Journal of Pathology Informatics

RESEARCH ARTICLE
Year
: 2019  |  Volume : 10  |  Issue : 1  |  Page : 39-

Whole-slide image focus quality: Automatic assessment and impact on ai cancer detection


Timo Kohlberger1, Yun Liu1, Melissa Moran1, Po-Hsuan Cameron Chen1, Trissia Brown2, Jason D Hipp3, Craig H Mermel1, Martin C Stumpe4,  
1 Google Health, Palo Alto, CA, USA
2 Work done at Google Health via Advanced Clinical, Deerfield, IL, USA
3 Google Health, Palo Alto, CA; Current Affiliation: AstraZeneca, Gaithersburg, MD, USA
4 Google Health, Palo Alto, CA; Current Affiliation: Tempus Labs, Chicago, IL, USA

Correspondence Address:
Dr. Timo Kohlberger
Google LLC, 1600 Amphitheatre Parkway, Mountain View, CA
USA

Abstract

Background: Digital pathology enables remote access or consults and powerful image analysis algorithms. However, the slide digitization process can create artifacts such as out-of-focus (OOF). OOF is often only detected on careful review, potentially causing rescanning, and workflow delays. Although scan time operator screening for whole-slide OOF is feasible, manual screening for OOF affecting only parts of a slide is impractical. Methods: We developed a convolutional neural network (ConvFocus) to exhaustively localize and quantify the severity of OOF regions on digitized slides. ConvFocus was developed using our refined semi-synthetic OOF data generation process and evaluated using seven slides spanning three different tissue and three different stain types, each of which were digitized using two different whole-slide scanner models ConvFocus's predictions were compared with pathologist-annotated focus quality grades across 514 distinct regions representing 37,700 35 μm × 35 μm image patches, and 21 digitized “z-stack” WSIs that contain known OOF patterns. Results: When compared to pathologist-graded focus quality, ConvFocus achieved Spearman rank coefficients of 0.81 and 0.94 on two scanners and reproduced the expected OOF patterns from z-stack scanning. We also evaluated the impact of OOF on the accuracy of a state-of-the-art metastatic breast cancer detector and saw a consistent decrease in performance with increasing OOF. Conclusions: Comprehensive whole-slide OOF categorization could enable rescans before pathologist review, potentially reducing the impact of digitization focus issues on the clinical workflow. We show that the algorithm trained on our semi-synthetic OOF data generalizes well to real OOF regions across tissue types, stains, and scanners. Finally, quantitative OOF maps can flag regions that might otherwise be misclassified by image analysis algorithms, preventing OOF-induced errors.



How to cite this article:
Kohlberger T, Liu Y, Moran M, Chen PHC, Brown T, Hipp JD, Mermel CH, Stumpe MC. Whole-slide image focus quality: Automatic assessment and impact on ai cancer detection.J Pathol Inform 2019;10:39-39


How to cite this URL:
Kohlberger T, Liu Y, Moran M, Chen PHC, Brown T, Hipp JD, Mermel CH, Stumpe MC. Whole-slide image focus quality: Automatic assessment and impact on ai cancer detection. J Pathol Inform [serial online] 2019 [cited 2020 Aug 14 ];10:39-39
Available from: http://www.jpathinformatics.org/text.asp?2019/10/1/39/272775


Full Text



 Introduction



Digital pathology is advancing into clinical workflows,[1],[2] enabled by the recent regulatory approval of the first whole-slide image (WSI) scanner for primary diagnosis in the U.S.,[3] as well as wider availability of cheaper storage and large technical infrastructure to manage gigapixel-sized image files. Digitization has several compelling use cases: archiving, telepathology for remote consults or diagnosis, teaching, and increasingly, facilitation of powerful image analysis algorithms.[4],[5] The process of digitization, however, can add artifacts to the imaging process, including color or contrast variations and out-of-focus (OOF) areas. These artifacts, particularly OOF areas, can hinder the rendering of accurate diagnoses by pathologists,[1],[6] or impact the accuracy of automated image analysis.[7]

We here distinguish between three general categories of OOF in decreasing severity: global OOF that affects the entire slide, regional OOF that affects larger tissue patches, and local OOF that affects individual cells or subcellular structures [Figure 1]. Global OOF can be caused by the whole-slide scanner erroneously focusing on coverslip debris. Regional OOF can be particularly prominent in specific tissue types (such as fat), or affect an entire section of the slide (e.g., “scan lane”). Finally, local OOF is common because typical tissue sections of approximately 5 μ are thicker than the depth of field of standard digital pathology scanners at high magnification. While a potential solution is the use of multiple focal depths (“z-stacking”), such techniques currently increase scan time and file sizes to an impractical extent for routine use.{Figure 1}

The different OOF categories have different impacts on the clinical workflow. Pathologists or histotechnicians may flag images with global OOF as low quality and order a rescan, which potentially results in reporting and workflow delays. Regional and local OOF can be much more difficult to detect consistently, particularly when reviewing a slide at low magnification. Rescans may only be requested after the pathologist has spent significant time reviewing other areas in the slide, before proceeding to evaluate specific areas of concern at higher magnification. Importantly, a technician can prescreen all digital slides for global OOF, but that would be impractical for manual review for smaller OOF regions. OOF image artifacts can have even more severe consequences in automated image analysis by directly impacting detection and classification. Some studies found that systematic errors can be attributed to suboptimal focus quality, such as OOF germinal centers being mistaken for tumor metastases by an algorithm.[7]

While every WSI scanner has built-in focus evaluation that can be used for automatic rescans of the affected regions or for quality reporting,[8] there are several shortcomings in existing methods: (1) despite built-in auto-focus control, most currently available WSI scanners still produce scans with focus issues,[9],[10] (2) focus evaluation methods across scanners are different, inhibiting comparison across devices, (3) focus metrics are not typically exported to the user with detail regarding the spatial distribution of focus quality, and (4) evaluation does not take the clinical relevance of the focus quality into account. For example, cytology diagnoses that more strongly rely on subcellular details require a higher focus quality than diagnoses that are based primarily on tissue architectural patterns, such as prostatic adenocarcinoma Gleason grading.

A sizable body of work exists for automated focus quality assessment in microscopy images using manually engineered image features.[11],[12],[13],[14],[15] Deep convolutional neural network approaches, on the other hand, learn the discriminative features and yield higher accuracies,[7],[16],[17],[18],[19],[20] for disease diagnosis and tissue classification for image quality classification in particular.[21],[22],[23] The main challenge, particularly for the neural networks approaches, is the availability of training data to help generalization across a large variety of tissue morphologies and stain properties. To help generalization, previous works have generated simulated OOF examples using real in-focus images through synthetic Gaussian blurring,[23] and additional perturbations such as brightness perturbations and artificial sensor noise.[22]

This work further improves the synthetic data generation approach by more closely mimicking the image acquisition process of real OOF artifacts. These improvements allowed us to build a robust, highly sensitive and fine-grained OOF predictor (ConvFocus) that distinguishes a large spectrum of focus degrees across different tissue and stain types. Our approach provides a generally applicable metric that is highly concordant with manually annotated focus quality and provides information about focus quality across every region in a WSI. Moreover, we quantify the sensitivity of a leading tumor detection algorithm to focus quality, which has not been shown in prior related work.

 Methods



Overview

Given a gigapixel-sized image of a whole pathology slide, our goal was to automatically detect and grade OOF regions at an accuracy level matching that of a pathologist across tissue, biopsy and stain types. To that end, we employed a convolutional neural network approach, which has shown superior performance over more classical machine learning approaches using hand-crafted features.[23] We term our approach ConvFocus [Figure 2] and [Figure 3].{Figure 2}{Figure 3}

Training data

Developing accurate neural networks requires large volumes of training data. Unfortunately, accurate delineation and grading of OOF areas in WSIs is a highly time-intensive task. For example, our pathologist spent 8–16 h per slide to find, delineate, and grade OOF regions (see the “Test Data” section for more details). Consequently, generating a sufficiently large pool of OOF training examples purely by manual annotation is impractical. Instead, we first manually assessed whether patches were completely in-focus or not. For each patch that was labeled “in-focus” by three independent raters, we then synthetically generated multiple additional OOF patches with varying but known degrees of OOF. The real in-focus image patches and the synthetically “de-focused” versions served as training examples. We next describe this process in detail.

Whole-slide scans used for training

26,526 different slides spanning a large variety of tissue, biopsy and stain types were used to train and validate our convolutional neural network. Of these, 8,135 slides had surgical sectioning site information available in a structured form [Table 1]. 26,099 slides were scanned with an Aperio AT2 (Leica Biosystems, Germany) at × 40 (pixel size: 0.251 μm × 0.251 μm), using a semi-automatic scan mode, and 427 slides were scanned using a Hamamatsu NanoZoomer XR (Hamamatsu Photonics, Japan) also at × 40 (pixel size: 0.227 μm × 0.227 μm).{Table 1}

Manual annotation of in-focus image patches

From each of the aforementioned WSIs, 300 × 300 pixel-sized patches at × 40 were sampled from tissue-containing regions, and subsequently evaluated to be either in-focus or OOF by three independent human raters. First, a “tissue mask” was generated by applying the criterion 0.0 < Y ≤ 0.8. Y was defined as the average intensity and computed from the RGB values: Y = 0.212×R + 0.715×G + 0.072×B. Next, 300 × 300 pixel-sized patches were randomly sampled from tissue-containing areas. Of those, the top eight (for the AT2) and top 24 (for the NanoZoomer XR) patches with the lowest average luma value were picked for the manual assessment of whether the patch was in-focus. More patches were selected from NanoZoomer XR images because fewer images from that scanner were available in our database, thus this was necessary to achieve a more balanced training set with respect to scanner. This procedure was designed to select the densest tissue parts, which in most cases provided the highest density of sharp image gradients and best exposed OOF issues.

These “densest” patches were then sent to trained nonpathologist raters for independent evaluation. Raters applied one of the following labels to each patch: “in-focus” if at least 75% of tissue area was in-focus, “out-focus” if at least 75% was OOF to any degree, and “undecided” otherwise. We selected approximately 166,000 patches that all three raters labeled as “in-focus” (i.e., consensus rating) to create semi-synthetic training data.

Semi-synthetic training data generation

For each rater-determined in-focus patch (which is labeled with OOF class 0), we synthetically generated 29 additional pseudo-OOF versions with increasing blur magnitude, labeled as OOF classes 1–29. Classes 1–28 represented fine-grained, exponentially increasing OOF magnitudes, and class 29 covered a large range of strong blurring artifacts. Early empirical results indicated that linearly increasing blurring magnitudes for classes 1–28 yielded inferior detection results.

We explored two kinds of blur filters: either a convolution with a 2D Gaussian kernel of size σ - as was done by prior work - or, alternatively, with a two-dimensional (2D) heaviside step function of radius r (both measured in pixels). The latter simulates “bokeh” defocus blur[24],[25] which, based on qualitative observations, were found to resemble more the OOF artifacts observed in scanned WSIs. “Bokeh” defocus blur artifacts are well-known in photography, where an OOF point light source generates an approximately homogeneously illuminated image of the lense aperture, rather than a 2D Gaussian distribution.

Specifically, to generate a blurred patch for an intermediate OOF class c ∈ [1, 28] a σ-value for Gaussian blurring was chosen randomly from the interval [0.926 × exp ((c − 1) × 3/28), 0.926 × exp (c × 3/28)], while the radius r for simulated bokeh blurring was chosen from [1.4 × exp ((c − 1) × 3/28), 1.4 × exp (c × 3/28)]. Strong OOF patches for class 29 were generated by randomly choosing Gaussian σ-values from the interval (0.926 × exp (3), 132), and bokeh radii r from (1.4 × exp (3), 200), respectively. The rationale for randomly sampling from class-specific intervals over just using each interval's center value was to help the network capture a wider range of blur magnitude during training. In the formulae above, the scaling factors 0.926 and 1.4 and the maximum values 132 and 200 were calibrated to yield similar blur strengths for either method or the same OOF class. Specifically, these scaling factors were determined by minimizing the sum of squared differences between RGB values of the images over all OOF classes [Figure 4].{Figure 4}

Postblurring Joint Photographic Experts Group and noise artifacts

In initial experiments, convolutional neural networks trained on Bokeh or Gaussian-blurred synthetic examples yielded poor prediction accuracy on real OOF images; erroneously predicting almost all OOF test patches as in-focus (see Results). We hypothesized that this was caused by the artificial smoothing removing other types of real artifacts. For example, grid-like artifacts at the edges of scan lanes and Joint Photographic Experts Group (JPEG) blocks are smoothed out in artificially blurred images, but would be present in real OOF images. Thus, several categories of other artifact types were re-added after synthetic blurring [Figure 5].{Figure 5}

Visual comparison at high magnification of synthetically blurred real in-focus images to real OOF revealed other artifacts beyond OOF: pixel noise, likely from the image sensor; and JPEG compression artifacts, likely from the lossy JPEG compression applied by scanners postdigitization (scanner quality settings in our data ranged from 70 to 90). In synthetically blurred images however, both of these artifacts ranged from faint to absent, depending on the synthetic blurring magnitude, even if they were present in the original in-focus input images. This is because both pixel noise and JPEG artifacts typically consist of high spatial frequencies, which are diminished by blurring.

Consequently, simulated JPEG compression artifacts were added back into the synthetically blurred images, implemented through JPEG encoding and decoding and an image quality parameter chosen between 70% and 90%. In addition, pixel noise was added because image sensors collect Poisson-distributed noise. Pixel-wise poisson noise was simulated separately for each color channel value [INSIDE:1], by applying the mapping f(xc) = P(xc/ S) × S to each, where P is the probability mass function of the Poisson distribution, and S is a parameter that inversely controls the signal-to-noise ratio. As the latter was observed to vary significantly between different scanners and objective magnifications, the noise portion during training was varied by randomly sampling S from the interval (0.01, 64.0) for each training patch.

Test data

The OOF algorithms were evaluated using pathologist-graded real OOF artifacts (as opposed to synthetic OOF in the training set). Specifically, three prostate resection and four lymph node biopsy slides were used, including hematoxylin and eosin (H and E) and immunohistochemistry stains. Each WSI was scanned using two scanners at × 40: A Leica Aperio AT2 (pixel size: 0.251 μm × 0.251 μm) using its semi-automatic mode, and a Hamamatsu NanoZoomer S360 scanner (pixel size: 0.230 μm × 0.230 μm) using its fully automatic mode. This resulted in two test sets: seven WSIs per scanner.

A pathologist then manually and nonexhaustively identified, delineated, and graded in-focus and OOF regions, using integral grades ranging from 0 (in-focus) to 6 (very strong OOF), [Figure 6]. Half-grades (e.g., 1.5) were occasionally used when the degree of OOF was interpreted as between integral grades. Annotations were corrected by the pathologist as desired (blinded to algorithm predictions), to help achieve grading consistency across the different tissue and stain types, and scanner models. The delineated regions were rasterized using a 128 × 128 pixel-spaced grid to enable direct comparison with the predictions of the OOF classifier (see below). To ensure label purity, only patches completely contained within a delineated OOF region were used. These annotations produced a total of 37,715 patches.{Figure 6}

In addition to pathologist annotations, we collected “z-stack” scans to further evaluate ConvFocus performance on a set of images that have real OOF with a consistent pattern. We digitized a lymph node biopsy slide with z-axis increments of 0.4 μm, and spanning +4 μm to −4 μm (relative to the scanner-determined in-focus depth), [Figure 7]a. This produced a total of 21 WSIs of the same glass slide.{Figure 7}

Convolutional neural network-based algorithm

Network architecture

Our convolutional neural network architecture was a series of convolutions and pools, equivalent to the Inception (v3) architecture[7],[26] truncated at a lower layer (“MaxPool_3a_3 × 3”[26]) and with a reduced number of filters per layer (“depth_multiplier” = 0.1) to reduce computation [Figure 2]. For this task, we empirically observed that neither modification deteriorated performance (Results). The last layer was attached to a 30-way classification layer to predict the 30 OOF classes. To handle the large image sizes of each WSI, we adopted a patch-based approach. Small crops of 139 × 139 pixels-sized patches (corresponding to approximately 35 μm × 35 μm at × 40) for each WSI were used for training. To help improve performance, we applied data augmentation techniques described in,[7] randomly perturbing the orientation, brightness, contrast, hue, saturation, and adding random translational jitter. The network was trained using the softmax cross-entropy loss function, using the same learning rate schedule and other hyperparameters as previously described.[7]

Inferring out-of-focus heatmaps on whole-slide images

To obtain algorithm predictions of OOF for every region on the slide, we applied ConvFocus in a sliding window fashion across each WSI to produce algorithm predictions for every 128 × 128 patch at × 40. This stride was chosen to match the stride used for cancer detection (see below), and is adjustable. The final predicted OOF class for each patch was the class with maximal probability (“activation”) in the final layer of the network. To visualize these predictions, we used the “jet” colorspace, ranging from blue for class 0 (in-focus) to red for class 29 (strongly OOF) [Figure 3].

ConvFocus evaluation

Each algorithm's performance was evaluated by comparing the predicted OOF class (in the range 0–29) to the corresponding pathologist-annotated OOF grades (0–6 in half-point increments, 13 grades overall) among all patches with annotations within the two test sets (one for each scanner: AT2 and S360). However, the numbers of patches across the 13 different pathologist-assigned grades were unevenly distributed within each test set, and the distributions differed between the two test sets. Therefore, for each test set and grade, 3000 patches were randomly sampled with replacement to obtain evenly distributed classes.

Similar to prior work,[23] the Spearman's rank correlation coefficient was used as the main evaluation metric. To assess the statistical significance of the correlation, we used a two-sided test.[27] Next, we computed a linear regression to evaluate deviation of the intercept from 0.0, since an “in-focus” patch (annotated grade: 0.0) should ideally be predicted as “in-focus” (ConvFocus class: 0). The regression's slope was used to assess deviations from 4.83 (=29/6), which reflects a linear mapping of the ConvFocus OOF class range of (0, 29) to the graded OOF range of (0, 6).

In addition, ConvFocus was further evaluated on the z-stack test set by plotting the predicted OOF class against the z-level for a number of patches. Two aspects were qualitatively assessed. First, we checked for the presence of a “v”-shaped plot indicating that predicted OOF class increases (towards poorer focus) as the z-level goes further from in-focus (in either direction). Second, we checked if the “in-focus” plane (z = 0) was generally predicted as the lowest OOF class.

Measuring cancer detection algorithm performance as a function of focus quality

After validating ConvFocus, we used the algorithm to study the impact of OOF (both real and synthetic) on the performance of a cancer detection algorithm on the publicly available Camelyon 2016 challenge test dataset,[16] consisting of 80 non-tumor and 48 tumor slides with pixel-level annotations of tumor locations. In the first experiment (real OOF), we employed the current best-performing breast cancer metastasis detector “LYNA,”[7],[28] which we configured to provide binary predictions at the same granularity as the OOF classifier. That is, for each image patch, predictions from both the breast metastasis and ConvFocus were available. Next, we stratified the patches by predicted OOF class. We then measured the LYNA algorithm's performance for patches in each OOF class, merging the higher OOF classes to ensure sufficient numbers of both tumor and nontumor patches. In a second experiment, we added artificial Bokeh blur at a selected strength to all patches in the test set, ran LYNA on all patches, and evaluated LYNA's performance. We repeated this for a range of Bokeh-blur strengths. In both experiments, we used the patch-level area under receiver operating characteristic curve (AUC[29]). Confidence intervals for the patch-level AUC were computed using the nonparametric bootstrap approach,[30] with n = 500 samples at the slide level.

 Results



ConvFocus evaluation

Sample visualizations of ConvFocus's predictions are shown in [Figure 3] and [Figure 8], depicting regional OOF patterns such as OOF scan lanes and OOF tissue edges. Quantitative evaluation using a total of 37,715 pathologist-graded OOF patches is reported in [Table 2] and visualized in [Figure 9], showing high Spearman rank correlations >0.8 (p < 0.001) between pathologist-graded OOF and ConvFocus predictions for slides scanned on both scanners tested: AT2 and S360.{Figure 8}{Table 2}{Figure 9}

Each step in our proposed semi-synthetic data generation approach was derived using qualitative assessments of whether regions with strong OOF were appropriately detected [Figure 8]. To quantify the improvements derived from each step, we also performed ablation experiments using several configurations [Table 2]. Using Bokeh synthetic blurring instead of Gaussian yields only a small improvement in correlation metrics, but a large increase in the regression's slope parameter values. Since the synthetic Gaussian and Bokeh blurring magnitudes were visually and quantitatively aligned (Methods), similar slope values were expected for the Gaussian-trained configuration. Instead, the Gaussian configuration classified patches annotated with OOF grades 5.0–6.0 as mid-range OOF classes 14–20 [Figure S1 [SUPPORTING:1]] compared to 20 and above with the Bokeh configuration [Figure 9]. Moreover, the majority of patches with no or weak OOF (annotated Grades 0.0–1.5) were predicted as OOF classes 0–4 by the Bokeh configuration (ConvFocus), compared to 5 and higher with the Gaussian configuration.

We next assessed ConvFocus's predictions for patches from a set of 21 “z-stack” images [Figure 10]. Almost all patches were predicted to be in-focus at a z-level, though not necessarily at z = 0 [Figure 7]. A strong “v”-shaped trend was also observed: patches were generally predicted to be more OOF at z-levels further from the predicted in-focus z-level (on average at z = 0). However, using the Gaussian configuration, few patches were predicted to have OOF class Impact of focus quality on cancer detection algorithm performance

[Figure 11] shows the effect of OOF on patch-level AUC for the LYNA metastatic breast cancer detector. Results for the first experiment (effects of real OOF) are shown in black dots with gray bars that indicate 95% confidence intervals. Since few image patches in the test set were predicted as higher OOF classes, we bucketed the classes into five groups: 0–4, 5–9, 10–14, 15–19, and 20–29. Results indicated a gradual drop in cancer detector performance with predicted OOF, as evidenced by decreasing AUC values and the associated upper bounds of the confidence intervals.{Figure 11}

To assess the effects of finer-grained OOF, we conducted a second experiment: synthetically adding OOF, as depicted by the blue curve. In contrast to the first experiment, since synthetic blur at each specific degree was applied uniformly across all slides, no class-merging was necessary. Results similarly showed the gradual and consistent drop in cancer detection with increasing OOF.

 Conclusions



We have developed and evaluated an automated focus quality detector (ConvFocus) to accurately locate and classify OOF regions as small as 32 μm × 32 μm in gigapixel-sized WSIs. Our algorithm correlates well with pathologist-annotated focus quality grades on real OOF regions, and across multiple scanners despite different imaging characteristics (e.g., pixel size in the S360 is 9% smaller than in the AT2). Because different scanners do not generally digitize the same slide locations with the same focus quality, the actual image patches used in this evaluation are different. As a result, a direct and fair inter-scanner performance comparison is not possible.

In a second evaluation using z-stacks to produce digitized images of varying OOF degrees, ConvFocus similarly produced sensible trends of increasing predicted OOF classes in both directions away from the scanner-determined in-focus plane. In this evaluation, while the scanner-determined in-focus plane (z = 0) was on average predicted to be close to or the most in-focus, this was not necessarily true of all image patches. Further analysis is needed to determine whether this is caused by inaccuracies in focus detection by the scanner's algorithm, or by ConvFocus. The asymmetry in this “v”-shaped trend may be because the z-levels' span of 8 μm exceeds the thickness of a tissue section (5 μm), and thus, some z-levels will overlap the coverslip or glass slide. This effect may be further exacerbated if the z = 0 plane is not centered within the tissue section.

A significant advantage of our semi-synthetic data generation approach is an easy extension to additional scanner models, tissues, and stain types. Specifically, input data can be generated by non-expert raters, improving the speed and cost of data collection. Our results indicate that our method of synthesizing OOF examples of different degrees enables better generalization to real OOF images. In terms of granularity, ConvFocus was developed for fine-grained discrimination between 30 OOF classes, significantly more than both our pathologist's grading scale and prior works: six classes[23] and twelve classes.[22],[23] When the granularity was further increased in our experiments (e.g., to 40 classes, data not shown), training accuracy decreased, indicating that 30 classes is close to a meaningful upper bound of granularity. Our qualitative [Figure 4] and quantitative [Figure S1] analyses also found that human perception of OOF followed an exponential relationship with blur magnitude, as opposed to linear. The exponential relationship expands the granularity of fine-grained OOF classes, and is particularly relevant for use cases like lymphoma diagnosis, where even mild OOF may interfere with the ability to resolve crucial nuclear details.

Prior works in synthetic data generation for training focus quality predictors have leveraged Gaussian blurring for histopathology[23] and Gaussian blurring with Poisson noise for cytology specimens from transmission light microscopes.[22],[23] By contrast, our work systematically demonstrates the value of adding Poisson noise and JPEG compression artifacts in histopathology, and using a synthetic blur (Bokeh) that more closely mimics real OOF. Our results indicate that generating synthetic OOF data using Gaussian blurring underestimates strong OOF. Based on these improvements, we showed that focus quality can be accurately predicted on a fine-grained scale using fairly shallow neural networks, which improves the speed of applying ConvFocus to large WSIs.

We further investigated the impact of focus quality on an otherwise accurate algorithm's interpretation,[7] showing that algorithm performance was lower in image patches identified by ConvFocus as OOF. Furthermore, our controlled experiment of synthetically “de-focusing” the image patches showed similar trends, demonstrating that synthetic OOF causes degradations in algorithm performance. These findings suggest that focus quality can negatively impact algorithms, and that focus quality should be comprehensively assessed at the region level to avoid both false-positive and false-negative interpretations by an algorithm. This is particularly important for algorithms because most algorithms are developed to assess smaller cropped image patches from large WSIs (e.g., 300 pixels for a cropped patch compared to 100,000 pixels across for the entire slide). By contrast, pathologists assessing an image generally review the tissue at lower magnifications, and also have the option of zooming out to assess a region of interest at a lower magnification.

The present work contains some limitations. First, the cost of manual grading of focus quality has limited our test dataset size and diversity in terms of stains, tissue types, scanner models, as well as number of annotators. Further work will be needed to further validate the generalizability of ConvFocus against multiple rater annotations,[23] and quantifications of z-stack prediction performance. Second, using our algorithm to compare focus quality across different scanners was confounded by the fact that different image patches are identified by our pathologist as OOF. Additional work will be needed to obtain scanner-agnostic or tissue-independent metrics of focus. Finally, we have focused on the high magnification, ×40, which most clearly shows OOF artifacts. The extension to the next-lower ×20 is the subject of future work.

Acknowledgments

We would like to thank members of the Google AI Healthcare Pathology team for software infrastructural and logistical support, and slide digitization. Gratitude also goes to Daniel Fenner for assistance with Bokeh blur, and Samuel J. Yang for helpful discussions.

Financial support and sponsorship

Nil.

Conflicts of interest

T.K., Y.L., M.M., P.C., J.H., C.M., and M.S. are/were employees of Google, LLC and own/owned Alphabet stock. T.B. was compensated by Google, LLC for the participation as a pathologist in these studies.

References

1Pantanowitz L. Digital images and the future of digital pathology. J Pathol Inform 2010;1:15.
2Hashimoto N, Bautista PA, Yamaguchi M, Ohyama N, Yagi Y. Referenceless image quality evaluation for whole slide imaging. J Pathol Inform 2012;3:9.
3Mukhopadhyay S, Feldman MD, Abels E, Ashfaq R, Beltaifa S, Cacciabeve NG, et al. Whole slide imaging versus microscopy for primary diagnosis in surgical pathology: A multicenter blinded randomized noninferiority study of 1992 cases (Pivotal study). Am J Surg Pathol 2018;42:39-52.
4Janowczyk A, Madabhushi A. Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases. J Pathol Inform 2016;7:29.
5Ghaznavi F, Evans A, Madabhushi A, Feldman M. Digital imaging in pathology: Whole-slide imaging and beyond. Annu Rev Pathol 2013;8:331-59.
6Stathonikos N, Veta M, Huisman A, van Diest PJ. Going fully digital: Perspective of a Dutch academic pathology lab. J Pathol Inform 2013;4:15.
7Liu Y, Kohlberger T, Norouzi M, Dahl GE, Smith JL, Mohtashamian A, et al. Artificial intelligence-based breast cancer nodal metastasis detection: Insights into the black box for pathologists. Arch Pathol Lab Med 2019;143:859-68.
8Montalto MC, McKay RR, Filkins RJ. Autofocus methods of whole slide imaging systems and the introduction of a second-generation independent dual sensor scanning method. J Pathol Inform 2011;2:44.
9Pantanowitz L, Farahani N, Parwani A. Whole slide imaging in pathology: Advantages, limitations, and emerging perspectives. Pathol Lab Med Int 2015;7:23-33.
10McKay RR, Baxi VA, Montalto MC. The accuracy of dynamic predictive autofocusing for whole slide imaging. J Pathol Inform 2011;2:38.
11Ameisen D, Deroulers C, Perrier V, Bouhidel F, Battistella M, Legrès L, et al. Automatic Image Quality Assessment in Digital Pathology: From Idea to Implementation. Proc. of IWBBIO; Grenada, Spain, 2014.
12Zerbe N, Hufnagl P, Schlüns K. Distributed computing in image analysis using open source frameworks and application to image sharpness assessment of histological whole slide images. Diagn Pathol 2011;6 Suppl 1:S16.
13Lahrmann B, Valous NA, Eisenmann U, Wentzensen N, Grabe N. Semantic focusing allows fully automated single-layer slide scanning of cervical cytology slides. PLoS One 2013;8:e61441.
14Moles Lopez X, D'Andrea E, Barbot P, Bridoux AS, Rorive S, Salmon I, et al. An automated blur detection method for histological whole slide imaging. PLoS One 2013;8:e82710.
15Walkowski S, Szymas J. Quality evaluation of virtual slides using methods based on comparing common image areas. Diagn Pathol 2011;6 Suppl 1:S14.
16Ehteshami Bejnordi B, Veta M, Johannes van Diest P, van Ginneken B, Karssemeijer N, Litjens G, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA 2017;318:2199-210.
17Litjens G, Sánchez CI, Timofeeva N, Hermsen M, Nagtegaal I, Kovacs I, et al. Deep learning as a tool for increased accuracy and efficiency of histopathological diagnosis. Sci Rep 2016;6:26286.
18Xu J, Luo X, Wang G, Gilmore H, Madabhushi A. A deep convolutional neural network for segmenting and classifying epithelial and stromal regions in histopathological images. Neurocomputing 2016;191:214-23.
19Nagpal K, Foote D, Liu Y, Chen PC, Wulczyn E, Tan F, et al. Development and validation of a deep learning algorithm for improving Gleason scoring of prostate cancer. NPJ Digit Med 2019;2:48.
20Chen PH, Gadepalli K, MacDonald R, Liu Y, Nagpal K, Kohlberger T, et al. An augmented reality microscope with real-time artificial intelligence integration for cancer diagnosis. Nat Med 2019;25:1453-7.
21Senaras C, Niazi MK, Lozanski G, Gurcan MN. DeepFocus: Detection of out-of-focus regions in whole slide digital images using deep learning. PLoS One 2018;13:e0205387.
22Yang SJ, Berndl M, Ando DM, Barch M, Narayanaswamy A, Christiansen E, et al. Assessing microscope image focus quality with deep learning. BMC Bioinform 2018;19:1,77.
23Campanella G, Rajanna AR, Corsale L, Schüffler PJ, Yagi Y, Fuchs TJ. Towards machine learned quality control: A benchmark for sharpness quantification in digital pathology. Comput Med Imaging Graph 2018;65:142-51.
24Nasse HH. Depth of Field and Bokeh. Carl Zeiss AG. Available from: https://www.lenspire.zeiss.com/photo/app/uploads/2018/04/Article-Bokeh-2010-EN.pdf. [Last accessed on 2019 Oct 18].
25Wadhwa N, Levoy M, Garg R, Feldman B, Kanazawa N, Carroll R, et al. Synthetic depth-of-field with a single-camera mobile phone. ACM Trans Graph 2018;37:1-13.
26Szegedy C, Vanhoucke V, Ioffe S, Shlens J, Wojna Z. Rethinking the inception architecture for computer vision. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition; 2016.
27Kokoska S, Zwillinger D. CRC Standard Probability and Statistics Tables and Formulae. Boca Raton, FL, USA: CRC Press; 2000.
28Liu Y, Gadepalli K, Norouzi M, Dahl G, Kohlberger T, Boyko A, et al. Detecting Cancer Metastases on Gigapixel Pathology Images. arXiv 2017; abs/1703.02442.
29Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982;143:29-36.
30Chihara LM, Hesterberg TC. Mathematical Statistics with Resampling and R. 2nd ed. Hoboken, NJ, USA: Wiley; 2018.