2.1x, JPEG-LS: ~7.6x, JPEG2000: ~9.2x). However, individual frames can be read and decoded fast (JPEG:~4.2 ms, JPEG-LS: ~4.7 ms, JPEG2000: ~49.6 ms). Conclusions: We implemented the DICOM standard for digital pathology in a multi-site healthcare network. The complexity of the standard is initially daunting and poses several IT-challenges; however, once implemented it facilitates efficient storage and access of image data and clinical metadata.
| Private Blockchain Distribution Network for Blood-Derived Biologic Products|| |
Thomas Durant1,2, H. Patrick Young2, Alex Ryder3, Eric Gehrie4 , Christopher Tormey1, Wade Schulz1,2
1Department of Laboratory Medicine, Yale School of Medicine, 2Center for Outcomes Research and Evaluation, Yale New Haven Hospital, New Haven, CT, 3Children’s Foundation Research Institute and University of Tennessee Health Science Center, Memphis, TN, 4Department of Pathology, the Johns Hopkins Medical Institutions, Baltimore, MD, USA. E-mail: [email protected]
Content: Production and distribution of biologic-derived blood products (BDBP) involves a complex set of federally mandated logistics protocols that are unique to the field of blood banking. Specific complexities arise in the context of quality assurance standards governing BDBP distribution, storage and infusion, and protection of protected health information (PHI) between disparate entities throughout a distribution network. The application of asymmetric cryptography protocols, such as blockchain, would permit the use of customized and immutable protocols that would allow users to standardize, validate, and automate the look-back process across entities within a distribution network. Technology: Components include MultiChain software (v1.0.2), based on the Bitcoin protocol, for the implementation of a secure, private blockchain. A web service implemented in Python (version 3.5) is used to interface BDBP related metadata with the blockchain. Custom Python scripts are executed within NiFi (v1.4.0) to hash BDBP message transactions. All scripts and software were deployed within Docker containers to increase portability and security of the application infrastructure. Design: A NiFi workflow monitors source systems for messages relating to the production, modification, and transfer of BDBPs. Messages are hashed to create a cryptographic representation of the message transaction. The hash, file metadata, and timestamp are then sent to a Python web service that logs the data into the private blockchain. Data streams based on product metadata are instantiated into the application – BloodChain – to allow query access by authorized users. Results: Implementation of BloodChain would allow for the maintenance of associated product and donor information through cryptographically verified message transactions across a distributed ledger. As a semi-private network, permissioned access to the distributed ledger allows users to perform queries of the blockchain data, specifically regarding BDBP supply, product state, and donor state, while maintaining the entity and donor anonymity. Conclusion: Semi-private blockchains can be used to create cryptographically-secure data governance for BDBPs. This would address the unique need of sharing information across BDBP distribution networks, while maintaining the security and privacy of entities and donors, respectively.
| Development and Implementation of a Comprehensive Transfusion Management and Utilization Platform|| |
Burak Bahar1, Nathan B. Price2, William M. Byron2, Thomas J. S. Durant1,3, Amit Gokhale1, Eric A. Gehrie4, Edward L. Snyder1, Wade L. Schulz1,3
1Department of Laboratory Medicine, Yale School of Medicine, 2Department in Information Technology Systems, Yale New Haven Health, New Haven, CT, 3Center for Outcomes Research and Evaluation, Yale New Haven Hospital, 4Department of Pathology, Johns Hopkins School of Medicine, Baltimore, MD, USA. E-mail: [email protected]
| Predictive Analytics to Identify Patients that May Benefit from Pharmacogenomic Screening|| |
Dustin R. Bunch1,2, Thomas J. S. Durant1,2, Burak Bahar1, Nathaniel Price2, Charles J. Torre Jr2, H. Patrick Young2, William Byron2, Joe M. El-Khoury1, Wade L. Schulz1,2
1Department of Laboratory Medicine, Yale-New Haven Hospital, 2Center of Computational Health, Yale University, New Haven, CT, USA. E-mail: [email protected]
Content: Pharmacogenomics can play a major role in an individual’s response to drug therapy. Improper dosing in patients with genetic interactions can reduce efficacy or increase morbidity and mortality. Real-time genetic screening is not feasible due to turnaround time and cost restraints. Therefore, our group attempted to generate a predictive algorithm that could identify patients who may benefit from preemptive genetic screening based on the likelihood of requiring medications with possible genetic interactions in the future based on prior healthcare utilization. Technology: Data were extracted from the Epic Caboodle data warehouse to a Hadoop-based data science environment. A data sandbox for downstream analytics was generated from these data with Spark, and supervised machine learning (extreme gradient boosting, logistic regression, and random forest) was done in Python with the pandas, NumPy, and scikit-learn libraries. Design: Electronic health record data screened for patients ages 18-89 who were prescribed warfarin or haloperidol between 1-1-2013 and 10-30-2017. For this retrospective analysis, we extracted demographics, prescribed medications, encounter types, and ICD10 diagnoses, consolidated into clinical classifications software codes, present at least 30 days prior to the medication being prescribed. Encounter types and other medications were summarized by occurrences in periods of 30, 120, 365, and >365 days prior to target medication start. Negative control patients had other common medication prescriptions, minus warfarin or haloperidol, over the same period. To improve feature learning, training datasets were balanced and included warfarin (7,309/7,363 patients/controls) and haloperidol (7,461/7,485 patients/controls). Validation datasets consisted of warfarin (3,171/54,524 patients/controls) and haloperidol (3,215/54,425 patients/controls). Results: The three algorithm accuracy for possible warfarin use was 76-77% for our validation data with area under the receiver operator curves (AUROC) 75-79%. For haloperidol, the accuracy was 83-84% with AUROC 75-80%. Conclusion: We found with this preliminary study that machine learning algorithms may predict patients who are likely to receive medications that have known genetic interactions and therefore may benefit from preemptive pharmacogenomics screening. While the algorithm may be more predictive of high healthcare utilization than specific medication prescription, it appears that it may be an effective measure to indicate who may benefit most from pharmacogenomic screening.
| Laboratory Information System Storage Requirements for Whole Slide Imaging in the Histology Laboratory|| |
Thomas Chong1, Angela Chen1, W. Dean Wallace1, Chris Khacherian1
1Pathology, David Geffen School of Medicine, University of California, Los Angeles, California, USA. E-mail: [email protected]
Content: As pathology laboratories contemplate the benefits and challenges of digitization of glass histology slides, there are important factors to consider that may stress IT capabilities or prove costly if not appropriately evaluated. Having a clearer understanding of the system storage requirements will help laboratories make informed management decisions. One significant consideration is the size of the digital image files produced during Whole Slide Image (WSI) creation. An image file from one slide may be larger than 1 gigabyte, and routine high throughput scanning could exhaust storage capacity. Management of these large WSI databases may impact WSI retrieval and affect performance. The aim of the study was to assess storage requirements and performance characteristics, for a typical workday volume at our laboratory. Technology: We scanned 1115 consecutive anatomic pathology biopsies and excisions from various organ sites with a Leica Aperio AT2 400-slide capacity slide scanner. Design: Glass histology slides were randomly selected from our pathology archive from the previous 12 months. In total, 1115 slides (109 cases) were scanned. This approximates 1 day of work for our in-house pathology service. The slides were scanned at 20x resolution, creating an SVS image file per slide. For each slide, a single rectangular region of interest (ROI) for maximum resolution capture was automatically determined by the scanner software. A subset of these ROI’s required subsequent manual adjustment to correct specimen coverage and minimize blank areas. The following data were captured: specimen type, image size (megabytes, MB), compressed file size (MB), image compression ratio, ROI height (pixels), ROI width (pixels), and scanning time. Results: The image files required an average of 0.39 terabytes (TB) of storage per day, or 142TB per year, with a total scan time of 22 hours and 54 minutes per day. Average/median image size: 4717MB/4414MB uncompressed Average/median file size: 349MB/255MB compressed Average compression ratio: 17.9 Average/median scan time: 1:14/1:10 Average image width: 30.8mm, 61341 pixels Average image height: 13.8mm, 27500 pixels (multi-modal distribution). Conclusions: The implementation of routine and high throughput whole-slide image scanning in the pathology laboratory requires considerable IT infrastructure and support. Our data provides useful information for developing guidelines for laboratory needs.
| Development of a Deep Watershed Transform Instance Segmentation Method for Nuclei Segmentation in Histopathologic Images of Breast Cancer|| |
Maozheng Zhao1, Rajarsi Gupta2, Le Hou1, Dimitris Samaras1, Tahsin Kurc2, and Joel Saltz2
Departments of 1Computer Science and 2Biomedical Informatics, State University of New York at Stony Brook, NY, USA. E-mail: [email protected]
Content: Robust and reliable segmentation and classification of nuclei in breast cancer continues to be a very challenging problem in digital histopathologic image analyses due to the inherent complexity of overlapping and pleomorphic nuclei that are present in highly complex and invasive tumor growth patterns that distort tissue microarchitecture. Unfortunately, commonly utilized segmentation methods have not yet resolved the problems of clumping aggregates of nuclei. This study proposes a deep watershed transform instance segmentation method to predict a segmentation mask to better separate and localize nuclei by utilizing the local maximal points of the predicted depth map. Technology: There are two networks in the pipeline, where the first network predicts the gradient direction map of the watershed transform of each instance and the second network predicts the watershed depth map from the direction map. The deep watershed transformation performs instance level segmentation by taking the RGB image gated by the class level segmentation mask as the input, where the final output is a mask for each instance with improved separation of nuclei. The final segmentation masks are based on the thresholds generated from the depth map. Design: We developed a deep watershed transform instance segmentation method for nuclei segmentation in a subset of whole slide images of high-grade infiltrative breast cancer images obtained from the public The Cancer Genome Atlas (TCGA) repository, where previously employed segmentation algorithms resulted in clumping nuclear aggregates with poor separation of nuclei. Results: In [Figure 1], panel A shows a representative area of overlapping tumor nuclei, panel B shows the predicted center of the nuclei from the local maximal points from the predicted depth map, panel C shows the depth map generated from the ground truth segmentation mask, and panel D shows the predicted depth map with local maximal points highlighted by blue dot. The final segmentation mask is resulted by taking a threshold of 1 on the depth map. The quantitative results are shown in [Table 1]. The instance level segmentation by deep watershed transform reduced the nuclei counting error by 3.62% and improved the DICE average by 0.0017. Conclusions: Commonly used segmentation methods mostly focus on class level segmentation of nuclei that segment areas containing nuclei areas versus areas that do not contain nuclei, which cannot get a good instance level segmentation for close and overlapping nuclei in histopathologic images of breast cancer. The deep watershed transformation instance segmentation method appears to be a feasible alternative to predict reasonably good segmentation mask with better localization of nuclei that can be used to generate improved counts of nuclei for quantitative immunohistochemistry analyses. Since breast cancer is a heterogeneous disease with morphologically and genetically distinct histologic types, molecular profiles, and biological behaviors, this method can potentially improve the quantification of extracted nuclear features that cannot be performed by a pathologist using traditional light microscopy in order to further enhance our understanding of breast cancer prognosis and response to therapy.
|Figure 1: (a) shows a representative area of overlapping tumor nuclei, (b) shows the predicted center of the nuclei from the local maximal points from the predicted depth map, (c) shows the depth map generated from the ground truth segmentation mask, (d) shows the predicted depth map combined with local maximal points. The final segmentation mask is a result of setting a threshold of 1 on the depth map|
Click here to view
| Support Models for Laboratory Information Systems: A Survey of Pathology and Non-Pathology Informaticist Leaders|| |
Christina Gutierrez1, Walter H. Henricks2, Myra Wilkerson3, Alexis B. Carter4
1Department of Pathology and Laboratory Medicine, Emory University, 4Department of Pathology and Laboratory Medicine, Children’s Healthcare of Atlanta, Atlanta, Georgia, 2Cleveland Clinic, Cleveland, Ohio, 3Laboratory Medicine, Geisinger Health System, Danville, Pennsylvania, USA. E-mail: [email protected]
Content: Recent emphasis on electronic health record (EHR) implementation has led many institutions to shift Laboratory Information System (LIS) support from the Laboratory to the Central Information Technology Department (CITD). The purpose of this survey is to gain insight into current support models and determine informaticist satisfaction with LIS support models. Technology: A 40-question survey was created using an online survey tool (Survey Monkey, Palo Alto, California, USA). Design: Participants were recruited through Informatics professional society listservs (API, AMDIS, AMIA) and excluded if they had no clinical systems oversight or no experience in an institution with an EHR and a LIS. Demographic, architecture, staffing, and satisfaction data were collected for up to 3 LISs per respondent. Results: Twenty-six complete responses were included. Fifty-one LISs were reported, including forty-two (82.4%) separate interfaced LISs. Results are summarized in [Table 1]. Support staff and maintenance fees in the CITD were attributed to networks, EHR interfaces, desktop support, and printing, while those in the Laboratory were attributed to the LIS application, instrument/middleware interfaces and label printing. Conclusions: Despite higher CITD employees responsible for LIS support, the pathology group had higher satisfaction. A potential contributing factor may be the higher number of Laboratory employees supporting LISs in the pathology group. In fact, the greatest recommendation for staffing improvement among non-pathologists was to hire more Laboratory employees to support the LIS. These data indicate higher satisfaction when a greater proportion of LIS support staff have laboratory domain expertise.
|Table 1: Summary of laboratory information system support survey results, comparing pathology and nonpathology informaticists|
Click here to view
| Time and Motion Study for Single Sign-On Solution (Caradigm) in Subspecialty Practice|| |
Douglas J. Hartman1, Jeffrey Fine1, Chelsea Watkins1, Jennifer Picarsic1, Liron Pantanowitz1
1University of Pittsburgh Medical Center, Pittsburgh, PA, USA. E-mail: [email protected]
Content: Due to more complex medical care, it has become important to review clinical records from multiple electronic health records so that a comprehensive pathology diagnostic report can be issued. In our health system, we have adopted a “best of breed” model and so pertinent clinical data may be contained within disparate systems. We implemented a single sign-on solution to share patient context across these various systems. This study examined the time impact for retrieval of patient records utilizing this solution for surgical pathology and cytopathology practice when signing out cases. Technology: Caradigm (version 6.2.46546.18723) for single sign-on and context management (Caradigm, Bellevue, WA, USA). Design: Two pathologists (gastrointestinal subspecialist, cytopathologist) collected time stamps when retrieving data from electronic systems (e.g. CoPath, Cerner, laboratory information system, Powerchart, Cerner, electronic medical record) both without and using Caradigm. Time to retrieve records for 200 gastrointestinal cases (100 with/without Caradigm) and 40 cytology cases (20 with/without Caradigm) were collected. For the gastrointestinal sign-out workflow, the same location within the patient chart was accessed for each subsequent case; while different locations were accessed for cytology workflow. Results: For the gastrointestinal cases, the average time to pull up records using Caradigm was 9.1762 seconds (range 6.71-11.40 secs) compared to 14.3196 seconds (range 11.42-23.01 secs) without Caradigm. For cytology cases, the average time to pull up records using Caradigm was 70 seconds (range 28-94 secs) compared to 205 seconds (range 95-291 secs) without Caradigm. For all cases the patient chart selected was automatically and correctly opened for all requisite applications. Conclusion: Use of a streamlined single sign-on system decreases the time for pathologists to retrieve information from multiple electronic health records. Depending on the subspecialty, there can be an approximately 0.5-3X time savings when signing out cases. Patient-centric navigation between the lab information and electronic medical record systems also indirectly promotes patient safety.
| Successful Deployment of Decision Support and Data-Mining in Patient Blood Management|| |
Jason Kang1, Polina Imas1, Catherine Saporito1, Peter Arroyo1, Gregory Wright1, James T. Perkins1
1Pathology, NorthShore University Health System, Evanston, Illinois, USA. E-mail: [email protected]
Content: Modern patient blood management requires 1) an effective alternative to traditional methods of consensus transfusion guideline dissemination and 2) institutional verification of provider adoption of these guidelines. Our institution deployed computer-assisted decision support programming and data-mining in order to meet these requirements. Technology: Our decision support tool guides blood ordering on a foundation of consensus transfusion guidelines via an integrated best practice alert (BPA) at the point of blood ordering. The data-mining algorithm logic searches for transfused red blood cell units (RBCs) and the associated pre-and post-transfusion hemoglobin values. The logic also searches for back-to-back RBC transfusions and identifies these as RBC transfusions that occur within eight hours of each other without an interim hemoglobin. Design: Our institution implemented a patient blood management program in 2012. This program introduced the integrated BPA decision support programming tool into the hospital electronic medical record. Retrospective data-mining commenced in order to confirm compliance with guidelines. Provider notifications and interventions, if necessary, took place retrospectively and at the point of care. Compliance data reporting to the departmental Chair, institutional Transfusion Committee, and institutional Medical Executive Committee took place on a quarterly and/or annual basis. Results: We observed the following: 38% reduction in RBCs transfused per 100 discharges between Q3 FY2013 and Q3 FY2017; 83% reduction in RBCs transfused at hemoglobin>8 between Q3 FY2013 and Q3 FY2017; 93% reduction in back-to-back RBCs transfused between Q3 FY2013 and Q3 FY2017. Taken together, the reduction in transfused RBCs per inpatient encounter was associated with a $2.1M or 46% inflation-adjusted reduction in cost of goods (COGs) between FY2012 and FY2017, i.e. a 57% inflation-adjusted reduction in COGs per inpatient encounter between FY2012 and FY2017 (R2=.99). Conclusions: Decision support programming can replace traditional methods of consensus guideline dissemination. Data-mining can aid verification of institutional and individual provider adherence to consensus transfusion guidelines. More broadly, our institution’s blood management experience demonstrates that computer-assisted decision support and data-mining are integral, complementary components of any stewardship initiative.
| Comparative Analysis of Changes in Pathologists’ Diagnostic Approaches Over Time|| |
Mikhail Kovalenko1, Richard Hammer1, Dmitriy Shin1
1Department of Pathology and Anatomical Sciences, University of Missouri, St. Louis, Missouri, USA. E-mail: [email protected]
Content: A lymphoma diagnosis involves a complex mental process influenced by a multitude of factors ranging from pathologist’s experience to circumstances of the clinical case. Our research is focused on revealing diagnosis-related details and heuristics that can be used to quantify and potentially improve the diagnostic process in pathology using whole-slide imaging and analytical tools. In this work we look at changes in pathologists’ approach to diagnosis as a function of time by capturing their diagnostic activities and comparing them to an earlier set captured two years prior, and whether we can quantify them as “experience.” Technology: We used our PathEdEx whole-slide imaging platform with realistic diagnostic workflow, whole-slide imaging viewing capability, and gaze tracking capture to record user activities related to diagnosing a cancerous tissue slide. Design: Four pathologists with experience ranging from expert to trainee recorded their diagnostic sessions in 2016 and then again in 2018 over the same lymphoma cases using PathEdEx. We represented diagnostic activities and clues recorded during tissue examination as graphs and matrices for easier comparison. We applied novel frameworks developed concurrently in our lab, such as REDESIGN differential signaling framework, to see if there are any changes in individual diagnostic approaches and results over time. Results: As expected, changes in experts’ results were minimal. Junior pathologists showed a small improvement in time and choice of stains considered. We were not able to detect significant differences in individual pathologist’s approaches to diagnosis over time. Conclusions: We suspect that the information captured did not represent a complete set of descriptive parameters to judge learning progress over time with complete confidence. Further research is needed to determine the necessary details for a better evaluation of pathologists’ performance in order to create personalized educational recommendations for diagnostic improvement.
| Process Excellence in a Large Volume Whole Slide Scanning Scanning Facility|| |
M. C. Lloyd1, D. Kellough1, T. Shanks1, A. Deshpande1, W. L. Frankel2, A. Mason1, A. V. Parwani2
1Inspirata Inc., Bengaluru, Karnataka, India, 2Department of Pathology, Wexner Medical Center, Ohio State University, Columbus, Ohio, USA. E-mail: [email protected]
Content: By harnessing the value of a quantifiable new digital pathology data modality in histological images through deep learning and similar computational advancements, it is reasonable that we can uncover both the genomic and the realized morphological changes occurring within tumors and their surrounding microenvironment. To facilitate meaningful artificial intelligence, large data sets are required. Our group is using process excellence to maximize the number of scans which can be acquired rapidly. Technology: Using the Inspirata Solution-as-a-Service workflow and Whole Slide Scanning (WSI) instruments from multiple vendors, we have created a high-throughput scanning facility for the acquisition of hundreds of thousands of WSIs. Design: Our group designed and executed a project to retrospectively acquire images of slides from every cancer case produced by our histology laboratory for the past five years (2012-2016). After 9 months of scanning at full capacity we embarked on a process excellence exercise to continue to improve our scanning throughput. The number of WSIs will exceed 500,000 in less than one calendar year of scanning. We designed the project planning, slide scanning facility, standard operating procedures, records, workflows, staffing, stakeholder engagement activities, system integrations, file storage and IT support, project governance, communication, quality management, risk management, financial planning. Results: 560 square feet of space was allocated and renovated to assemble seven high-throughput slide scanning instruments. Lean processes and waste reduction practices were applied to the layout and workflows through the facility. Since our go-live date we have begun a follow-up process excellence review to enhance the following processes and increase our scanning volume: lab layout (storage and labeling procedures); slide touches (>2,000/day), slide inspection and cleaning procedures (>30sec/slide), scan error rate reduction (>2%) and QC reviews at the scanner and for the WSI (>3%). Conclusion: Computational analyses are emerging as a reliable tool to perform novel research in pathology however creating a massive pipeline of image data has not previously been achieved at this scale. We are the first organization to create a repository of WSIs of this magnitude. We continue to work to enhance our processes to achieve the highest quality WSIs in the shortest period possible.
| Automated Identification of Mitoses using a Hybrid Approach: Combining Deep Learning with Classical Domain-Based Detection|| |
Dave Harding1, Nishant Verma1, Amir Mohammadi1, Ajay Basavanhally1, James Monaco1, Mark Lloyd1, Gary Tozbikian2, Ziabo Li2, Anil Parwani2
1Inspirata Inc., Bengaluru, Karnataka, India, 2Ohio State University, Columbus, Ohio, USA. E-mail: [email protected]
Content: Measures of mitotic activity are highly prognostic of survival times in patients with node-negative breast cancers. However, identifying individual mitotic figures can be a time-consuming task for pathologists and may be subject to significant inter-observer variability. Adoption of digital pathology allows the development of automated image analysis algorithms to assist pathologists in efficiently and reliably identifying mitotic figures. Recent major advances in performance of automated image analysis systems have been seen due to the application of machine learning techniques such as deep learning to these problems. These deep learning systems do not directly embed any of the collective domain knowledge of pathologists in their design, instead finding features in the image data directly through an optimization process. We combined a traditional system using domain-specific features with a modern deep image segmentation network to see if this hybrid system could outperform them both. Technology: The deep image segmentation network we applied was Retinanet (Facebook AI Research, Menlo Park, CA). The traditional system and hybrid systems were implemented and evaluated using MATLAB 2016b (Natick, MA). The digitized image and annotation data was made available by the AMIDA13 MICCAI Grand Challenge and was digitized on an Aperio ScanScope XT at the University Medical Center Utrecht (Utrecht, Netherlands). Design: Each of our systems was individually trained on the training subset of the AMIDA13 data. The mitotic figures detected by each system were combined and reclassified using the confidence values reported by both systems. Performance data was gathered over the disjoint testing subset of the AMIDA13 data. Results: As expected, the deep learning system outperformed the traditional system over this homogeneous challenge dataset. However, when operating over a range of specificities, we found sensitivity improve by 10% by combining the two systems. Specifically, at a rate of.25 false detections per true positive, the traditional, deep learning, and hybrid systems’ sensitivities were 38%, 48%, and 53%, respectively. Conclusion: We find that there is information captured by the domain-specific features of the traditional system that is not represented in the deep learning system. A combination of the two techniques outperformed either system individually.
| Features Used in Computerized Image Analysis Systems Could Benefit Pathologists: Mitosis Detection|| |
James Monaco1, Nishant Verma1, Amir Mohammadi1, Ajay Basavanhally1, Dave Harding1, Mark Lloyd1, Gary Tozbikian2, Ziabo Li2, Anil Parwani2
1Inspirata Inc., Bengaluru, Karnataka, India, 2Ohio State University, Columbus, Ohio, USA. E-mail: [email protected]
Content: A significant limitation of manual histological analysis is the inherent inter-observer variability. One factor contributing to this variability is the imprecise, qualitative nature of the established protocols. For example, in mitosis detection – which is the focus of this abstract – the criteria describing a mitosis are ill-defined, lacking objective, quantitative descriptors. By contrast, computer-aided detection systems employ quantitative features and follow precisely defined protocols. If these features and protocols are both effective (i.e. they perform their tasks well) and intuitive (i.e. they are intelligible to humans), they could be used to better define the protocols followed by pathologists, thus reducing inter-observer variability. Technology: Using MATLAB 2016b (Natick, MA), we developed a system that accurately detects mitoses in digitized H&E stained breast tissue. The system first identifies and segments potential mitoses in each image. Subsequently, the algorithm extracts eight intuitive, biologically relevant features from each detection. These features measure object color, shape, boundary smoothness, and surface shape (e.g. donut, dome, valley). A classifier then leverages these features to label each detection as a mitosis or not. Design: We ran our system over 1179 breast cancer images (600 x 600 microns) from 94 different patients from four institutions (with unique staining protocols), digitized on four different whole-slide scanners. Pathologists identified 3301 mitoses, labeling the phase of each. We extracted the eight feature values for each mitosis. We then determined 1) each feature’s relative importance in identifying mitoses and 2) the mean and standard deviation of each feature for each mitotic phase. Results: Some relevant results are as follows: Boundary smoothness was the most important feature, followed by width and color. Metaphase width values have mean of 3.60 microns with a standard deviation of 0.92, while prophase mitoses have a mean of 5.87 microns and standard deviation of 1.19. Compared with mitoses, detections that are not mitoses have boundaries that are 20% smoother. Conclusions: The mitotic features described herein are biological, intuitive, and effective. Such features could be used to begin to create a quantitative framework for establishing a quantified definition of what constitutes a mitotic figure.
| Application of Image J in evaluating Ki-67 proliferation index in gastrointestinal neuroendocrine tumors|| |
Nitin Marwaha1,2, Vijayalakshmi Padmanabhan1, LindaK. Green2
1Department of Pathology and Immunology, Baylor College of Medicine, 2Department of Pathology, Michael E. DeBakey Veterans Affairs Medical Center, Houston, TX, USA. E-mail: [email protected]
Content: The Ki-67 proliferation index play pivot role in grading many tumors including neuroendocrine tumors of gastrointestinal system. It is a widespread practice to estimate the Ki-67 index by visual inspection, which is very subjective and has poor reproducibility. We wanted to come up with a simple tool which can be used easily by all pathologists and is also cost effective. In this study, we use Image J application, which can be downloaded from NIH website at no cost, to analyze Ki-67 proliferation index in neuroendocrine tumors of gastrointestinal system. Technology: An image J 1.51j8, Windows 64bit (National Institute of Health, Bethesda, Maryland USA) application was downloaded. Photomicrographs were taken using Nikon DS-Fi3 camera mounted on Nikon Eclipse E400 microscope. Design: The cases of neuroendocrine tumors of gastrointestinal system were retrieved by using the Houston VAMC computerized SNOMED system from 1989-2018. Twenty consecutive most recent cases were selected to avoid selection bias. Out of 20 cases, Ki-67 immunohistochemical stain was performed on 10 cases and they were selected for the study. Glass slides were reviewed, and photomicrographs were taken. The images measuring 2880 X 2048 pixels were uploaded on the Image J. The images were converted to 8-bit and the image threshold was set at 0.30%. Analyze particles, function was utilized by setting up the size at 200-infinity and circularity at 0.00-1.00. The photomicrographs were also printed, and manual counting was performed by the senior author, who was blinded regarding the results of Ki-67 index reported on the surgical pathology reports and values calculated by Image J. Results: The image J assisted Ki-67 values were compared with the manual counts (n=10) using paired two tailed t-test and no statistical difference was noted (p=0.62, t stat =0.51, t critical=2.3). Strong positive correlation (Pearson correlation coefficient =0.97) was also noted. Conclusion: Image J can be effectively used in counting Ki-67 index in many tumors including neuroendocrine tumors, gastrointestinal stromal tumors, and lymphomas. This tool can also be used as a quality assurance measure in retrospective analysis of Ki-67 proliferation index.
| Quantified Analysis of Hematoxylin Stain Variations|| |
Yao Nie1, Maria V. Sainz de Cea1
1Roche Tissue Diagnostics, Digital Pathology, Mountain View, CA, USA. E-mail: [email protected]
Content: Performance of image analysis algorithms in digital pathology is often hampered by cross-image stain variations caused by factors including tissue preparation, staining protocol and scanning process. However, few works have been reported for identifying the major sources of the variations. This study investigates the contributions of several factors to the hematoxylin stain variations in immunohistochemistry slides through quantified analysis. Technology: Hue-Saturation-Density transformation is used to convert image pixels’ RGB values to stain’s chromatic and density components. The Kullback-Leibler divergence, which measures the distance between probability distributions, is used to assess stain variations among different biomarkers, scanners, tissue blocks and concentrations. Design: A dataset which contains breast cancer immunohistochemistry slides stained for 3 different biomarkers (i.e., human-epidermal-growth-factor-2, estrogen-receptor and Ki-67) is collected. For each biomarker, 12 tissues blocks are used and each block has 3 sections. From each section, 9 slides are stained at hematoxylin concentration levels ranging from titer 1 to titer 9, which are then scanned by 3 different iScanHT scanners, resulting in 2916 slide in total. The mean Hue-Saturation-Density values of 1.5 million randomly selected blue pixels in each slide are used to represent hematoxylin stain’s chromatic and density properties of that slide. Results: The mean Kullback-Leibler divergence between two groups of pixels randomly sampled from the same staining/scanning condition is 0.05. For a fixed scanner, the mean Kullback-Leibler divergence between two different biomarkers is 1.0, 3.2 and 0.82 for the three scanners, respectively. For a fixed biomarker, the mean Kullback-Leibler divergence between two different scanners is 2.77 for human-epidermal-growth-factor-2, 4.67 for estrogen-receptor, and 4.27 for Ki-67, respectively. For scanner1/human-epidermal-growth-factor-2 combination, the mean Kullback-Leibler divergence is 0.89 between two different hematoxylin titer levels; and is 47.81 between two different tissue blocks, which is significantly greater than other factors. Similar values are observed for other scanner/biomarker combinations. Conclusion: Hematoxylin titer level has the minimal impact to stain distribution in Hue-Saturation-Density space other than shifting along density axis. The impact of biomarkers is also small, followed by that from the scanners. Tissue preparation process is the dominant source of stain variations. Stain normalization algorithm is needed to reduce the impact of such variations.
| Fuzzification of Cancer Staging in Pathology for a Precision Medicine System|| |
1Icahn School of Medicine at Mount Sinai, Mount Sinai Institute for Systems Biomedicine, New York, USA. E-mail: [email protected]
Content: Finding quantitative measures for categorization in pathology is a key element for biomarker discovery and establishing precision medicine protocols. Pathologic staging of cancer is done through qualitative methods and despite our boosting knowledge of bimolecular alterations coming from transcriptomics and genomics of tumors; it is still not a common practice to use a systems-level approach for pathology staging in cancer. One of the challenges of systems-level staging system which considers many parameters including cytology and transcriptomic profiles of cancer tissue is lack of in silico and mathematical methods to handle this task. Technology: Here a mathematical formulation is presented to combine all parameters affecting the accurate staging of cancer using fuzzy logic. Design: Through fuzzification of the staging system instead of having multiple Boolean platforms, a fuzzy membership function acts to assign the staging for a specific tissue. Results: This mathematical methodology is capable of transforming qualitative pathology stating into quantitative setting which is useful for practice in a precision medicine system. Conclusion: Fuzzification of cancer staging in pathology can be a helpful step towards establishing a clinical system for precision medicine.
| Genomic Landscape of Autophagy System Based on Histopathology of Lung Cancers: Lessons from the Cancer Genome Atlas|| |
1Icahn School of Medicine at Mount Sinai, Mount Sinai Institute for Systems Biomedicine, New York, USA. E-mail: [email protected]
Content: Autophagy is a cellular process which plays a key role in homeostasis of normal and cancer cells. Autophagy is involved in both cell survival and cell death and functions as a module in cells to transduce the cellular stress response. Autophagy has a context-dependent function in different tissues under different conditions and biomarkers related to autophagy can be predictive of a tumor response to different therapeutic regimens. Technology: In this project we have analyzed The Cancer Genome Atlas (TCGA) and Cancer Cell Line Encyclopedia (CCLE) using a computational methodology to decode the genomic landscape of autophagy in lung cancer based on histopathology of lung cancers. Design: Genomic and transcriptomic data, histopathology and patients’ data were analyzed in different types of cancers. Results: This landscape is capable of showing a set of molecular biomarkers related to autophagy in different types of lung cancer. Conclusions: This set of biomarkers combined with histopathology is able to provide a map of autophagy system in individual patients which has applications in precision oncology.
| A New Era for Intraoperative Neuro-Telepathology: Impact of Migrating to LVI from Zeiss|| |
Swikrity Upadhyay Baskota1, Liron Pantanowitz1, Clayton A. Wiley1
1Department of Pathology, University of Pittsburgh Medical Center, Pittsburgh, PA, USA. E-mail: [email protected]
Content: With advances in technology improved devices have been developed with better optics, hardware and software to support telepathology. For real-time intraoperative consultation newer instruments offer hybrid robotic and whole slide imaging, faster scans, more user-friendly software, and simultaneous viewing. The aim of this study was to assess the impact on the intraoperative neuro-telepathology (NTP) service when migrating from an old Zeiss (previously Trestle) robotic microscopy platform to a new Aperio LVI desktop scanner. Technology: Outdated Zeiss robotic microscopy system including SL-4 (4-slide capacity) robotic microscope and MedMicrosocpy application (no longer vendor supported). Aperio LV1 4-slide capacity scanner with desktop console (Leica Biosystems, Vista, CA, USA). Design: A retrospective study was performed at UPMC Shadyside and Children’s Hospitals comparing deferral and diagnostic concordance rates for NTP when using Zeiss and Aperio LVI devices. A total of 169 neuropathology cases were reviewed in which intraoperative NTP was performed, including 86 cases (69 adults, 17 children) using the Zeiss system and 85 cases (68 adults, 17 children) using LVI. Concordance between intraoperative and final diagnoses was divided into 5 categories: (defer/non-diagnostic, lesional tissue identified, favor/compatible with diagnosis, correct category and exact diagnosis). Results: The deferral rate to permanent and ancillary studies when using the Zeiss system was 17.64% compared to 15.47% for LVI. Concordance rates are summarized in [Table 1]. Incorrect diagnoses were rendered in 4 cases (4.65%) using Zeiss and in 1 (1.17%) case when using LVI. Conclusion: Use of the newer LVI instrument for intraoperative NTP lowered the deferral rate and resulted in an improvement of our concordance rate. Despite the fact that neuropathologists at our institution have gained experience and familiarity with these systems over time, it is likely that technologic advances (e.g. faster scanning and viewing, enhanced picture quality and more user-friendly software) contributed to these improved intraoperative NTP results.
| Reference Interval Validation for Urine Protein Concentration Following a Vendor Notice|| |
J. Abel1, D. Bowman2, E. Pearlman3
1Department of Pathology and Laboratory Medicine, University of Tennessee Health Sciences Center, 2Department of Mathematical Sciences, University of Memphis, 3Department of Pathology, Veterans Affairs Medical Center, Memphis, TN, USA. E-mail: [email protected]
Content: In January, 2017 our chemistry equipment vendor (Ortho Clinical Diagnostics [OCD]; Rochester, NY, USA) informed us of manufacturing problems involving its UPC reagents with the possibility of a positive bias. This caused us to re-evaluate our use of the manufacturer’s suggested reference interval (MSRI) of 0-12 mg/dL. Technology: UPC was assayed with the Ortho-5600 analyzer. Calculations used Peak Fit and Table Curve-2D software (Systat, San Jose, CA). Design: We retrieved all UPC results (n=8691) from two calendar years preceding receipt of the vendor letter (2015/2016). RI validation was done in two steps. The data contained both “normal” and pathologic results and was modelled as a mixture of gamma distributions. Since UPC is expected to be minimal in healthy individuals, the left end of the RI was assumed = 0 mg/dL. Instrument results were integers and left censored at 5 mg/dL. Results less than 5 mg/dL. were simulated by sampling from a uniform (0, 5) distribution and rounding to the next highest integer. The right end of the first step interval was selected as the first x-intercept larger than the upper end of the MSRI (21 mg/dL) thus encompassing the MSRI. In the second step, all data points greater than 21 mg/dL. were discarded as “obviously” high and the remaining data was fit by splines and the curve was integrated from zero to an upper limit enclosing 95% of the AUC. This second step interval was taken as our putative RI. Results: The initial model contained 27 peaks with an r-squared of 0.93. Smoothing with a low pass filter reduced the number of peaks to 13 and increased the r-squared to 0.99. In either case the first step interval was 0-21 mg/dL and the suggested RI was 0-16.5 mg/dL. Conclusion: Our result was a RI that was 35% larger than the MSRI. This may be related to differences between our patient population and that used for derivation of the MSRI but may indicate that vendor manufacturing problems may have positively shifted the RI.
| Effect of Resampling on Validation of the Manufacturer’s Suggested Reference Interval for Urine Protein Concentration|| |
J. Abel1, D. Bowman2, E. Pearlman3
1epartment of Pathology, University of Tennessee Health Sciences Center, 2Department of Mathematical Sciences, University of Memphis, 3Department of Pathology, Veterans Affairs Medical Center, Memphis, TN, USA. E-mail: [email protected]
Context: We studied the effect of resampling on reference interval validation. Technology: UPC was assayed with the Ortho-5600 chemistry analyzer (OCD; Rochester, NY, USA). Calculations used Peak Fit and Table Curve-2D software (Systat, Chicago, IL). Resampling used the package “GSM” (R Foundation; Vienna, Austria). Design: Both original and resampled data were subject to the same procedure for RI validation. Estimation of the RI for UPC involved two steps. First, all patient UPC results (n=8691) from calendar years 2015/2016) were retrieved. This data contained both “normal” and pathologic results and was modelled as a mixture of gamma distributions. Since UPC is minimal in healthy individuals, the left end of the reference interval was assumed = 0 mg/dL. As assay data were integers and left censored at 5 mg/dL., results less than 5 mg/dL. were simulated by sampling from a uniform (0, 5) distribution and rounding the result to the next highest integer. The right end of the first step interval was the first x-intercept greater than the upper end of the MSRI thus encompassing the latter. In the second step, all data points greater than the right end of the initial interval were discarded as “obviously” high and the remaining data fit by splines. Our putative RI was defined by numerical integration that encompassed 95% of the AUC. Alternatively, these steps were performed with 10,000 resampled points. Results: The effect of the resampling was to delete more extreme values. The initial gamma fit resulted in 27 and 24 peaks with r-squared values of.93 and.934 respectively. The maximum and next highest UPCs were 2891 and 2203 mg/dL in the original data set and 1623 and 1589 mg/dL. in the resampled data. The length of the RI was decreased from 0-16.5 mg/dL to 0-15.5 mg/dL. Conclusion: Our proposed RI for UPC was narrowed by resampling, related to the increased number of points in the resampled data and deletion of more extreme values. The latter may ignore cases with important pathophysiologic information but can be desirable in sharpening reference intervals by sampling from the most frequently occurring values.
| External Laboratory Results: Provider Needs and Health Information Exchange Capabilities|| |
Joleen R. Borg1, Patrick C. Mathias1, David Chou1, William B. Lober2
1Department of Laboratory Medicine, University of Washington, 2Department of Biobehavioral Nursing and Health Informatics, University of Washington, Seattle, Washington, USA. E-mail: [email protected]
Content: Laboratory results are vital to clinical decision making. Testing may be performed at a laboratory that is not affiliated with the provider caring for the patient. These external results often are not electronically transferred into the provider’s electronic health record (EHR). Providers struggle to deliver timely, high quality, cost effective care when laboratory data are not readily available. Health information exchanges (HIEs) promised to improve availability of external laboratory results but in many situations this has not occurred. This research examines the needs of providers accessing external laboratory results and the capabilities of HIE solutions in a five state health care system, to describe gaps and recommendations to improve the availability of external results. Technology: Qualitative methods were used to assess needs of health care providers, and HIE capabilities from laboratory informatics and HIE experts in Providence Health and Services. NVivo (https://www.qsrinternational.com/nvivo/home) software was used to identify and code themes. Design: For this initial work we conducted semi-structured interviews with five providers from different specialties and with nine informatics experts. We performed a gap analysis to describe any gaps between needs and capabilities and to formulate a set of recommendations for strategies to incorporate external lab results in the EHR. Results: Providers received external laboratory results for 15% to 50% of their patients. Providers want external results in the EHR as structured data to support efficient review, trending, templated notes, and population health/quality metrics. Metadata can offer important information such as abnormal flags, and reference ranges. While HIEs are technically capable of transferring external lab results to the EHR, experts described many barriers. Technical barriers included poor data quality for patient matching, challenges with data normalization due to lack of standard application of structured terminologies, and data context. Non-technical barriers included fear of losing lab revenue, establishing contracts, and the uncertain future of HIEs. Conclusions: Providers need external laboratory results transferred into the EHR, however, many barriers stand in the way. Standard patient identification, improved coding of lab data, and standard approaches to HIE may lead to increased transfer of external lab results.
| Biomarker Colocalization Analysis of a Virtual 12-Plex using Discovery Chromogenic Dyes and Tissuealign™ Co-registration Software|| |
Benjamin Freiberg1, T. Regan Baird1
1Visiopharm Inc. E-mail: [email protected]
Content: The tumor micro-environment plays an important role in the diagnosis, prognosis and treatment regimens that patients receive. As such, understanding the specific changes in the immune-oncology (I/O) milieu presentation for tumors is key to developing novel therapeutics, new treatment regimens and ultimately increasing the survival and long-term prognosis for cancer patients. Imaging the tumor milieu has inherent problems as there are many different types of immune cells to identify. In order to increase the number of biomarkers that can be analyzed concurrently, we propose a combination of multi-chromogen dyes and software for co-registration of sequential serial sections. These co-registered images produce highly multiplexed virtual images (8-, 10-, 15-plex or more) in which the tumor microenvironment can then be interrogated. Technology:3 sequential 4um serial sections were labeled with Roche Ventana chromogenic assay kits designed for multiplexing in the tumor microenvironment. The purple label is identical in all 3 sequential serial sections that were subsequently imaged using an Aperio scanner. The resulting whole slide images were then co-registered using Visiopharm’s Tissuealign module and subject to localization analysis in each section. Design: Three tonsil tissue sections cut at 4-micron intervals were stained with Discovery Purple CD8 plus combinations of Discovery Teal, Discovery Yellow Ki67, and hematoxylin counterstain, then digitized using an Aperio scanner and co-registered using Visiopharm’s Tissuealign™ module. Results: The results of the study are described in the charts below. We observe no relationship between cells in sections that are not aligned (data not shown). However, data that are aligned with Tissuealign™ show high degrees of cellular correlation throughout the virtual stack of 3 serial sections. The occurrence of CD8+ cells in all three images in the virtual stack was very high (77%) if the cell was present in the middle section. As expected, cells that were identified at the top or bottom plane were identified to much less of a degree (36% top section and 44% bottom section throughout all three sections. Sampling error, measured by the presence of cells in the top section and bottom section was minimal (14%) and is most likely explained by the presence of a second cell appearing in the stack. Conclusion: We have demonstrated that it is practically feasible to work with high-plex study designs, using a combination of physical and virtual multiplexing. This approach requires the ability to achieve highly accurate/precise alignment of several serial sections, close to cell-to-cell alignment, even in the presence of non-linear tissue deformations across serial sections. The virtual multiplexing software developed by Visiopharm provided both a high level of precision and speed, which made it well suited as a tool for this type of study designs. The image analysis software, makes it efficient/feasible to analyze high (even hyper)-plexed datasets regardless of whether it is physical, virtual or hybrid multiplex designs.
| Virtual Case Seminars Using Whole Slide Digital Scans and Screencast Technology to Enhance Pathology Education|| |
Joseph S. Frye1, Mary Wong1, Stacey Kim1, Alberto Marchevsky1
1Department of Pathology, Cedars Sinai, Angeles, California, USA. E-mail: [email protected]
Content: Many residency programs are currently based on a subspecialized model where residents work intermittently with a relatively small number of surgical pathology cases during their 4 years of training. Slide seminars where residents are asked to diagnose “unknown” cases have traditionally supplemented exposure to challenging diagnostic problems. There has been little interest in developing virtual case seminars using digital tools. Screencasts are videos with digital recordings of computer screen output with interactive features. Technology: Aperio scanner (Leica Biosystems, Buffalo Grove, IL) was utilized to produce whole slide images. Images were exported, stored and managed with Aperio eSlide Manager (Leica Biosystems, Buffalo Grove, IL). To create screen casts, whole slide digital scans were viewed using Aperio ImageScope (Leica Biosystems, Buffalo Grove, IL; v18.104.22.16856). Screencasts were created using Camtasia 2.0 (TechSmith, Okemos, MI; v8.6) software. Design: Twenty virtual pulmonary pathology cases were prepared using whole slide digital scans utilizing Aperio scanner (Leica Biosystems, Buffalo Grove, IL) and Camtasia 2.0 (TechSmith, Okemos, MI) screencast software. They included: brief clinical history, video of chest CT scan or chest X-rays, video showing virtual slide at various magnifications, a pretest, video with a description of the virtual slide by an attending pathologist and a posttest. Results of quizzes were automatically emailed to an attending pathologist. Screencasts were stored in the internal departmental server and at screencast.com. Screencasts links were emailed to 12 residents and ancillary pathology staff who were asked to provide feedback to various questions on a 0-5 scale. Statistical analysis was performed using the paired t test. Results: The screencasts included digital multimedia container file format (MP4) ranging from 10 to 90 megabytes. They lasted up to 5 minutes unless viewers elected to stop the videos to replay portions representing virtual slides. They could be viewed with desktop or laptop computers using Windows and Mac operating systems, and mobile devices using Android and iOS platforms. Participant feedback was very favorable, showing average scores ranging from 4.38-4.75 to survey questions. Participants viewed the screencasts as a valuable and fun to use addition to their education. Nine of 12 participants favored screencasts when compared to previous pilot studies, due to its multi-platform functionality, namely android and iOS. Of those nine users, the average score ranged from 4.47-4.83. Using the paired t test, there was a significant difference in the pretest (48.5% ±31.2%) and posttest (87.0% ± 21.6%) results (P<0.0001). Participants favored viewing the screencasts on computers over mobile devices and preferred the length to be up to 5 minutes long. Conclusion: Virtual case seminars using whole slide digital scans and screencasts can be easily produced by pathology staff and housestaff without the need for specialized technical support. Using software that has multi-platform functionality was important to users, especially in an era of increased handheld and tablet use. Screencasts were well received by our housestaff and provide a useful tool to enhance surgical pathology education in the digital era. Further studies with larger cohorts and external controls are needed to ascertain its true efficacy in education.
| Next Generation Sequencing Identifies Mutational Differences between Primary and Metastatic Colorectal Carcinoma: Therapeutic Implications for Specimen Selection|| |
Joseph S. Frye1, Eric Vail1, Andy Pao1, Jean R. Lopategui1
1Cedars Sinai Medical Center, Los Angeles, CA, USA. E-mail: [email protected]
Content: Colorectal carcinoma (CRC) is a common and aggressive malignancy for which standard chemotherapy is of limited benefit in the metastatic setting. While a few studies have suggested that KRAS, NRAS, and BRAF mutations are highly concordant in paired primary and metastatic CRC, others have shown considerable heterogeneity. It has not been previously elucidated whether this is due to tumor heterogeneity or clonal evolution. We aim to explore the mutational landscape in paired primary and metastatic colorectal tumors. Technology: 50-gene AmpliSeq cancer panel v2 on an Ion Torrent PGM sequencing instrument (Thermo Fisher, Inc., Canoga Park, CA). Sunquest PowerPath (Sunquest Information Systems, Tucson, AZ; v10.0.1.42) laboratory information system. Epic (Verona, WI) electronic medical record. Design: A total of 5 patients were selected with mutational discordance between the primary tumor and metastasis. Their primary tumor and metastasis were multiply sampled to obtain 21 formalin-fixed paraffin-embedded tumors for sequencing. The patient’s tumors included 5 primary tumors, 1 regional lymph node metastasis and 5 distant organ metastases. Samples were sequenced with the 50-gene AmpliSeq cancer panel v2 on an Ion Torrent PGM sequencing instrument (Thermo Fisher, Inc.), targeting 2855 hotspot mutations. Results: Four of five patients had partially discordant results (80.0%). Of these, there were multiple point mutations detected within the same gene of both the primary and metastases. One of the five patients had completely discordant results with a deletion in the primary and gain of the same tumor suppressor in the metastasis. In the six cases with discordant variants, clinically relevant mutations in APC, NRAS, and PIK3CA were either gained (n=3) or lost (n=3) in distant metastases. Across all patients, 42 mutations were identified involving 12 genes: TP53, APC, KRAS, NRAS, PIK3CA, FLT3, FGFR3, STK11, CTNNB1, JAK3, KDR and CDKN2A. Thirty-three mutations (78%) were concordant between primary tumors and distant metastases. Conclusion: Evidence from our study shows mutational discordance in paired primary and metastatic tumors making it unclear whether primary CRCs or their metastases should be selected for NGS. Our data suggests that while some mutations are present in both the primary and metastasis there is significant mutational tumor heterogeneity amongst the samples. Perhaps more clinically relevant is that those samples that exhibited clonality had gains and loss in metastases suggestive of clonal evolution. A tentative interpretation is that treatment selection is ideally based on multiplex sampling and genetic profiling of the most current or recent tissue samples (including metastases) rather than chronologically distant samples.
| Incorporation of SNOMED CT and International Classification of Diseases for Oncology-3 Codes for Topographical and Morphological Maps in College of American Pathologists Electronic Cancer Checklists|| |
Keren I. Hulkower1, Eric M. Daley1, Gemma Lee2, Richard Moldwin1, Jaleh Mirza1
1College of American Pathologists, Northfield, IL, USA, 2Cancer Care Ontario, Toronto, Ontario, Canada. E-mail: [email protected]
Content: The College of American Pathologists (CAP) and Cancer Care Ontario collaborated to create mapping links between CAP electronic Cancer Checklist items and SNOMED CT and International Classification of Diseases for Oncology (ICD-O) codes. Maps provide links between particular concepts or terms in one system and concepts or terms in another system. Mapping allows interoperability among international classification and code systems. ICD-O and CT coding have been used historically by cancer registries, for coding the topography and the morphology of the tumor taken from pathology reports. Technology: SNOMED CT and ICD-O-3 codes for Histologic Type, Tumor Site and Primary Tumor Site answer items were sourced from the July 2017 SNOMED CT International Release, WHO Classification of Tumours, and the International Classification of Diseases for Oncology Online website. Additional fields were added to our Template Editor tool to capture ICD-O-3 Match Type, WHO Full Term, and SNOMED CT Match Type. Design: SNOMED CT and ICD-O-3 codes and WHO Full terms were embedded in the XML files of the CAP electronic Cancer Checklists as metadata associated to each unique morphologic type or topographic site. Results: [Table 1] shows the number of codes mapped in each category across the CAP eCC Resection and Biopsy templates.Match types were provided to assess how closely aligned the eCC histologic type and anatomic site list items were to the SNOMED CT and ICD-O-3 codes. These included: Exact Match, Semantic Match, and Item More Specific than Code. Conclusions: Embedding the ICD-O-3 and SNOMED CT codes within the CAP eCC templates has improved the efficiency and accuracy of coding and enabled the creation of standardized epidemiologic data.
|Table 1: The number of codes mapped in each category across the College of American Pathologists Electronic Cancer Checklists resection and biopsy templates|
Click here to view
| Real Time Student Assessment in Pathology Teaching Laboratories|| |
1Pathology, Medical School, University of Connecticut, Connecticut, USA. E-mail: [email protected]
Content: Early identification of students at risk in pathology teaching laboratories is essential in providing rapid remediation so that students have sound foundation for subsequent laboratory sessions. A web based student assessment system was developed to rapidly identify any students at risk in second year pathology teaching laboratories. Technology: A web services solution stack consisting of Apache web server, MySQL database and PHP (hereinafter, web stack) was used to develop a real time student assessment system. Initially Uniform Server (http://uniformserver.com) and latter Bitnami (http://bitnami.com/stacks) web stacks were employed. The Uniform Server, a Microsoft Windows operating system web stack, can run from a thumb drive or any USB storage device and requires no installation. Thus the system can be used on any Windows computer with a static IP address. The Bitnami web stack can run on Windows, OS X or Linux environments as well as virtualized environments and popular cloud platforms. Design: Using the notation feature of a virtual microscope laboratory preceptor indicate various pathological processes/cells. The notations (arrows, circles, etc.) are seen by students on their laptop computer virtual microscopes. In addition to viewing the notated virtual microscope field, students connect to the assessment systems that displays a list of possible answers. Students select the appropriate answer which is then database stored. Preceptor view student answers as they are submitted. Results: When a significant number of students do not correctly identify a process/cell the teaching objective is reviewed for the entire laboratory, when one or a few students have difficulty, these students can be flagged for individual remediation. Conclusion: The real time pathology laboratory student assessment system helps identifying students at risk so that remediation can be instituted immediately rather than waiting for med-term or final exams. By identifying at risk student’s early remediation can be done so that students do not fall behind.
| Update from Latin America in the Era of Digital Pathology|| |
Andres Mosquera-Zamudio1, Javier A. Arias-Stella1, Matthew Hanna2, Emilio Madrigal3, Carmen Lis Ugalde4, Jorge Ugalde4, Osvaldo Spinelli5, Mauro Saieg6, Rafael Parra-Medina7, Paula Toro8, Gonzalo de Toro9,10, Juan Hernandez-Prera11, Dra. Vilma Rivas Lemus12, David Jaramillo13, Liron Pantanowitz14
1Clínica Universitaria Colombia-Colsanitas, Department of Pathology, Pathology expert professor, Fundacion Universitaria Sanitas Bogota, Colombia, 2Memorial Sloan Kettering Cancer Center, Department of Pathology, New York, 3Emory University/Massachusetts General Hospital Atlanta, GA, Boston, MA, 4Labpato.com, Cuenca – Ecuador, 5Facultad de Ciencias Médicas, Universidad Nacional de La Plata, La Plata, Argentina, 6Santa Casa Medical School, Sao Paulo, Brazil, 7Fundación Universitaria Ciencias de la Salu, Bogotá, Colombia, 8Universidad Nacional de Colombia, Bogota, Colombia, 9Hospital del Puerto Montt, Chile, 10Universidad Austral sede Puerto Montt, Puerto Montt, Chile, 11Moffitt Cancer Center, Tampa, USA, 12Hospital San Juan de Dios de Santa Ana, El Salvador, CA, 13Fundacion Universitaria Sanitas, Bogota, Colombia, 14UPMC Shadyside, Pittsburgh, PA, USA. E-mail: [email protected]
Content: Digital Pathology (DP) endeavors have been conducted in Latin America (LATAM) for many years. However, there has been no comprehensive review of all of these efforts. Therefore, the aim of this study was to do undertake an extensive review of DP activity in all LATAM countries. Technology: Email correspondence, videoconferences and web searches were used to collect data. Design: Leading authorities of DP and telemedicine were contacted in 16 LATAM countries, including those who originated from these countries but were now living abroad. Responders were asked to submit a manuscript summarizing the use of DP (contributions, appraisal, limitations) in their country since its inception. For countries without a representative, an exhaustive search was made online and of publications for DP content. Results: The study was challenged by finding pathologists in LATAM with adequate knowledge and experience about DP. Ten countries participated, which covered 80% of LATAM territory [Figure 1]. 6 countries successfully use whole slide imaging mostly for teleconsultation and academic purposes. Others rely on static images and/or employ messenger apps on smartphones for teleconsultation. One country was actively involved with image analysis and creating DP software and hardware. One country because of economic restrictions has no experience in DP at all. Conclusion: In LATAM technological, infrastructure and economic limitations interfere with the desired progress in DP. Despite these barriers, important progress has been made in at least half of these countries. The major activity encountered is the application of telepathology using both old (e.g., static images) and new (e.g., whole slide imaging) modes of practice. We encourage LATAM pathologists to embrace DP, prepare themselves to take greater advantage of novel tools (e.g. image analysis), and share their experiences (e.g., scientific publications) with the rest of the world. We consider the creation of a DP LATAM network to be important. This will help those countries with limited or no experience in DP and offer timely and expert anatomic pathology services to those territories that are difficult to access.
|Figure 1: Latin America countries reporting digital pathology activity (green indicates active response)|
Click here to view
| Circular Visualization of Digitized Pathology Data and Derived Tissue Morphometrics Offers a New Approach to Presentation of Extensive Data Sets|| |
David G. Nohle1, Weiqiang J. Zhao1, Leona W. Ayers1
1Department of Pathology, The Ohio State University, Columbus, OH, USA. E-mail: [email protected]
Content: We digitized diagnostic pathology biomarker tissue test slides and generated morphometric data commonly used to characterize tumors for diagnosis and prognosis. Recorded clinical data included patient survival. Tumor cell test results were displayed separately from similar reactivity of background cells. Data included tissue morphometric values for cell or nuclear size and biochemical results for CD20, KI67, cyclinD1, CD5, Sox11, CD31, PRMT5, CD10, P53, CD34 as well as genomic data from a diverse study group of 31 mantle cell lymphomas. Technology: Tissue Studio (Definiens AG, Munich, Germany), Circos software (http://circos.ca/). Design: Our circular diagrams feature rings to represent the results for each biomarker test type or patient treatment and survival data. Rings can contain text, glyphs or data represented by various plots (two dimensional tracks): line, scatter, histogram, heat map and combinations. Each patient’s test results and clinical data are represented in a wedge that contains one patient’s portion of all rings. In some rings, a cell (ring-wedge intersection) has the percent positive produced in biomarker tests displayed as amplitude while color indicates staining intensity. An adjacent ring displays background reaction positivity. Based on selected test results, patient wedges sort into subtype wedge groups. Data for mantle cell lymphomas were visualized using these constructs to identify subtypes. Results: Our patient wedges were grouped by increasing numerical values (i.e. cell nuclear size, cell proliferation index, and immunohistochemistry intensity). Static biomarkers and genomic, epigenetic and genetic results are each represented by a wedge ring result for tumor types and genetic groupings or by patient survival indicators (figure). Analyses that do and do not correlate with clinical predictions can be immediately recognized. Image analysis algorithms group lymphoma variants within a heterogeneous study population. We have developed a technique that can be used to display various pathology data. Conclusions: Circular data visualization can incorporate all quantitative and qualitative tissue results of a diagnostic entity using various ring representations. Used for a mantle cell lymphoma study, such visualizations highlighted correlations and progressions in expression and intensity of expression. Our constructs can extend to handle large numbers of cases/tests as are increasingly available with digital Pathology.
| Development of Pltvxm, A Clinical Decision Support Tool for Computer-Aided Selection of Platelet Units for Platelet Transfusion-Refractory Patients|| |
Andrew P. Norgan1, Justin E. Juskewitch1
1Department of Laboratory Medicine and Pathology, Mayo Clinic, Rochester, MN, USA. E-mail: [email protected]
Content: Immune-mediated platelet transfusion–refractoriness, or failure to appropriately increment platelet count after platelet transfusion due to anti-human leukocyte antigen (HLA) antibodies, presents a significant challenge to supporting affected patients. Immune-mediated refractoriness is addressed through specific antibody testing and precise platelet unit selection. Optimization of this selection process can result in fewer overall transfusions, thus conferring institutional resource and cost savings. Such optimization may be difficult to achieve manually, when large platelet inventories (or donor pools) need to be matched against complex and evolving antibody profiles. The aim of this work was to develop pltVXM, a computer-aided decision support tool that enables consistent selection of optimal platelet units for immune-mediated platelet transfusion-refractory patients. Technology: R programming language (https://www.r-project.org/), Shiny Server Pro (RStudio Inc., Boston, MA, USA), Tableau (Tableau Software, Inc., Seattle, WA, USA), SafeTrace (Haemonetics Corporation, Braintree, MA, USA). Design: To facilitate platelet unit selection, information from several clinical data sources has to be retrieved and integrated. Platelet inventory information was obtain using a Tableau workbook to query the SafeTrace inventory database when needed. The workbook is populated with available platelet units, and their associated metadata (blood type, expiration data, HLA type). The pltVXM application then pulls in the inventory information and combines it with patient specific data (blood and HLA types, anti-HLA antibody results) retrieved from a clinical data warehouse using a web-based application programming interface. The patient’s antibody profile algorithmically cross-matched against the available platelet inventory, with suggested compatible units highlighted. The user can refine unit selection by filtering for antibody strength, blood type, or expiration date, and automatically generate a patient-specific platelet reservation request. The system also allows for a similar matching process using the platelet donor pool at our institution to facilitate specific donor recruitment. Results: In initial testing, platelet unit selection times for complex patients (>4 anti-HLA antibodies) were reduced from 10-30 minutes to <5 minutes and created actionable platelet reservation requests without errors. Conclusions: Using an algorithm-driven approach, pltVXM allows for consistent and timely optimal platelet support to be selected for platelet transfusion-refractory patients, regardless of the user’s expertise level.
| Effect of Dedicated Cytology Protocol Setting for Scanning ThinPrep Slides|| |
Liron Pantanowitz1, Jon Duboy2, Jennifer Picarsic1, Jeffrey Fine1, Douglas J. Hartman1
1Department of Pathology, University of Pittsburgh Medical Center, 2Divison of Information Services, University of Pittsburgh Medical Center, Pittsburgh, Pennsylvania, USA. E-mail: [email protected]
Content: Digitizing cytology glass slides is more challenging because of the need to focus on 3D cellular material. As a result, default settings for single focal plane image acquisition on whole slide scanners may not generate good quality digital slides. Our aim was to determine the consequence of routinely scanning ThinPrep slides using a dedicated cytology protocol. Technology: Aperio AT2 scanner (Leica Biosystems, Vista, CA, USA) with console version 22.214.171.124. Cytology protocol that restricts the scan area to the ThinPrep circle, automatically drops shotgun focus points over the slide [Figure 1]a, uses the unsharp mask setting, and ignores individual failed focus points. Design: 30 ThinPrep glass slides with varied cytologic material were scanned at 40x on an AT2 scanner using the default setting (with few autoselected focus points on the entire slide) and dedicated cytology protocol setting without and with 3 Z-planes. Scan times, file sizes and image quality for these scan settings were compared. Results: [Table 1] summarizes the differences between scans. Scanning slides with the cytology protocol takes only slightly longer and generates digital slides that are 3-4x larger. Z-stacking increases file size equally with both scan settings. Image quality of cytologic material is better when slides are scanned with the cytology protocol [Figure 1]b; left with cytology protocol, right using default setting].Conclusion: Scanning ThinPrep slides using a dedicated cytology protocol produces digital slides with better focus and image quality. However, whilst the time to digitize slides with this setting does not take that much longer it does generate image files that are 3-4x larger compared to scans employing the default setting.
|Figure 1: (a) Cytology protocol that restricts the scan area to the ThinPrep circle, automatically drops shotgun focus points over the slide; (b) Left with cytology protocol, right using default setting|
Click here to view
| Evaluation of Quantitative Phase Imaging Compared to Urovysion for Urine Cytology|| |
Dinesh Pradhan1, Hoa V. Pham2, Jalil Nasibli2, Yang Liu2, Liron Pantanowitz1
1Department of Pathology, University of Pittsburgh Medical Center, 2Department of Medicine and Bioengineering, Biomedical Optical Imaging Laboratory, University of Pittsburgh, Pittsburgh, Pennsylvania, USA. E-mail: [email protected]
Technology: A custom-built QPI system was based on diffraction phase microscopy, where a transmission grating was placed at the image plane to generate multiple copies of the emerging field. The filtered zeroth-diffraction order interferes with the entire first-order diffracted beam on the camera to create an interferogram used to reconstruct the quantitative phase image. Image acquisition was performed on unstained slides by automatically scanning the entire slide. Thereafter, the slide was stained (Papanicolaou stain) and re-imaged. Design: Urine specimens (voided and instrumented) collected from 102 patients (mean age 75.4 years; range 35-93) and fixed in 95% ethanol were submitted for cytologic evaluation and Urovysion FISH testing. QPI was performed on an additional unstained ThinPrep slide prepared from each patient. Based on the quantitative phase map, nuclear and cell dry mass, nuclear entropy and nucleus-to-cell dry mass ratio were calculated for several hundred cells in each patient. Findings were correlated with cytology interpretation, Urovysion results, and available cytohistologic follow-up diagnoses. Results: Nuclear mass ratio and nuclear entropy of urothelial cells showed significant difference between negative and positive cytology groups. For patients with atypical or suspicious cytology results (n=44), nuclear mass ratio and entropy in cases with a follow-up diagnosis of malignancy (n=10) were significantly higher [Figure 1] than those patients without subsequent malignancy on follow-up (n=12). However, QPI parameters were not highly concordant with Urovysion results [Figure 2]. Conclusion: These data confirm that QPI is a novel tool with potential to improve the diagnostic accuracy of urine cytology. Whilst the parameters derived by QPI appear to correlate well with cytomorphology, they did not correlate with the chromosomal abnormalities detected by the UroVysion FISH assay. This suggests that there may be other genetic changes important in bladder cancer pathogenesis.
| Using Image-Assisted Manual Counting as a Quality Assurance Tool in Pancreatic Neuroendocrine Tumors|| |
Andrey Prilutskiy1, Qing Zhao1
1Department of Pathology, Boston Medical College, Boston, MA, USA. E-mail: [email protected]
Content: The three-tier histologic grading system for well-differentiated pancreatic neuroendocrine tumors has been established based on the Ki67 index: low grade (G1, <3%), intermediate grade (G2, 3-20%) and high grade (G3, >20%). There is still no consensus on how to quantify positive and negative cells on Ki67-immunostained sections. The most popular method of estimation (“eyeballing”) is neither reliable nor reproducible. The quantitative imaging-based methods are however requiring significant resources to implement and often face practical resistance. In the 2018 College of American Pathologists (CAP) guidelines, there is a new recommendation on using manual cell counting on a printed photograph. Technology: SPOT Insight CCD camera mounted on Olympus BH-2 microscope and SPOT 5.2 software were used for pictures taken in hot-spots on Ki67-stained slides, under 400x magnification. Pictures printed on HP LaserJet M553 color printer using letter-sized plain paper. Manual cell counting was performed on printed images by circling positive cells and crossing out negative cells. Design: Total of 14 cases of pancreatic neuroendocrine tumors were collected in our pathology department from 2013 to 2018. CAP recommendation for number of cells to be counted is 500 to 2000. We counted between 530 and 1720 cells for each case, with an average number of 833 cells. All results of manual counting were compared to the reported proliferation index (estimated by “eyeballing”). Results: Based on the manual count, 3 out of 14 cases that were reported as Grade 1 were upgraded to Grade 2, and 1 case that was reported as Grade 2 was downgraded to Grade 1 [Table 1]. All 4 cases had a lower proliferative index close to the 3% cutoff. Conclusion: Using the described method for assessing Ki67 index in pancreatic neuroendocrine tumors is very easy to implement and does not require additional personnel, instruments or significant workflow changes. We found this method to be a powerful quality assurance tool and also something that can be adopted quickly in daily practice. It can become a successful bridge between imprecise estimation and more advanced automatic image-based quantitative methods in the future.
|Table 1: Results of manual cell counting in 14 pancreatic neuroendocrine tumor cases|
Click here to view
| Pap smear More Details Based Cervical Cancer Screening"> Low Cost Imaging System for Pap smear Based Cervical Cancer Screening|| |
Srikanth Ragothaman1, Sridharakumar Narasimhan1, Basavaraj M. Gurappa1, Rajan Dewar2
1Department of Chemical Engineering, Indian Institute of Technology Madras, Chennai, Tamil Nadu, India, 2Department of Pathology, University of Michigan, Michigan, USA. E-mail: [email protected]
Content: Cervical cancer is one of the leading causes of cancer death among women from Low and Middle Income countries (LMIC). Screening at early stages using the popular Pap smear test has been demonstrated to reduce fatalities significantly in the western countries. However, availability of conventional cytology in LMICs is challenging due to cost and manpower constraints. Thus, there is a need to develop suitable low cost automatic screening systems that are robust enough as a primary screen of cytology smears. Technology: Developed a prototype of an imaging system with low cost portable digital microscope with single magnification. With magnification of nearly 400x, pap smear cells are imaged and analyzed. The slide is placed on motorized linear stage to perform whole slide scanning. Various fields of view images are acquired and stored in database. Image acquisition is followed by image analysis which involves algorithm to spot potentially abnormal cells in given field of view image. Image analysis involves segmentation of cell images into 3 regions i.e. Nucleus, cytoplasm and Background. The image segmentation is performed using Gaussian mixture model based clustering to estimate the nuclear size. Design: The data set consisted of 246 images with 106 normal (negative) and 140 abnormal (positives) which includes categories such as atypical squamus of undetermined significance (ASCUS), low grade squamus intra epithelial lesion (LSIL) and high grade squamus intra epithelial lesion (HSIL). The prediction was done based on specific nucleus size above which the image is flagged as positive. Results: The experimental results yielded sensitivity of 0.75 and specificity of 0.40 which indicates the device feasibility to act as screening test device. Conclusion: The device performance can be increased by tuning the nuclear size parameters and validating on huge data set. Robust features apart from nuclear size has to be identified and included as features in classifier to increase the performance of this imaging device.
| Predictive Values of Common Laboratory Tests for Data Driven Patient Management|| |
Ziemba C. Yonah1, Kalpana S. Reddy1, Lomsadze Liya1, Haghi Nina1, Scott Duong1, Feifan Chen1
1Department of Pathology and Laboratory Medicine, Hofstra Northwell School of Medicine, East Garden City, New York, USA. E-mail: [email protected]
Content: Positive predictive values (PPV) and negative predictive values (NPV) of laboratory tests are crucial in forming data-driven clinical decisions, yet these values are usually not available to clinicians. This is because they vary with the disease prevalence in each region, making it nearly impossible for a test manufacturer to publish standard predictive values. In this absence, clinical response to test results may become guided by anecdotal experience, clinician habit, and untested assumptions. Herein, we demonstrate a simple method that can be used by any hospital laboratory to accurately assess predictive values for their inpatient population. Technology: Microsoft Excel. Design: Screening and confirmatory test were defined as outlined in [Table 1]. We included all screening tests that were followed by a confirmatory test during the same hospital stay in 2017 at a medium size hospital in our health system. Results: As displayed in Table 1, the results are not consistent with the natural assumption that common tests have high predictive values. Rather, the PPVs were considerably lower that the NPVs. Most of the NPVs were 90% or above, while most of the PPVs were 50% or below, reflecting the broad thresholds that are commonly chosen for screening tests. Conclusions: Measurement of PPV and NPV for individual hospitals can be accomplished easily and accurately at any hospital laboratory. This will produce accurate predictive values, which will be valuable to hospitals in several ways. Firstly, it can guide patient management when a patient with a positive screen is waiting for results of confirmatory testing. For example, if urine analysis shows positive leukocyte esterase, the decision of whether to begin antibiotics while waiting for urine culture results should be guided by the PPV. In addition, this will identify tests that have very low predictive values, and thus avoid unnecessary workups that cause patient risk, discomfort and hospital expense. Finally, in the era of computer-assisted clinical management, computer algorithms will use these predictive values to determine when further follow up testing is indicated, and which tests are likely to be most helpful.
[Figure 1], [Figure 2], [Figure 3], [Figure 4], [Figure 5], [Figure 6], [Figure 7], [Figure 8], [Figure 9], [Figure 10], [Figure 11], [Figure 12], [Figure 13], [Figure 14], [Figure 15], [Figure 16], [Figure 17], [Figure 18], [Figure 19], [Figure 20], [Figure 21], [Figure 22], [Figure 23], [Figure 24], [Figure 25], [Figure 26]
[Table 1], [Table 2], [Table 3], [Table 4], [Table 5], [Table 6], [Table 7], [Table 8], [Table 9], [Table 10]