|
 |
| ORIGINAL ARTICLE |
|
|
|
| J Pathol Inform 2016,
7:29 |
Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases
Andrew Janowczyk, Anant Madabhushi
Department of Biomedical Engineering, Case Western Reserve University, Cleveland, OH 44106, USA
| Date of Submission | 18-Nov-2015 |
| Date of Acceptance | 18-Mar-2016 |
| Date of Web Publication | 26-Jul-2016 |
Correspondence Address:
Andrew Janowczyk
Department of Biomedical Engineering, Case Western Reserve University, Cleveland, OH 44106
USA
 Source of Support: None, Conflict of Interest: None  | Check |
DOI: 10.4103/2153-3539.186902
 Abstract | | |
Background: Deep learning (DL) is a representation learning approach ideally suited for image analysis challenges in digital pathology (DP). The variety of image analysis tasks in the context of DP includes detection and counting (e.g., mitotic events), segmentation (e.g., nuclei), and tissue classification (e.g., cancerous vs. non-cancerous). Unfortunately, issues with slide preparation, variations in staining and scanning across sites, and vendor platforms, as well as biological variance, such as the presentation of different grades of disease, make these image analysis tasks particularly challenging. Traditional approaches, wherein domain-specific cues are manually identified and developed into task-specific "handcrafted" features, can require extensive tuning to accommodate these variances. However, DL takes a more domain agnostic approach combining both feature discovery and implementation to maximally discriminate between the classes of interest. While DL approaches have performed well in a few DP related image analysis tasks, such as detection and tissue classification, the currently available open source tools and tutorials do not provide guidance on challenges such as (a) selecting appropriate magnification, (b) managing errors in annotations in the training (or learning) dataset, and (c) identifying a suitable training set containing information rich exemplars. These foundational concepts, which are needed to successfully translate the DL paradigm to DP tasks, are non-trivial for (i) DL experts with minimal digital histology experience, and (ii) DP and image processing experts with minimal DL experience, to derive on their own, thus meriting a dedicated tutorial. Aims: This paper investigates these concepts through seven unique DP tasks as use cases to elucidate techniques needed to produce comparable, and in many cases, superior to results from the state-of-the-art hand-crafted feature-based classification approaches. Results : Specifically, in this tutorial on DL for DP image analysis, we show how an open source framework (Caffe), with a singular network architecture, can be used to address: (a) nuclei segmentation (F-score of 0.83 across 12,000 nuclei), (b) epithelium segmentation (F-score of 0.84 across 1735 regions), (c) tubule segmentation (F-score of 0.83 from 795 tubules), (d) lymphocyte detection (F-score of 0.90 across 3064 lymphocytes), (e) mitosis detection (F-score of 0.53 across 550 mitotic events), (f) invasive ductal carcinoma detection (F-score of 0.7648 on 50 k testing patches), and (g) lymphoma classification (classification accuracy of 0.97 across 374 images). Conclusion: This paper represents the largest comprehensive study of DL approaches in DP to date, with over 1200 DP images used during evaluation. The supplemental online material that accompanies this paper consists of step-by-step instructions for the usage of the supplied source code, trained models, and input data. Keywords: Classification, deep learning, detection, digital histology, machine learning, segmentation
How to cite this article:
Janowczyk A, Madabhushi A. Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases. J Pathol Inform 2016;7:29 |
| Introduction | |  |
Digital pathology (DP) is the process by which histology slides are digitized to produce high-resolution images. DP is becoming increasingly common due to the growing availability of whole slide digital scanners. [1] These digitized slides afford the possibility of applying image analysis techniques to DP for applications in detection, segmentation, and classification. Already algorithmic approaches have shown to be beneficial in many contexts as they have the capacity to not only significantly reduce the laborious and tedious nature of providing accurate quantifications (e.g., tumor extent, nuclei counts), but to act as a second reader helping to reduce inter-reader variability among pathologists. [2],[3]
A number of image analysis tasks in DP involve some sort of quantification (e.g., cell or mitosis counting) or tissue grading (classification). As shown in [Figure 1], these tasks invariably require identification of histologic primitives (e.g., nuclei, mitosis, tubules, epithelium, etc.). For example, while the spatial arrangement of nuclei in oropharyngeal [4] and breast cancers [5] has been correlated with outcome, these approaches still initially requiring deep annotations (i.e., various entities identified at different scales) to extract features from. As a result, there is a strong need to develop efficient and robust algorithms for analysis of DP images. | Figure 1: The flowchart shows a typical workflow for digital pathology research. Histologic primitives (e.g. nuclei, lymphocytes, mitosis, etc.,) are identified, after which biologically relevant features are extracted for subsequent use in higher order research directives. Typically, the tasks in the red box are undertaken by the development and upkeep of individual task specific approaches. The premise of this tutorial is that these tasks can be performed by a single generic deep learning approach, which can be easily maintained and extended upon
Click here to view |
While there have been a number of papers in the area of computational image analysis of DP images for the purposes of object detection and quantification in the last few years, there appear to be two main drawbacks to existing approaches. First, the development of task specific approaches tends to require long research and development cycles. For example, to develop a nuclei segmentation algorithm, one must first understand all of the possible variances in morphology, texture, and color appearances. Subsequently, an algorithmic scheme needs to be developed which can account for as many of these variances as possible while not being too general as to result in false positive results or too narrow as to result in false negative errors. This process can become quite unwieldy as it is often infeasible to view all of the outlier cases a priori, and thus an extensive iterative trial and error approach needs to be undertaken. Unfortunately, once a suitable set of operating parameters is found for a specific dataset, it is unlikely to directly translate to a second independent dataset, typically requiring additional parameter tweaking and tuning. This leads to the second drawback with existing approaches; the implicit knowledge of how to find or adjust optimal parameters often resides solely with the developers of the algorithms and thus are not intuitively understood by external parties. In addition, note the process above describes only a single task, in the case where a DP suite is created, consisting of a single approach for each desired task (e.g., segmentation of nuclei, detection of mitosis, etc.), there is a multiplicative burden of both steep learning curves of the nuances of each algorithm as well as the general maintenance and upkeep of multiple software projects. Together, these create a strong hindrance for researchers to leverage or extend the available technology to investigate their clinical hypothesis.
Deep learning (DL) is an example of the machine learning paradigm of feature learning; wherein DL iteratively improves upon learned representations of the underlying data with the goal of maximally attaining class separability. This is to say that every DL network begins with the same assumption of random initialization, and for each iteration, data are propagated through the network to compute its respective output. This output is compared to the desired output (e.g., determining if a pixel in question belongs to a nucleus or not), and an error is computed per parameter so that it can be adjusted to better dichotomize that training sample into the correct class. We note that there are no preexisting assumptions about the particular task or dataset, in the form of encoded domain-specific insights or properties, which guide the creation of the learned representation. The DL approach involves deriving a suitable feature space solely from the data itself. This is a critical attribute of the DL family of methods, as learning from training exemplars allows for a pathway to generalization of the learned model to other independent test sets. Once the DL network has been trained with an adequately powered training set, it is usually able to generalize well to unseen situations, obviating the need of manually engineering features.
DL is thus uniquely suited to analyze big data repositories (e.g., TCGA, which currently comprises over 1 petabyte worth of digital tissue slide images), as it is ideally suited to learn in an implicit fashion the diversity of image patterns embedded within large datasets. On the other hand, employing a feature engineering or "hand-crafted" approach might require several algorithmic iterations and substantial effort to capture a similar range of diversity. Many manually engineered or hand-crafted feature-based approaches are not implicitly poised to manipulate and distil large datasets into classifiers in an efficient way. DL approaches, on the other hand, function well under these circumstances.
DL algorithms also have the potential for being the unifying approach for the many tasks in DP, having previously been shown to produce state-of-the-art results across varied domains, including mitosis detection, [6],[7],[8] tissue classification, [9] and immunohistochemical staining. [10] While we are seeing a wide adoption of DL technology, the burden of entry for (i) DL experts with minimal DP experience and (ii) DP and image processing experts with minimal DL experience, remains quite high. The challenges specific to the context of the DP domain, such as (a) selecting appropriate magnification at which to perform the analysis or classification, (b) managing errors in annotation within the training set, and (c) identifying a suitable training set containing information rich exemplars, have not been specifically addressed by existing open source tools [11],[12] or by the numerous tutorials for DL. [13],[14] The previous DL work in DP performed very well in their respective tasks though each required a unique network architecture and training paradigm. As this manuscript is intended to be an introductory tutorial and not a thorough review of the current literature, we direct the interested reader to a number of outstanding recent papers on the use of DL for specific tasks in the context of DP. In particular, detection of invasive ductal carcinomas (IDCs), [9] mitosis detection, [8] neuron segmentation, [15] colon gland segmentation, [16] nuclei segmentation [17],[18],[19] and detection, [20] brain tumor classification, [21] epithelial tumor nuclei identification, [22] epithelium segmentation, [23] and glioma grading [24] have been previously tackled via DL strategies. However, since these approaches were originally developed in the context of specific contexts, the architecture and approach may not readily generalize to other DP tasks. As such, the focus of this manuscript is to discuss the usage of a single framework, which can be marginally tweaked to apply to a diverse set of unique use cases.
We developed this tutorial to focus specifically on the critical components often needed by DP researchers in automating tasks (e.g., grading) or investigating clinical hypothesis (e.g., prognosis prediction). The seven use cases examined in this tutorial, (a) nuclei segmentation, (b) epithelium segmentation, (c) tubule segmentation, (d) lymphocyte detection, (e) mitosis detection, (f) IDC detection, and (g) lymphoma classification, demonstrate how DL can be applied to a spectrum of the most common image analysis tasks in DP. We subdivide our seven tasks into three categories of detection (e.g., mitotic events, lymphocytes), segmentation (e.g., nuclei, epithelium, tubules), and tissue classification (e.g., IDC, lymphoma sub-types) as the approaches used within each analysis category are similar. Each task is cast into a well-studied problem, to leverage not only the open source DL framework Caffe, [25] but also using the well-known CIFAR-10 AlexNet network schema [26] (notably smaller and easier to train than the full 101 × 101 Version), [27] provided by it.
We show how a single training and model-building paradigm can be applied to each task, solely by modifying the patch selection technique, and yet still generate results that are either comparable or superior to existing handcrafted approaches. Understanding these unique patch selection techniques allows for the elucidation of best practices needed for researchers to re-apply these approaches to their own tasks. At the same time, this convergence to a unified approach not only allows for a low maintenance overhead but also implies that image analysis researchers or DP users face a minimal learning curve, as the overall learning paradigm and hyperparameters remain constant across all tasks.
As this manuscript is intended to be a didactic tool, aimed at enabling imaging and machine learning scientists to apply DL to DP problems, we are also concomitantly releasing (a) an online step-by-step guide on the implementation of the various approaches, (b) supporting source code, (c) trained network models, and (d) the data sets themselves. [28] We strongly encourage the readers to review the material, as they are intended as a supplement to the manuscript presented here. Leveraging these resources should allow readers to not only easily reproduce the results presented in this tutorial but also to have a strong basis from which to modify these approaches and align these approaches toward their own datasets and tasks. We note as well that many of the released datasets are the first of their kind to be disseminated publicly, and thus we hope these datasets will serve as an important resource by the community for use in benchmarking task specific algorithms (e.g., epithelium segmentation).
The rest of the paper is outlined as follows: Section 3 provides an overview of the DP tasks and datasets used in this tutorial. Section 4 illustrates the DL setup used, Section 5 provides the main context of the paper via the 7 different use cases, and Section 6 presents concluding remarks.
| Digital pathology tasks addressed | | |
[Table 1] presents a list of the seven different tasks addressed in this paper. These tasks have been chosen as they represent the ensemble of critical components necessary for most of the pertinent pathology tasks (e.g., disease grading, mitotic counting) and thus span the current challenges in the DP image analysis space. This is evidenced by large numbers of papers and grand challenges, which have been proposed to address these problems. [7],[29] | Table 1: Descriptions of the digital pathology tasks undertaken in this tutorial using seven different use cases
Click here to view |
Segmentation and Detection Tasks
A segmentation task is defined as the requirement of delineating an accurate boundary for histologic primitives (i.e., nuclei, epithelium, tubules, and IDC) so that precise morphological features can be extracted. Detection tasks (i.e., lymphocyte and mitosis detection) are different from segmentation tasks in that the goal is typically to simply identify the center of the primitive of interest and not explicitly extract the primitive contour or boundary. Segmentation typically tends to be more challenging than detection, especially in the cases where the primitives of interest have multiple possible manifestations (e.g., mitotic cycles). Thus, a single monolithic classifier or model may not be able to capture the full range of diversity in presentation (i.e., the constellation of visual queues and features used to identify a particular histologic primitive).
Tissue-Based Classification Task
Another set of use cases we tackle in this paper is tissue level classification (i.e., lymphoma subtype identification). As opposed to explicitly identifying individual tissue-based primitives (e.g., mitoses, nuclei) and trying to identify primitive specific features to make predictions regarding tissue class, an alternative strategy is to directly learn the set of features representative of the tissue class via DL. The DL classifier could thus be trained to self-discover the nuanced disease patterns within each class. This approach thus obviates the need for explicit primitive identification and provides a more direct pathway to the final classification while not at the same time requiring a comprehension of the (potentially unknown) domain specific relationships of the primitives. In fact, in this setting, the DL approach only needs the image patches which have been tagged with the class label to learn the most discriminating representations for class separability.
Manual Annotation for Ground Truth Generation
Well annotated exemplars are an important prerequisite for DL schemes; unfortunately, the main challenge in performing any digital histopathology work is to obtain high-quality annotations. These ground truth annotations, typically done by an expert, involves delineating object boundaries or annotating pixels corresponding to a region or tissue of interest. In computational approaches, this level of annotation precision is critical so that supervised classification systems, and more specifically learn-from-data approaches (where domain knowledge is not explicitly implemented in the algorithm), can be optimized. Generating these annotations, though, is a cumbersome and laborious process, and often quite onerous, due to the large amount of time and effort needed. For example, the nuclei annotation dataset used in this work took over 40 hours to annotate 12,000 nuclei, and yet represents only a small fraction of the total number of nuclei present in all images.
There have been discussions previously in the literature, [9],[30] regarding the challenges associated with supervised learning classifiers that have to rely on large swathes of deeply annotation data. The findings from [30] show that metrics computed from a single resolution appear to degrade at a finer resolution, not because the tissue classifier presented was performing worse. On the contrary, the classifier became so sophisticated at the higher magnification that it began to tease apart regions that were too subtle to be captured by an expert approximate delineations [a similar situation is shown in [Figure 5]]. A large contributory reason is the fact that pathologists are typically not available to perform the large amounts of laborious manual annotations at the high resolutions needed for training and evaluating supervised object detection and classification algorithms. As a result, annotations are (a) rarely pixel level precise, (b) usually done at a lower magnification, and (c) tend to contain numerous false positives and negatives. For example, the annotation of the IDC dataset took place at ×2, while there are many subregions visible at ×5 and ×10 that are clearly not IDC, but since the delineation happens quickly and at a high level, those regions are falsely included in the positive class.
There can also be an issue with the ambiguity naturally present in biological images, especially where three-dimensional (3D) objects are represented in 2D, further confounding the annotation process. For example, annotating clumps or overlapping nuclei is a challenge since it is not always clear where the boundaries between intersecting nuclei lie. This is an unfortunate artifact of tissue sectioning and representation of fundamentally 3D tissue sections as a 2D planar image on a glass slide. In the discussion section below, we discuss an approach that aims to optimize the process of ground truth generation and annotation construction.
| Deep learning methods | | |
This section is divided into two parts. The first sub-section discusses the typical workflow used when applying DL to a DP image analysis task. The second subsection briefly describes the components typically used in constructing a DL architecture (i.e., network); instantiated as the popular CIFAR-10 version of AlexNet. [26]
Overview of Deep Learning Workflows
The DL approach employed in conjunction with the 7 use cases can be thought of as comprising the following four high-level modules.
Casting
Typically, one needs to make various decisions to design an appropriate network such as input patch size, number of layers, and convolutional attributes. We attempt to mitigate this dependency by instead opting to leverage the popular and successful AlexNet network (described below). The main reasons for using an existing architecture are 2-fold. First, finding the most successful network configuration for a given problem can be a difficult challenge given the total number of possible configurations one could avail of and also the concomitant amount of time for training and testing the network. By choosing an existing proven network, we can measure the performance of other configurations against a known benchmark. Second, since the patch size of 32 × 32 is associated with a well-known image benchmark challenge CIFAR-10, [31] and since we use a popular open-source DL framework (Caffe), [25] we create a situation by which future upgrades are essentially obtained for "free." As newer, more efficient/accurate networks and training produces become available and integrated into Caffe, they can be directly leveraged. If we were to design our own network, without regard to input sizes and software, we would require significant upkeep to leverage any future advances in DL techniques.
Patch generation
Once the network is defined, which involves locking down input sizes, image patches need to be generated to construct the training and validation sets. This stage requires modest domain knowledge in order to ensure a good representation of diversity in the training set. Since our chosen network has limited discrimination ability (drastically reducing the likelihood of over-fitting the model), selecting appropriate image patches for the specific task could have a dramatic effect on the outcome. Especially in the domain of histopathology, there can be substantial variance present within a single target class, such as nuclei. This is especially pronounced in breast cancer nuclei, where nuclear areas can vary upwards of 200% between nuclei. Ensuring that a sufficiently rich set of exemplars is extracted from the images is perhaps one of the most key aspects of effectively leveraging and utilizing a DL approach. In Section 5: Use Cases, we present a detailed description of approaches that can allow for tailoring of training sets toward specific tasks.
Training
The training procedure for all tasks is essentially the same and follows the well-established paradigm laid out in. [32] This strategy utilizes a stochastic gradient descent approach, with a fixed batch size, (a) a series of mean corrected image patches are introduced to the network over a series of epochs, (b) an error derivative calculated, and (c) back-propagated through the network by updating the network weights. The learning rate is annealed over time so that a local minimum is reached. The resulting learned weights (i.e., the model) are stored to be used later at test time.
Testing
By submitting image patches to the network, of the same size used during training, we obtain a class prediction from the learned model.
Review of AlexNet Network Architecture
Although a full DL primer is out of the scope of this paper, we briefly discuss the components which make up the popular AlexNet and then follow on by describing the full network. We strongly encourage the reader to review, [26],[27] for a complete understanding of the network. We assume the input image to be of size w × w × c, where w is both the width and height and c is the number of channel. In addition, one represents grayscale, and three represents red-green-blue.
Convolutional layer
This layer type takes a square kernel of size k × k, which is smaller than the input w, and is then convolved with the image to obtain network activations. A number of these kernels are learned such that they minimize the training error function (discussed below). Due to the static nature of natural images, especially in histopathology, a single bank of filters can optimally represent the many components present in an image. A convolutional filter is often likened to a local receptor field, where spatially proximal inputs are mapped to a single value through a filter activation. Convolutional layers are interesting because they minimize the number of individual variables required (since k2 <<w2 ), while still displaying strong representational ability. As seen in previous papers, [33] the first layer tends to be similar to edge detectors, Gabor filters, and other first order filters. Artificially augmenting an image to a specified size (i.e., padding) often takes place to ensure favorable computational properties, such as the number of elements being processing coinciding with a power of 2 for improved GPU efficiency. Padding can be done by appending zeros but often times involve simply mirroring adjacent pixels. In addition, there is an optional stride component which specifies the intervals at which to apply the filter. Output of this is of size:.
Pooling layer
These layers are used as a way of summarizing the information created from the layer above. Two types of pooling layers are typically used, max and average, which summarize an area of k × k into either the maximal value or the mean value. The output size is computed in a manner similar to the convolutional layer.
Inner product (fully connected)
This is the traditionally fully-connected layer where every input is fed into a unique output after being multiplied by a learned weight. Inner products are easily represented by matrix multiplications of a weight matrix and the input vector to produce a vector output, which is the same size as that of the previously specified number of neurons.
Activation layer
This layer operates on each element individually (i.e., element-wise) to introduce nonlinearity into the system. In past approaches, [34] a sigmoid function was typically used, but more recent implementations [35],[36] have shown that a rectified linear (ReLu) activation has more favorable properties. These properties include sparser activation, elimination of vanishing/exploding gradient issues, and more efficient computation as the underlying function consists of only a comparison, addition, and multiplication. In addition, one can argue that this type of activation is more biologically plausible, [37] allowing for more consonance with the way the human brain functions. A ReLu activation is of the form f (x) = max (0, x).
Dropout layer
Performed on the fully connected layers, dropout [38] is the process of randomly excluding different neurons during each iteration of training. This has been shown to improve generalizability of the classifier to unseen cases while also eliminating overfitting as weights cannot become co-dependent. It has been shown that simply using this procedure, which could improve the training time as additional computations are simply avoided, is on par with training multiple nets with different initialization points and averaging their resulting probabilities.
Softmax layer
The entire network is optimized to minimize a loss function. In all cases discussed here, we use the softmax loss function which computes the multinomial logistic loss of values presented to it. The purpose of using softmax, as opposed to a regular argmax function is that the softmax function has the favorable property of a smooth gradient, so that the back-propagated error is not subject to discontinuities, allowing for easier training.
Using these components, the AlexNet describes a complete DL architecture according to [Table 2]. As we can see, the model accepts as input a 32 × 32 image and ends up by producing a class prediction using a softmax operation. | Table 2: The AlexNet configurations are used in this work. The network is identical to the one provided by Caffe. The dropout network is the same except layers 7 and 8 have an additional dropout combined with the ReLu
Click here to view |
| Use cases | | |
We use each of the individual tasks, described in [Table 1], as a vehicle to describing the unique challenges present in DP and the solutions we have implemented. We briefly discuss their clinical motivation, unique dataset characteristics, and the resulting patch generation schemas. Once the individual patches are created, in a manner that is unique to each task, the same network architecture (Section 4.2: Review of AlexNet Network Architecture) and hyperparameters are used. We present qualitative and quantitative results as well as comparisons to state-of-the-art hand-crafted classification approaches.
Deep Learning Parameters
The parameters used with the stock AlexNet architecture are shown in [Table 3]. They were held constant to further illustrate how parameter tweaking and tuning is not strictly necessary to yield good quality results. Also to alleviate the need for an annealmeant schedule of the learning rate, we use AdaGrad [39] (which is supplied by Caffe), where optimal learning rates on a per variable basis are continuously estimated. We note that the training time is about 22 h on a Tesla M2090 GPU using CUDA 5.0 without cuDNN, and about 4 h using a Tesla K20c with CUDA 7.0 using cuDNN for all experiments as the number of iterations and mini-batch size was fixed. | Table 3: Deep learning hyperparameter settings held constant for all experiments
Click here to view |
A subset of the tasks below were performed using the dropout network described above. There did not exist a case where dropout improved the metrics in any of the experiments so further investigation was not performed. This is unsurprising as the original dropout paper [38] discusses that the optimal dataset size for dropout usage is smaller than the ones we have created here. As our training sets are quite large, we saw no evidence of overfitting, which further reduces the motivation for the usage of a dropout approach.
Nuclei Segmentation Use Case
Challenge
Nuclei segmentation is an important problem for two critical reasons: (a) There is evidence that the configuration of nuclei is correlated with outcome, [5] and (b) nuclear morphology is a key component in most cancer grading schemes. [40],[41] A recent review of nuclei segmentation literature [42] shows that detecting these nuclei tends not to be extremely challenging, but accurately finding their borders and/or dividing overlapping nuclei is the current challenge. The overlap resolution techniques are typically applied as postprocessing on segmentation outputs, and thus outside of the scope of this paper. We have specifically chosen to look at the problem of detecting nuclei within hematoxylin and eosin (H&E) stained estrogen receptor positive (ER+) breast cancer images.
Manually annotating all of the nuclei in a single image is not only laborious but also does not generalize to all of the other variances present by other patients and their stain/protocol variances. As a result, time is better invested annotating sub-sections of each image for a number of minutes. Unfortunately, this creates a challenging situation for generating training patches. Typically, one would use the annotations as a binary mask created for the positive class, and the negation of that mask as the negative class, randomly sampling from both to create a training set. In this particular case, though while one can successfully randomly sample from the positive mask, the randomly sampling from the complement image may or may return unmarked nuclei belonging to the positive class.
Patch selection technique
An example of a standard approach for patch selection could involve selecting patches from the positive class, and using a threshold on the color-deconvolved image [43] to determine examples of the negative class (Examples of the patches are shown in [Figure 2]). This rationale is based on the fact that nonnuclei regions tend not to strongly absorb hemotoxin. [Figure 2] shows that while the patches would correctly correspond to their associated class, the negative class [Figure 2]a would not be particularly informative from the perspective of training the network. The resulting network consequently has very poor performance in correctly delineating nuclei, as shown in [Figure 3]d, since these edges are underrepresented in the training set.
To compensate, we extend the standard approach, discussed above, with intelligently sampled challenging patches for the negative class training set. [Figure 3]a shows an example image with its associated nuclear mask in [Figure 3]b. Note that only a subset of the nuclei is annotated. Using [Figure 3]b to identify positive pixels and the basic color deconvolution [43] thresholding approach to select random negative patches, we obtain the segmented nuclei in [Figure 3]d. However, as may be evidenced by the result in [Figure 3]d, the network is unable to accurately identify nuclear boundaries. To enhance these boundaries, an edge mask is produced by morphological dilation of [Figure 3]b, in turn yielding the result shown in [Figure 3]c. From the dilated mask, we select negative training patches, [Figure 2]c which are inherently difficult to learn due to their similarity with the positive class. We still include a small proportion of the stromal patches to ensure that these exemplars are well represented in the learning set. This patch selection technique results in clearly separated nuclei with more accurate boundaries, as seen in [Figure 3]e. | Figure 2: Typical patches extracted for use in training a nuclear segmentation classifier. Six examples of (a) the negative class show large areas of stroma which are notably different than (b) the positive nuclei class and tend to be very easily classified. To compensate, we supplement the training set with (c) patches which are exactly on the edge of the nuclei, forcing the network to learn boundaries better
Click here to view |
 | Figure 3: The process of creation of training exemplars to enhance the result obtained via deep learning for nuclei segmentation. The original image (a) only has (b) a select few of its nuclei annotated. This makes it difficult to find patches which represent a challenging negative class. Our approach involves augmenting a basic negative class, created by sampling from the thresholded color deconvoluted image. More challenging patches are supplied by (c) a dilated edge mask. Sampling locations from (c) allows us to create negative class samples which are of very high utility for the deep learning algorithm. As a result, our improved patch selection technique leads to (e) notably better-delineated nuclei boundaries as compared to the approach shown in (d)
Click here to view |
Results and Discussions
Each of the 5-folds in the cross-validation set had about 100 training and 28 testing images. We use a ratio of 1:1:0.3 in selecting positive patches, negative edge patches, and miscellaneous negative patches for a total of 130 k patches in the training set. We present metrics at both ×20 and ×40. For the detection rate,
represent true positives, false positives, and false negatives, respectively. Note that the probability map obtained via DL is thresholded at 0.5 to obtain a binary result.
Qualitatively, we can see in [Figure 4] that the quantitative results correspond to the visual results. The network yields crisper nuclear boundaries that are more accurately delineated at the ×40 magnification, compared to the ×20 resolution. | Figure 4: Nuclear segmentation output as produced by our approach wherein the original image in (a) is shown with (b) the associated manually annotated ground truth. When applying the network at × 40 probability map (c) is obtained, where the more green a pixel is, the higher the probability associated with it belonging to the nuclei class. The × 20 version is shown in (d)
Click here to view |
Quantitatively, from the [Table 4], we can see in all cases that the higher ×40 magnification testing performs better than the lower ×20 magnification. This is not unexpected owing to the higher strength signal embedded within the higher magnification. We also note that the detection rate, i.e., the ability to find nuclei in the image, is very high, with the network identifying 98% of all nuclei at the ×40 magnification. Dropout appears to negatively impact the metrics here. In addition, we note that in the recent review paper, [42] the performance measures are on par with several state-of-the-art nuclear detection algorithms. | Table 4: Results for both ×20 and ×40 magnifications showing detection accuracy, F-score, true positive rate, and positive predictive value. We can see that in all cases, operating at the higher magnification produces more accurate results, though at the cost of computation time. The variances of all reported metrics were <0.001
Click here to view |
Epithelium Segmentation Use Case
Challenge
The identification of epithelium and stroma regions is important since regions of cancer are typically manifested in the epithelium. In addition, recent work by Beck et al. [44] suggest that histologic patterns within the stroma might be critical in predicting overall survival and outcome in breast cancer patients. Thus, from the perspective of developing algorithms for predicting prognosis of disease, the epithelium-stroma separation becomes critical.
This task is unique in that it is less definitive than the more obvious tasks of mitosis detection and nuclei segmentation where the expected results are quite clear. Epithelium segmentation, especially the subcomponent of identifying clinically relevant epithelium, is typically done more abstractly by experts at lower magnifications. This has been discussed above in Section 3.3: Manual Annotation for Ground Truth Generation, but for a concrete example consider [Figure 5], which shows expert annotation versus our output. Due to such discrepancies, which can make both training and evaluating difficulties, we consider an additional expert evaluation metric to validate our results. | Figure 5: Epithelium segmentation output as produced by our approach where original images in (a and d) have their associated ground truth in (b and e) overlaid. We can see that the results from the deep learning, in (c and f), that a pixel level metric is perhaps not ultimately suited to quantify this task as deep learning is better able to provide a pixel level classification, intractable for a human expert to parallel
Click here to view |
Patch selection technique
Given that our AlexNet approach constrains input data to a 32 × 32 window, we need to appropriately scale the task to fit into this context. The general principal employed is that a human expert should be able to make an educated decision based solely on the context present in the patch supplied to the DL network. What this fundamentally implies is that we must a priori select an appropriate magnification from which to extract the patches and perform the testing. In this particular case, we downsample each image to have an apparent magnification of ×10 (i.e., a 50% reduction) so that sufficient context is available for use with the network. Networks which accept larger patch sizes could thus potentially use higher magnifications, at the cost of longer training times, if necessary.
Similar to the nuclei segmentation task discussed above, we aim to reduce the presence of uninteresting training examples in the dataset, so that learning time can be dedicated to more complex edge cases. Epithelium segmentation can have areas of fat or the white background of the stage of the microscope removed by applying a threshold at conservative level of 0.8 to the grayscale image, thus removing those pixels from the patch selection pool. In addition, to enhance the classifiers ability to provide crisp boundaries, samples are taken from the outside edges of the positive regions, as discussed above in Section 5.2: Nuclei Segmentation Use Case.
Results and Discussion
Each of the 5-fold cross validation sets has about 34 training images and 8 test images. We use a ratio of 5:5:1.5 in selecting positive patches, negative edge patches, and miscellaneous negative patches for a total of 765 k patches in the training set.
Quantitatively, we evaluate our results using the F-score after applying (a) a thresholding procedure to eliminate all the white pixels from the background, (b) an area threshold to remove all objects with an area <300 as these areas are not clinically relevant. Next, we aim to identify the optimal threshold using the 1 st fold and apply it to all other folds. In addition, we separately report the F-score of each fold the corresponding to the unique optimal threshold that was identified. These results are summarized in [Table 5]. | Table 5: F-scores for epithelium versus stroma segmentation task. We can see that the optimal thresholds of each fold are close to each other as are the F-scores
Click here to view |
We can see that the individual optimal thresholds are all very near each other. These findings appear to suggest that the network and the classifier are relatively robust to variations in the training set.
Qualitatively, from the above [Figure 5], we can see that pathologists often treat this task as a higher level abstraction instead of a pixel level classification. It becomes clear in panel (f) why we exclude white pixels from the metric computation, as these gaps correspond to white background which is rarely removed manually by the pathologist (as shown in (e)). We note that we are also able to identify smaller regions which are often ignored by pathologists, most likely since they are not believed to be clinically relevant.
While visually our results appear quite similar to the original ground truth, the additional pixel level detail that the DL segmentation yields are not quite captured by the quantitative metrics, as we discussed in Section 3.3: Manual Annotation for Ground Truth Generation.
Apart from the quantitative performance measures, we also had our results reviewed by our clinical collaborator and these results were then graded on a scale of 1-5, where 1 is "poor, not fit for purpose" and 5 is "definitely fit for purpose." On average, our images were scored a 4 with a standard deviation of 0.8. This implies that overall our results are suitable to be used in conjunction with other classification algorithms (e.g., prognosis prediction).
Interestingly, this is the first attempt, to our knowledge, to directly segment and quantify epithelium tissue in general and more specifically in breast tissue. We hope that with the release of our dataset, with annotations, other researchers will be interested in using it as a benchmark to quantify their respective segmentation approaches.
Tubule Segmentation Use Case
Challenge
The morphology of tubules is correlated with the aggressiveness of the cancer, where later stage cancers present with the tubules becoming increasingly disorganized, as seen in [Figure 6]. The Nottingham breast [40] cancer grading criteria divides scoring of the tubules into three categories according the area relative to a high power field of view: (i) >75%, (ii) 10-75%, and (iii) <10%. The benefits of being able to identify and segment the tubules are thus 2-fold, (a) automate the area estimation, decreasing inter-/intra-reader variances, and (b) provide greater specificity, which can potentially lead to better stratifications associated with prognosis indication. | Figure 6: The benign tubules, outlined in red, (a) are more organized and similar, as a result the deep learning can provide very clear boundaries (b), where the stronger green indicates a higher likelihood that a pixel belongs to the tubule class. On the other hand, when considering malignant tubules (c), the variances are quite large making it more difficult for a learn from data approach to generalize to all unseen cases. Our results (d) are able to identify a large portion of the associated pixels, but can be seen providing incorrect labeling in situations where traditional structures are not present
Click here to view |
Tubules are the most complex structures considered so far. They not only consist of numerous components (e.g., nuclei, epithelium, and lumen) but also the organizational structure of these components determines tubule boundaries. There is a very large variance in the way tubules present given the underlying aggressiveness and stage of cancer. In benign cases [Figure 6]a, tubules present in a well-organized fashion with similar size and morphological properties, making their segmentation easier, while in cancerous cases [Figure 6]c, it is clear that the organization structure breaks down and accurately identifying the boundary becomes challenging, even for experts. To further compound the complexity of the situation, tubules as an entity are much larger compared to their individual components, thus requiring a greater viewing area to provide sufficient context to make an accurate assessment.
Patch selection technique
In this use case, we introduce the concept of using cheap preprocessing to help identify challenging patches, which can help provide more informative and diverse exemplars to the DL system. Per image, we randomly select a number of pixels (e.g., 15,000) belonging to both classes to act as training samples, and compute a limited set of texture features (i.e., contrast, correlation, energy, and homogeneity). These features were chosen because they are available in MATLAB and are also very fast to compute. Next, we use a naïve Bayesian classifier to determine posterior probabilities of class membership for all the pixels in the image. In a matter of seconds, we are able to identify pixels which would potentially produce false positives and negatives and thus would benefit from additional representation in the DL training set. These pixels are selected based on their magnitude of confidence, such that false positives with posterior probabilities closer to 1 are selected with greater likelihood than those with .51. This approach further helps us to bootstrap our training set, by removing trivial samples, without requiring any additional domain knowledge.
Finally, knowing that benign cases are easier to segment than malignant cases, patches are disproportionally selected from malignant cases to further help with generalizability. While this dataset comes with the samples divided into benign and malignant cases, which is a valuable piece of knowledge to have ahead of time, an approach discussed in Section 5.5: Invasive Ductal Carcinoma Segmentation Use Case, could just as easily have been used to help dichotomize the training set.
Results and discussion
[Figure 6] shows that benign sections of tissue do well as a result of being able to generalize well from the dataset. Malignant tubules, on the other hand, are far more abstract and tend to have the hallmarks of a tubule, such as clear epithelial ring around a lumen, less obvious making them harder to generalize to. This is potentially one of the downfalls of machine learning techniques, which make inferences from training data; when insufficient examples are provided to cover all cases expected to be viewed in testing phases the approaches begin to fail. On the other hand, in this case, especially these challenges could be addressed by providing a larger database of malignant images.
Each of the 5-fold cross validation sets has about 21 training images and 5 test images. We use a ratio of 2:1 of malignant to benign patches whereas also including rotations of 180 and 270 to the malignant training set, for a total of about 320 k training patches. The mean F-score, using a threshold of 0.5, was 0.827 ± 0.05. When we optimized the threshold on a per fold basis, the measure rose slightly to 0.836 ± 0.05. To determine if this was suitable for clinical usage, we computed the difference in area between our results and the ground truth results. When combining all the test sets together, the P = 0.33, indicating that there was no significant difference between the expected clinical grade associated with our approach versus and expert's ground truth annotation. Two state-of-the-art approaches claim 86% accuracy [45] and 0.845 object-level dice coefficient, [46] indicating that our approach is on par with others currently in the field.
Invasive Ductal Carcinoma Segmentation Use Case
Challenge
Invasive Ductal Carcinoma (IDC) is the most common subtype of all breast cancers. To assign an aggressiveness grade to a whole mount sample, pathologists typically focus on the regions which contain the IDC. As a result, one of the common preprocessing steps for automatic aggressiveness grading is to delineate the exact regions of IDC inside of a whole mount slide.
We obtained the exact dataset, down to the patch level, from the authors of [9] to allow for a head to head comparison with their state-of-the-art approach, and recreate the experiment using our network. The challenge, simply stated is can our smaller more compact network produce comparable results? Our approach is at a notable disadvantage as their network accepts patches of size 50 × 50, while ours use 32 × 32, thus being provided 60% less pixels of context to the classifier.
Patch selection technique
To provide sufficient context (as discussed above in epithelium segmentation section), the authors have down sampled their original ×40 images by a factor of 16:1, for an apparent magnification of ×2.5. We attempted three different approaches of using these 50 × 50 patches, and casting them into our 32 × 32 solution domain:
Resizing
Using the entire 50 × 50 patch, we resize it down to 32 × 32.
Cropping : Each 50 × 50 image was cropped to a 32 × 32 sub-patch from exactly the center to ensure that the class label was correctly retained.
Cropping + additional rotations : To compensate for the heavily imbalanced training set, where the negative class is represented over 3 times as much, we artificially oversample the positive class by adding additional rotations. Since the provided patches are 50 × 50, we can rotate them around the center of the image origin, and still crop out a 32 × 32 image. As a result, we use rotations of 0, 45, 90, 135, 180 degrees, along with their mirrors to the training set for the positive class. We continue to use only the 2 rotations for the negative class as before. The totals patches available for training are about 157 k for the positive set and about 167 k for the negative set, nearly balancing the classes.
Results and discussion
Qualitatively, we can see from [Figure 7] that our results are quite similar to. [9] While the pathologist annotations are shown in green in [Figure 7]a, we note that in our results, the upper right corner is not a false positive, but simply a region underannotated by the pathologist. As we discussed above in Section 3.3: Manual Annotation for Ground Truth Generation, this continues to be one of the challenges in DP; computer algorithms can often be more fine grained as the computation time is cheap while performing the same level of annotation for a pathologist is simply too laborious.
Quantitatively, we present the F-score and the balanced accuracy for our methods to compare against [9] in [Table 6]. We can see that using our net provides a better F-score and also a slightly higher accuracy balance. Interestingly, resizing the images seems to produce the best results indicating that the selected field of view is critical to obtaining better results. While cropping the images produces better resolution patches, the field of view is smaller, most likely making certain areas tricky to differentiate without neighborhood information. Again, we note that dropout did not provide any improvement in generalization during test time. | Figure 7: Invasive ductal carcinoma segmentation where we see the original sample (a) with the pathologist annotated region shown in green. From (b) we can see the results generated by the resizing approach, (c) shows the same results without resizing, (d) shows the output when resizing and balancing the training set and (e) finally resizing with dropout, where the more red a pixel is, the more likely it represents an invasive ductal carcinomas pixel. We note that the upper half of the image actually contains true positives which were not annotated by the pathologist
Click here to view |
 | Table 6: F-score and balance accuracy for the various approaches. We note that resizing the larger patches to fit into our existing framework provided the best results, as well as improving upon previous results using the same dataset
Click here to view |
Lymphocyte Detection Use Case
Challenge
Lymphocytes, a subtype of white blood cells, are an important part of the immune system. Lymphocytic infiltration is the process by which the density of lymphocytes greatly increases at sites of disease or foreign bodies, indicating an immune response. A stronger immune response has been highly correlated to better outcomes in many forms of cancer, such as breast and ovarian. As a result, identifying and quantifying the density and location of lymphocytes has gained a lot of interest recently, particularly in the context of identifying which cancer patients to place on immunotherapy.
Lymphocytes present with a blue tint from the absorption of hemotoxylin, their appearance similar in hue to nuclei, making them difficult to differentiate in some cases. Typically, though, lymphocytes tend to be smaller, more chromatically dense, and circular. In this particular use case, our goal was to identify the center of lymphocytes, making this a detection problem (see Section 3: Digital Pathology Tasks Addressed).
Patch selection technique
At the original ×40 magnification, the average size of a lymphocyte is approximately 10 pixels in diameter, much smaller than the 32 × 32 patches used by our network. The focus here is on identifying lymphocytes without focusing on the surrounding tissue if the patches were to be extracted at ×40, only 10% of the input pixels would be of interest. The other 90% of the input pixels would eventually learn to be ignored by the network. This would have the unfortunate effect of reducing the discriminative ability of the network. Thus, to increase the predictive power of the system, we artificially resize the images to be ×4 as large, so that the entire input space, when centered around a lymphocyte, contains lymphocyte pixels, allowing more of the weights in the network to be useful.
Positive class exemplars are extracted by randomly sampling locations from a 3 × 3 region around the supplied centers of each lymphocyte. The selection of the negative class proceeds as follows, (a) a naïve Bayesian classifier is trained on 1000 randomly selected pixels from the image to generate posterior class membership probabilities for all pixels in the image, (b) for all false positive errors, the distance between the false positive pixels and the closest true positive pixels is computed, (c) iteratively, the pixel with the greatest distance between the false positive and true positive errors is chosen so that negative image patches can be generated from those locations. Since there are few positive samples available, the training set is augmented by adding additional rotations.
At test time, the posterior probabilities are computed for every pixel in the test image. To identify the location most likely to be the center of a lymphocyte, a convolution is performed with a disk kernel and the probability output so that the center of the probably regions are highlighted. Iteratively, the highest point in the image is taken as center and a radius is cleared, which is the same size as a typical lymphocyte to prevent multiple centers from being identified for the same lymphocyte.
Results and discussion
Each of the 5-fold cross validation sets has about 80 training and 21 test images. We use a ratio of 1:1 for the positive and negative classes, while also including rotations of 180° and 270° to the positive training set due to them being under-represented, for a total of about 700 k training patches. We used a single fold to optimize the variables (disk clearing, convolution disk size, and threshold) and applied them unchanged to the other 4 folds. The optimal threshold was found to be at 0.7066, convolution disk size of 6 and clearing disk of size 28. The mean F-score was found to be 0.90 ± 0.01, mean TPR 0.93 ± 0.01 and PPV of 0.87 ± 0.02, demonstrating a favorable comparison to the states of the art which show (a) a TPR of 86% and PPV of 64% [47] and (b) F-score of 88.48. [48] Qualitatively, as shown above in [Figure 8], we are able to detect most of the lymphocytes. The dataset itself has lymphocytes on the borders of the image, often times with over 50% of the lymphocyte not being visible [as shown above in [Figure 8]a], making detection difficult for such edge pixels. | Figure 8: Lymphocyte detection result where green dots are the ground truth, and red dots are the centers discovered by the algorithm. The image on the left (a) has 21 TP/2 FP/0 FN. The false positives are on the edges, about 1 o'clock and 3 o'clock. The image on the right (b) one has 11 TP/1 FP/2 FN. We can see the false negatives are quite small and not very clear making it hard to detect them without also encountering many false positives. The only false positive is in the middle at around 7 o'clock though this structure does look "lymphocyte-like
Click here to view |
Mitosis Detection Use Case
Challenge
The number of mitoses present per high power field is an important aspect of breast cancer grade. Typically, the more aggressive the cancer, the faster the cells are dividing which can be approximated by counting the mitotic events in a histologic snapshot. The current grading scheme divides the mitotic counts into three categories per 10 high-power fields, (i) ≤7 mitoses, (ii) 8−14 mitoses, and (iii) ≥15 mitoses. This is an active area of interest with a number of competitive grand challenges taking place in this space. [6],[7],[8]
In practice, pathologists rely on changing the focal length of an optical microscope to visualize 3 dimensionally the mitotic structure, allowing them to eliminate false positives from their estimates. As such, accurately identifying mitosis on a 2D digital histology image is very difficult but highly sought after as it would allow for the automatic interrogation of existing large, long-term, repositories. An open question in the field is trying to determine the minimal amount of accuracy necessary for clinical usage.
Patch selection technique
Since the network is smaller than the one used in [8] (32 × 32 as compared to 101 × 101), we modified and extended the approach in [8] accordingly. In order to provide enough context for each of the patches, we perform all operations at ×20 apparent magnification, such that an entire mitotic figure can be captured within a single image patch. This is most important in cases where the mitosis is in the anaphase or telophase [Figure 9]c, and the coordinates provided by the ground truth are actually in the middle of the two new cells. | Figure 9: Result of deep learning for mitosis detection, where the blue ratio segmentation approach is used to generate the initial result in (a). We take this input and dilate it to greatly reduce the total area of interest in a sample. (b) In the final image, (c) we can see that the mitosis is indeed located in the middle of the image, included our computational mask. We can see that the mitosis is in the telophase stage, such that the DNA components have split into two pieces (in yellow circle), making it more difficult to identify
Click here to view |
For the positive class we take each known mitosis location, and use a 4-pixel radius around it to construct the corresponding training patch. Since there are very few training pixels available, we add a large number of rotations to augment the training set, in this case, rotations of 0, 45, 90, 135, 180, 215, 270 degrees.
For the negative class, and to reduce both computational time, and in order to improve the selection of image patches, we leverage a well-known segmentation technique termed blue-ratio segmentation since there is evidence that mitoses are highlighted in regions identified by the blue ratio segmentation scheme [Figure 9]a. [49],[50] The results of the blue ratio segmentation approach are dilated into a 20 disk radial mask [Figure 9]b. This creates regions from which we will sample the negative patches, as it enables the natural elimination of trivial examples from the learning process. We sample 2.5 times as many patches as positive patches but only rotate each of them in 0, 90, 180, 270 degrees, so that we have more unique patches instead of simply rotated images.
Subsequently and modeled on the approach in, [8] a naïve Bayesian is employed in order to compute the probability masks for the training set. A new DL network is then trained by oversampling from the false positives produced by the first network. This is done so that we can focus the classification power of the network on the most difficult cases. In particular, for the ground truth, we use the same positive class selection except we increase the number of rotations to every 15°. For the negative class, we only consider probabilities which are in the blue ratio generated mask, and sample those according to their weights. This makes it possible to select pixels which were, incorrectly, strongly believed to be a mitotic event. This approach resulted in approximately 600 k patches for the first stage of training and 4 million patches for the second stage of training. To identify the final locations of the mitoses, we convolve the image with a kernel disk and identify a mitotic event as those image locations identified as being above a certain probability threshold.
Results and discussion
Our 5-fold analysis produced a mean F-score of 0.37 ± 0.2 when using the first round classifier and 0.54 ± 0.1 when using the second trained classifier, indicating a substantial improvement when using two sequential DL networks, where the second DL network is trained based off the false positive errors identified by the first DL network. Our F-scores are comparable to the state-of-the-art and only marginally lower than the winner of a recent grand challenge competition on mitosis detection. [8] The winners of that grand challenge (F-score = 0.61) use a 101 × 101 size patches which operates at ×40, and thus contains increased classification power as compared to our 32 × 32 approach at ×20. In our runs of cross-validation, the thresholds varied significantly across different folds suggesting that an independent validation set is needed for evaluating the trained network. Typical false and true positives can be seen in [Figure 10]a and b, respectively. | Figure 10: False positive samples of mitoses (a) with (b) true positive samples on the right. We can see that in many cases the two classes are indistinguishable from each other in the two-dimensional plane, thus requiring the common practice of focal length manipulation of the microscope to determine which instances are truly mitotic events
Click here to view |
Lymphoma Subtype Classification Use Case
Challenge
The NIA curated this dataset to address the need of identifying three sub-types of lymphoma: Chronic lymphocytic leukemia (CLL), follicular lymphoma (FL), and mantle cell lymphoma (MCL). Currently, class-specific probes are used in order to reliably distinguish the subtypes, but these come with additional cost and equipment overheads. Expert pathologists specializing in these types of lymphomas, on the other hand, have shown promise in being able to differentiate these sub-types on H&E, indicating that there is the potential for a DP approach to be employed. A successful approach would allow for more consistent and less demanding diagnosis of this disease. This dataset was created to mirror real-world situations and as such contains samples prepared by different pathologists at different sites. They have additionally selected samples which contain a larger degree of staining variation than one would normally expect [Figure 11]. | Figure 11: Exemplars taken from the (a) chronic lymphocytic leukemia, (b) follicular lymphoma, and (c) mantle cell lymphoma classes used in this task. There is notable staining difference across the three samples. Also, it is not intuitively obvious what the characteristics are which should be used to classify these images
Click here to view |
This use case represents the only classification use case of this manuscript: Attempting to separate images into 1 of 3 sub-types of lymphoma. In the previous tasks, we were looking at primitives and attempting to segmented or detect them. In this case, though, a high-level approach is taken, wherein we provide whole tissue samples to have the DL learn unique features of each class.
Patch selection technique
To generate training patches, a naïve approach was used. Images were split into 36 × 36 sub-patches with a stride of. [32] Caffe has the ability, at training time, to randomly crop out smaller 32 × 32 patches from the larger ones provided, artificially increasing the dataset. This approach could not be used in other tasks because there was no guarantee that the center pixel would retain the appropriate class label (consider an edge pixel of nuclei, an arbitrary translation could potentially change its underlying class). During testing time, patches were extracted using the same methodology, and a voting scheme per subtype was used where votes were aggregated based on the DLs output per patch. In a winner-take-all, the class with the highest number of votes became the designated class for the entire image.
Results and discussion
Each of the 5-fold cross validation sets had 300 training images and 75 test images, for a total of about 825 k training patches. The mean accuracy is 96.58% ±0.01% (on average 2.6 misclassified images in 75 tests). This is over a 10% improvement from the software package, wnd-chrm, [51] where the dataset was also used. Interestingly, both approaches encode no domain knowledge.
In the cases where images were incorrectly classified, there tends to be an overall poor quality of the slide, which would have resulted in either a rescan or a removal. For example, in [Figure 12], the images have significant artifacts which likely caused its misclassification. The voting for these types of images shows 814 patches assigned to the CLL category, 562 patches to the FL category, and 0 patches to the MCL category, a strong indication of uncertainty. When this is compared to other images, for example, in the FL category, the scoring is {5, 1357, 14}, respectively. This indicates that in the case where there is not a landslide voting victory, the slide should be reviewed manually. | Figure 12: (a and b) Misclassified image belonging to the follicular lymphoma subtype. We can see that when magnified, there appears to be some type of artifact created during the scanning process. It is not unreasonable to think that upon seeing this a clinician would ask for it to be rescanned
Click here to view |
| Discussion | | |
There are a few insights which can be gleaned from the experiments involving the use cases. First, there was no situation which dropout had improved the resulting metrics. Srivastava et al. [38] performed rigorous quantitative evaluation identifying dataset sizes which might benefit from dropout. Our datasets are larger than the recommended sizes discussed in their paper and are thus likely large enough that we do not suffer from overfitting. This potentially limited the utility of dropout in our use cases.
Second, it is of the utmost importance to select an appropriate magnification for each task. The rule of thumb we employed is that a human expert should be able to make the correct assessment given only the context presented in a patch. For this reason, tasks such as the epithelium segmentation were performed at very low magnification, while nuclei edge detection was performed at higher magnification. By the same token, if too low of magnification is selected for a task, only a few pixels supply the context needed for the class identification. As a result, the network becomes less powerful as the noncontext pixels have no use and yet still consume input variables.
Third, a majority of the work in this paper focused on finding simple, albeit robust ways of identifying challenging exemplars for training, in other words, those exemplars that would be most informative to the DL network. In situations where random selection was solely utilized, there are too many instances of trivial exemplars that ended up being selected, exemplars that did not enhance the learning capability of the network (e.g., nuclei segmentation task). Another technique for identifying important patches was to use a 2-stage classification stage (i.e., mitosis detection and lymphocyte detection), where false positives and negatives from the first round classifier were oversampled to form the second training set.
In addition, due to the nature of DL, where inferences are derived from data, improving ground truth annotations so that they are precise to the pixel level, would likely further improve the results. [Figure 5] illustrates the difference between a typical manual segmentation of a pathologist versus a pixel level output produced by an algorithm. DL has the potential to provide a first pass ground truth annotation of very high quality, thus allowing domain experts such as pathologists to solely focus on correcting errors made by the DL network.
Finally, while we have mentioned comparable current state-of-the-art metrics where applicable, we note that datasets complexity can vary greatly in digital histopathology, perhaps more so than other domains, making a direct comparison difficult if not impossible unless a single benchmark dataset used. Consequently, we are releasing our datasets and annotations, online for usage, and review by the community in hopes of creating more standardized benchmarks. However, we wish to emphasize that a single unified DL framework that was employed with little to no modifications across a variety of different use cases yielded results that compared favorably with the best-reported results for each of those domains, a remarkable result in light of the fact that little to no domain specific information was invoked.
| Conclusion | | |
We have shown how DL can be a valuable unifying tool for the DP domain due to its innate ability to learn useful features directly from data. Via seven use cases, (a) nuclei segmentation, (b) epithelium segmentation, (c) lymphocyte detection, (d) mitosis detection, and (e) lymphoma classification, we have outlined a guide containing the necessary insights for bridging the current knowledge gap between DL approaches and the DP domain. In particular, we have shown that a common, practical, and publicly available software framework can perform on par, or better, than several state-of-the-art classification approaches for several digital histopathology tasks. Using this tutorial in conjunction with our supplemental online resources, we believe researchers can rapidly augment their current tools by leveraging DL for their specific histological needs.
We do however acknowledge that this tutorial and the associated framework did have some limitations. At test time, using the mean of the output from many rotations of the same patch has been shown to further reduce the variance of the output. [8] Others have shown 15 that training multiple networks, with the same or different architectures, can work well in the form of a consensus of experts voting scheme, as each network is initialized randomly and does not derive the same local minimum.
Given that all of the approaches presented in this tutorial did not explicitly and specifically invoke domain specific information, additional improvements could be made by invoking additional handcrafted or domain pertinent features. For example, the nuclei segmentation approach discussed does not address the need to split clustered cells, but such postprocessing approaches tend to require the well-defined boundaries that our approaches provide. Hand-crafted features can also be used in parallel with DL to improve the quality of the classifier. [50] Conversely, recently presented approaches [52],[53] can potentially reverse engineer DL models to determine what relationships were discovered and thus supply valuable insights to the specific problem domain.
Computational efficiency is also a concern, given such large images. Hierarchical approaches have been discussed which greatly limit the number of patches, which must be classified by the network, improving efficiency. [54] In addition, approaches such as blue-ratio segmentation or color deconvolution could serve as a preprocessing step to identify locations for subsequent application of a DL network, for instance in the detection of nuclei.
The approaches presented here are not intended to be a final ending point towards all histological problems, but a surprisingly robust jumping off point for further research. In fact, given that the mitosis benchmark results, it is evident that a 32 × 32 network is not the optimal framework for all challenges. Yet with the source code and data at hand, it becomes possible to begin training and employing DL networks very rapidly and begin to modulate the approaches as appropriate for task specific settings.
Financial Support and Sponsorship
Research reported in this publication was supported by the National Cancer Institute of the National Institutes of Health under award numbers 1U24CA199374-01, R21CA167811-01, R21CA179327-01; R21CA195152-01 the National Institute of Diabetes and Digestive and Kidney Diseases under award number R01DK098503-02, the DOD Prostate Cancer Synergistic Idea Development Award (PC120857); the DOD Lung Cancer Idea Development New Investigator Award (LC130463), the DOD Prostate Cancer Idea Development Award; the Ohio Third Frontier Technology development Grant, the CTSC Coulter Annual Pilot Grant, the Case Comprehensive Cancer Center Pilot Grant VelaSano Grant from the Cleveland Clinic the Wallace H. Coulter Foundation Program in the Department of Biomedical Engineering at Case Western Reserve University. .
Conflicts of Interest
Dr. Madabhushi is an equity holder in Elucid Bioimaging and in Inspirata Inc.. He is also a scientific advisory consultant for Inspirata Inc and also sits on its scientific advisory board. He is also an equity holder in Inspirata Inc. Additionally his technology has been licensed to Elucid Bioimaging and Inspirata Inc. He is also involved in a NIH U24 grant with PathCore Inc.
| References | | |
| 1. | Gurcan MN, Boucheron LE, Can A, Madabhushi A, Rajpoot NM, Yener B. Histopathological image analysis: A review. IEEE Rev Biomed Eng 2009;2:147-71.  |
| 2. | Veta M, Pluim JP, van Diest PJ, Viergever MA. Breast cancer histopathology image analysis: A review. IEEE Trans Biomed Eng 2014;61:1400-11. |
| 3. | Bhargava R, Madabhushi A. A review of emerging themes in image informatics and molecular analysis for digital pathology. Annu Rev Biomed Eng 2016;18. [Last accessed on 2016 Apr 19]. |
| 4. | Lewis JS Jr., Ali S, Luo J, Thorstad WL, Madabhushi A. A quantitative histomorphometric classifier (QuHbIC) identifies aggressive versus indolent p16-positive oropharyngeal squamous cell carcinoma. Am J Surg Pathol 2014;38:128-37. |
| 5. | Basavanhally A, Feldman M, Shih N, Mies C, Tomaszewski J, Ganesan S, et al. Multi-field-of-view strategy for image-based outcome prediction of multi-parametric estrogen receptor-positive breast cancer histopathology: Comparison to oncotype DX. J Pathol Inform 2011;2:S1. [ PUBMED]  |
| 6. | Veta M, van Diest PJ, Willems SM, Wang H, Madabhushi A, Cruz-Roa A, et al. Assessment of algorithms for mitosis detection in breast cancer histopathology images. Med Image Anal 2015;20:237-48. |
| 7. | Roux L, Racoceanu D, Loménie N, Kulikova M, Irshad H, Klossa J, et al. Mitosis detection in breast cancer histological images An ICPR 2012 contest. J Pathol Inform 2013;4:8. |
| 8. | Ciresan DC, Giusti A, Gambardella LM, Schmidhuber J. Mitosis detection in breast cancer histology images with deep neural networks. Med Image Comput Comput Assist Interv 2013;16(Pt 2):411-8. |
| 9. | Cruz-Roa A, Basavanhally A, González F, Gilmore H, Feldman M, Ganesan S, et al. Automatic detection of invasive ductal carcinoma in whole slide images with convolutional neural networks. In: SPIE Medical Imaging. Vol. 9041. ;2014. p. 904103-904103-15. |
| 10. | Chen T, Chefd′hotel C. Deep learning based automatic immune cell detection for immunohistochemistry images. In: Wu G, Zhang D, Zhou L, editors. Machine Learning in Medical Imaging. (Lecture Notes in Computer Science).Vol. 8679.: Springer International Publishing; 2014. p. 17-24. |
| 11. | Goodfellow IJ, Warde-Farley D, Lamblin P, et al. Pylearn2: A machine learning research library. arXiv preprint arXiv: 1308.4214; 2013. |
| 12. | Bastien F, Lamblin P, Pascanu R, Bergstra J, Goodfellow IJ, Bergeron A, et al. "Theano: New Features and Speed Improvements." Deep Learning and Unsupervised Feature Learning NIPS 2012 Workshop; 2012. |
| 13. | LeCun Y, Bottou L, Orr G, Muller K. Efficient backprop. In: Orr G, Müller KR, editors. Neural Networks: Tricks of the Trade. Springer; 1998. |
| 14. | Montavon G, Orr GB, Müller K, editors. Neural Networks: Tricks of the Trade. (Lecture Notes in Computer Science). 2 nd ed., Vol. 7700. Springer; 2012. |
| 15. | Ciresan D, Giusti A, Gambardella LM, Schmidhuber J. Deep neural networks segment neuronal membranes in electron microscopy images. In: Pereira F, Burges C, Bottou L, Weinberger K, editors. Advances in Neural Information Processing Systems 25. Curran Associates, Inc.; 2012. p. 2843-51. |
| 16. | Kainz P, Pfeiffer M, Urschler M. Semantic Segmentation of Colon Glands with Deep Convolutional Neural Networks and Total Variation Segmentation. CoRR, Vol. abs/1511.06919; 2015. |
| 17. | Maqlin P, Thamburaj R, Mammen J, Manipadam M. Automated nuclear pleomorphism scoring in breast cancer histopathology images using deep neural networks. In: Prasath R, Vuppala AK, Kathirvalavakumar T, editors. Mining Intelligence and Knowledge Exploration. (Lecture Notes in Computer Science). Vol. 9468. Springer International Publishing; 2015. p. 269-76. |
| 18. | Sirinukunwattana K, Raza S, Tsang YW, Snead D, Cree I, Rajpoot N. Locality sensitive deep learning for detection and classification of nuclei in routine colon cancer histology images. IEEE Trans Med Imaging 2016. |
| 19. | Zhou Y, Chang H, Barner KE, Parvin B. Nuclei Segmentation via Sparsity Constrained Convolutional Regression. In: Biomedical Imaging (ISBI), 2015 IEEE 12 th International Symposium on; April, 2015. p. 1284-7. |
| 20. | Xu J, Xiang L, Liu Q, Gilmore H, Wu J, Tang J, et al. Stacked sparse autoencoder (SSAE) for nuclei detection on breast cancer histopathology images. IEEE Trans Med Imaging 2016;35:119-30. |
| 21. | Xu Y, Jia Z, Ai Y, Zhang F, Lai M, Chang EIC . Deep Convolutional Activation Features for Large Scale Brain Tumor Histopathology Image Classification and Segmentation. In: Acoustics, Speech and Signal Processing (ICASSP), 2015 IEEE International Conference on; April, 2015. p. 947-51. |
| 22. | Sirinukunwattana K, Ahmed Raza S, Tsang Y, Snead D, Cree I, Rajpoot N . A spatially constrained deep learning framework for detection of epithelial tumor nuclei in cancer histology images. In: Wu G, Coupé P, Zhan Y, Munsell B, Rueckert D, editors. Patch-Based Techniques in Medical Imaging. Vol. 9467. (Lecture Notes in Computer Science). Springer International Publishing; 2015. p. 154-62. |
| 23. | Xu 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. |
| 24. | Ertosun MG, Rubin DL. Automated grading of gliomas using deep learning in digital pathology images: A modular approach with ensemble of convolutional neural networks. AMIA Annu Symp Proc 2015;2015:1899-908. |
| 25. | Jia Y, Shelhamer E, Donahue J, Karayev S, Long J, Girshick R, et al. Caffe: Convolutional Architecture for Fast Feature Embedding. arXiv preprint arXiv: 1408.5093; 2014. |
| 26. | Krizhevsky A. Convolutional Deep Belief Networks on Cifar-10; 2010. Available from: https://www.cs.toronto.edu/~kriz/conv-cifar10-aug2010.pdf.[Last accessed on 2016 Mar 30]. |
| 27. | Krizhevsky A, Sutskever I, Hinton GE. Imagenet classification with deep convolutional neural networks. In: Pereira F, Burges C, Bottou L, Weinberger K, editors. Advances in Neural Information Processing Systems 25. Curran Associates, Inc.; 2012. p. 1097-105. |
| 28. | Janowczyk A. Deep Learning for Digital Pathology Image Analysis: A Comprehensive Tutorial with Selected Use Cases. Technical Report; 2015. Available from: http://www.andrewjanowczyk.com/deep-learning. [Last accessed on 2016 Mar 30]. |
| 29. | Gurcan MN, Madabhushi A, Rajpoot N. Pattern recognition in histopathological images: An ICPR 2010 contest. In: Ünay D, Çataltepe Z, Aksoy S, editors. Recognizing Patterns in Signals, Speech, Images and Videos. (Lecture Notes in Computer Science).Vol. 6388: Springer Berlin Heidelberg; 2010. p. 226-34. |
| 30. | Doyle S, Feldman M, Tomaszewski J, Madabhushi A. A boosted Bayesian multiresolution classifier for prostate cancer detection from digitized needle biopsies. IEEE Trans Biomed Eng 2012;59:1205-18. |
| 31. | Krizhevsky A. Learning multiple layers of features from tiny images. Technical Report. University of Toronto; 2009. |
| 32. | LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521:436-44. |
| 33. | Lee H, Grosse R, Ranganath R, Ng A. Convolutional Deep Belief Networks for Scalable Unsupervised Learning of Hierarchical Representations. In: Proceedings of the 26 th Annual International Conference on Machine Learning, ICML ′09. New York, USA: ACM; 2009. p. 609-16. |
| 34. | LeCun Y, Kavukcuoglu K, Farabet C. Convolutional Networks and Applications in Vision. In: International Symposium on Circuits and Systems (ISCAS 2010), May 30 - June 2, 2010, Paris, France; 2010. p. 253-6. |
| 35. | Nair V, Hinton GE. Rectified linear units improve restricted boltzmann machines. In: Fürnkranz J, Joachims T, editors. ICML. Omni Press; 2010. p. 807-14. |
| 36. | Dahl GE, Sainath TN, Hinton GE. Improving Deep Neural Networks for LVCSR Using Rectified Linear Units and Dropout. In: IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2013, Vancouver, BC, Canada; 26-31 May, 2013. p. 8609-13. |
| 37. | Glorot X, Bordes A, Bengio Y. Deep sparse rectifier neural networks. In: Gordon GJ, Dunson DB, editors. Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics (AISTATS-11). Vol. 15. Workshop and Conference Proceedings; 2011. p. 315-23. [Journal of Machine Learning Research]. |
| 38. | Srivastava N, Hinton GE, Krizhevsky A, Sutskever I, Salakhutdinov R . Dropout: A simple way to prevent neural networks from overfitting. J Mach Learn Res 2014;15:1929-58. |
| 39. | Duchi JC, Hazan E, Singer Y. Adaptive subgradient methods for online learning and stochastic optimization. J Mach Learn Res 2011;12:2121-59. |
| 40. | Genestie C1, Zafrani B, Asselain B, Fourquet A, Rozan S, Validire P, et al. Comparison of the prognostic value of scarff-bloom-richardson and nottingham histological grades in a series of 825 cases of breast cancer: Major importance of the mitotic count as a component of both grading systems. Anticancer Res 1998;18:571-6. |
| 41. | Humphrey PA. Gleason grading and prognostic factors in carcinoma of the prostate. Mod Pathol 2004;17:292-306. |
| 42. | Irshad H, Veillard A, Roux L, Racoceanu D. Methods for nuclei detection, segmentation, and classification in digital histopathology: A review-current status and future potential. IEEE Rev Biomed Eng 2014;7:97-114. |
| 43. | Ruifrok AC, Johnston DA. Quantification of histochemical staining by color deconvolution. Anal Quant Cytol Histol 2001;23:291-9. |
| 44. | Beck AH, Sangoi AR, Leung S, Marinelli RJ, Nielsen TO, van de Vijver MJ, et al. Systematic analysis of breast cancer morphology uncovers stromal features associated with survival. Sci Transl Med 2011;3:108ra113. |
| 45. | Basavanhally A, Yu E, Xu J, Ganesan S, Feldman M, Tomaszewski J, et al. Incorporating domain knowledge for tubule detection in breast histopathology using o′callaghan neighborhoods. In: SPIE Medical Imaging. (Computer-Aided Diagnosis). Vol. 7963. SPIE; 2011. p. 796310. |
| 46. | Sirinukunwattana K, Snead D, Rajpoot N. A Random Polygons Model of Glandular Structures in Colon Histology Images. In: Biomedical Imaging (ISBI), 2015 IEEE 12 th International Symposium on; April, 2015. p. 1526-9. |
| 47. | Fatakdawala H, Xu J, Basavanhally A, Bhanot G, Ganesan S, Feldman M, et al. Expectation-maximization-driven geodesic active contour with overlap resolution (EMaGACOR): Application to lymphocyte segmentation on breast cancer histopathology. IEEE Trans Biomed Eng 2010;57:1676-89. |
| 48. | Arteta C, Lempitsky V, Noble JA, Zisserman A. Learning to detect cells using non-overlapping extremal regions. In: Ayache N, editor. International Conference on Medical Image Computing and Computer Assisted Intervention. (Lecture Notes in Computer Science). MICCAI, Springer; 2012. p. 348-56. |
| 49. | Chang H, Loss L, Parvin B. Nuclear Segmentation in H and E Sections via Multi-reference Graph-cut (mrgc). International Symposium Biomedical Imaging; 2012. |
| 50. | Wang H, Cruz-Roa A, Basavanhally A, Gilmore H, Shih N, Feldman M, et al. Mitosis detection in breast cancer pathology images by combining handcrafted and convolutional neural network features. J Med Imaging (Bellingham) 2014;1:034003. |
| 51. | Orlov N, Shamir L, Macura T, Johnston J, Eckley DM, Goldberg IG. WND-CHARM: Multi-purpose image classification using compound image transforms. Pattern Recognit Lett 2008;29:1684-93. |
| 52. | Zeiler MD, Fergus R. Visualizing and Understanding Convolutional Networks. In: Computer Vision - ECCV 2014 - 13 th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I; 2014. p. 818-33. |
| 53. | Erhan D, Bengio Y, Courville A, Vincent P. Visualizing Higher-layer Features of a Deep Network. Tech. Rep. 1341, University of Montreal, June 2009. ICML 2009 Workshop on Learning Feature Hierarchies, Montréal, Canada; 2009. |
| 54. | Janowczyk A, Doyle S, Gilmore H, Madabhushi A. A resolution adaptive deep hierarchical (radhical) learning scheme applied to nuclear segmentation of digital pathology images. In: Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization. 2016. |
[Figure 1], [Figure 2], [Figure 3], [Figure 4], [Figure 5], [Figure 6], [Figure 7], [Figure 8], [Figure 9], [Figure 10], [Figure 11], [Figure 12]
[Table 1], [Table 2], [Table 3], [Table 4], [Table 5], [Table 6]
| This article has been cited by | | 1 |
Self supervised contrastive learning for digital histopathology |
|
| Ozan Ciga, Tony Xu, Anne Louise Martel | | Machine Learning with Applications. 2022; 7: 100198 | | [Pubmed] | [DOI] | | | 2 |
Memory augmented convolutional neural network and its application in bioimages |
|
| Weiping Ding, Yurui Ming, Yu-Kai Wang, Chin-Teng Lin | | Neurocomputing. 2021; 466: 128 | | [Pubmed] | [DOI] | | | 3 |
Deep computational pathology in breast cancer |
|
| Andrea Duggento, Allegra Conti, Alessandro Mauriello, Maria Guerrisi, Nicola Toschi | | Seminars in Cancer Biology. 2021; 72: 226 | | [Pubmed] | [DOI] | | | 4 |
Development and evaluation of deep learning–based segmentation of histologic structures in the kidney cortex with multiple histologic stains |
|
| Catherine P. Jayapandian, Yijiang Chen, Andrew R. Janowczyk, Matthew B. Palmer, Clarissa A. Cassol, Miroslav Sekulic, Jeffrey B. Hodgin, Jarcy Zee, Stephen M. Hewitt, John O’Toole, Paula Toro, John R. Sedor, Laura Barisoni, Anant Madabhushi, J. Sedor, K. Dell, M. Schachere, J. Negrey, K. Lemley, E. Lim, T. Srivastava, A. Garrett, C. Sethna, K. Laurent, G. Appel, M. Toledo, L. Barisoni, L. Greenbaum, C. Wang, C. Kang, S. Adler, C. Nast, J. LaPage, John H. Stroger, A. Athavale, M. Itteera, A. Neu, S. Boynton, F. Fervenza, M. Hogan, J. Lieske, V. Chernitskiy, F. Kaskel, N. Kumar, P. Flynn, J. Kopp, J. Blake, H. Trachtman, O. Zhdanova, F. Modersitzki, S. Vento, R. Lafayette, K. Mehta, C. Gadegbeku, D. Johnstone, S. Quinn-Boyle, D. Cattran, M. Hladunewich, H. Reich, P. Ling, M. Romano, A. Fornoni, C. Bidot, M. Kretzler, D. Gipson, A. Williams, J. LaVigne, V. Derebail, K. Gibson, A. Froment, S. Grubbs, L. Holzman, K. Meyers, K. Kallem, J. Lalli, K. Sambandam, Z. Wang, M. Rogers, A. Jefferson | | Kidney International. 2021; 99(1): 86 | | [Pubmed] | [DOI] | | | 5 |
Deep learning identified pathological abnormalities predictive of graft loss in kidney transplant biopsies |
|
| Zhengzi Yi, Fadi Salem, Madhav C. Menon, Karen Keung, Caixia Xi, Sebastian Hultin, M. Rizwan Haroon Al Rasheed, Li Li, Fei Su, Zeguo Sun, Chengguo Wei, Weiqing Huang, Samuel Fredericks, Qisheng Lin, Khadija Banu, Germaine Wong, Natasha M. Rogers, Samira Farouk, Paolo Cravedi, Meena Shingde, R. Neal Smith, Ivy A. Rosales, Philip J. O’Connell, Robert B. Colvin, Barbara Murphy, Weijia Zhang | | Kidney International. 2021; | | [Pubmed] | [DOI] | | | 6 |
PAIP 2019: Liver cancer segmentation challenge |
|
| Yoo Jung Kim, Hyungjoon Jang, Kyoungbun Lee, Seongkeun Park, Sung-Gyu Min, Choyeon Hong, Jeong Hwan Park, Kanggeun Lee, Jisoo Kim, Wonjae Hong, Hyun Jung, Yanling Liu, Haran Rajkumar, Mahendra Khened, Ganapathy Krishnamurthi, Sen Yang, Xiyue Wang, Chang Hee Han, Jin Tae Kwak, Jianqiang Ma, Zhe Tang, Bahram Marami, Jack Zeineh, Zixu Zhao, Pheng-Ann Heng, Rüdiger Schmitz, Frederic Madesta, Thomas Rösch, Rene Werner, Jie Tian, Elodie Puybareau, Matteo Bovio, Xiufeng Zhang, Yifeng Zhu, Se Young Chun, Won-Ki Jeong, Peom Park, Jinwook Choi | | Medical Image Analysis. 2021; 67: 101854 | | [Pubmed] | [DOI] | | | 7 |
Residual cyclegan for robust domain transformation of histopathological tissue slides |
|
| Thomas de Bel, John-Melle Bokhorst, Jeroen van der Laak, Geert Litjens | | Medical Image Analysis. 2021; 70: 102004 | | [Pubmed] | [DOI] | | | 8 |
Fine-Tuning and training of densenet for histopathology image representation using TCGA diagnostic slides |
|
| Abtin Riasatian, Morteza Babaie, Danial Maleki, Shivam Kalra, Mojtaba Valipour, Sobhan Hemati, Manit Zaveri, Amir Safarpoor, Sobhan Shafiei, Mehdi Afshari, Maral Rasoolijaberi, Milad Sikaroudi, Mohd Adnan, Sultaan Shah, Charles Choi, Savvas Damaskinos, Clinton JV Campbell, Phedias Diamandis, Liron Pantanowitz, Hany Kashani, Ali Ghodsi, H.R. Tizhoosh | | Medical Image Analysis. 2021; 70: 102032 | | [Pubmed] | [DOI] | | | 9 |
Artificial intelligence applied to breast pathology |
|
| Mustafa Yousif, Paul J. van Diest, Arvydas Laurinavicius, David Rimm, Jeroen van der Laak, Anant Madabhushi, Stuart Schnitt, Liron Pantanowitz | | Virchows Archiv. 2021; | | [Pubmed] | [DOI] | | | 10 |
The role of machine learning in cardiovascular pathology |
|
| Carolyn Glass, Kyle J. Lafata, William Jeck, Roarke Horstmeyer, Colin Cooke, Jeffrey Everitt, Matthew Glass, David Dov, Michael A. Seidman | | Canadian Journal of Cardiology. 2021; | | [Pubmed] | [DOI] | | | 11 |
Impact of stain normalization and patch selection on the performance of convolutional neural networks in histological breast and prostate cancer classification |
|
| Massimo Salvi, Filippo Molinari, U Rajendra Acharya, Luca Molinaro, Kristen M. Meiburger | | Computer Methods and Programs in Biomedicine Update. 2021; 1: 100004 | | [Pubmed] | [DOI] | | | 12 |
The impact of pre- and post-image processing techniques on deep learning frameworks: A comprehensive review for digital pathology image analysis |
|
| Massimo Salvi, U. Rajendra Acharya, Filippo Molinari, Kristen M. Meiburger | | Computers in Biology and Medicine. 2021; 128: 104129 | | [Pubmed] | [DOI] | | | 13 |
Mitosis detection techniques in H&E stained breast cancer pathological images: A comprehensive review |
|
| Xipeng Pan, Yinghua Lu, Rushi Lan, Zhenbing Liu, Zujun Qin, Huadeng Wang, Zaiyi Liu | | Computers & Electrical Engineering. 2021; 91: 107038 | | [Pubmed] | [DOI] | | | 14 |
Deep learning powers cancer diagnosis in digital pathology |
|
| Yunjie He, Hong Zhao, Stephen T.C. Wong | | Computerized Medical Imaging and Graphics. 2021; 88: 101820 | | [Pubmed] | [DOI] | | | 15 |
PMNet: A probability map based scaled network for breast cancer diagnosis |
|
| Salman Ahmed, Maria Tariq, Hammad Naveed | | Computerized Medical Imaging and Graphics. 2021; 89: 101863 | | [Pubmed] | [DOI] | | | 16 |
A U-Net based framework to quantify glomerulosclerosis in digitized PAS and H&E stained human tissues |
|
| Jaime Gallego, Zaneta Swiderska-Chadaj, Tomasz Markiewicz, Michifumi Yamashita, M. Alejandra Gabaldon, Arkadiusz Gertych | | Computerized Medical Imaging and Graphics. 2021; 89: 101865 | | [Pubmed] | [DOI] | | | 17 |
A deep learning based multiscale approach to segment the areas of interest in whole slide images |
|
| Yanbo Feng, Adel Hafiane, Hélčne Laurent | | Computerized Medical Imaging and Graphics. 2021; 90: 101923 | | [Pubmed] | [DOI] | | | 18 |
Automated assessment of glomerulosclerosis and tubular atrophy using deep learning |
|
| Massimo Salvi, Alessandro Mogetta, Alessandro Gambella, Luca Molinaro, Antonella Barreca, Mauro Papotti, Filippo Molinari | | Computerized Medical Imaging and Graphics. 2021; 90: 101930 | | [Pubmed] | [DOI] | | | 19 |
Fractal Neural Network: A new ensemble of fractal geometry and convolutional neural networks for the classification of histology images |
|
| Guilherme Freire Roberto, Alessandra Lumini, Leandro Alves Neves, Marcelo Zanchetta do Nascimento | | Expert Systems with Applications. 2021; 166: 114103 | | [Pubmed] | [DOI] | | | 20 |
Application of Single-Cell Approaches to Study Myeloproliferative Neoplasm Biology |
|
| Daniel Royston, Adam J. Mead, Bethan Psaila | | Hematology/Oncology Clinics of North America. 2021; 35(2): 279 | | [Pubmed] | [DOI] | | | 21 |
Quick Annotator: an open-source digital pathology based rapid image annotation tool |
|
| Runtian Miao, Robert Toth, Yu Zhou, Anant Madabhushi, Andrew Janowczyk | | The Journal of Pathology: Clinical Research. 2021; 7(6): 542 | | [Pubmed] | [DOI] | | | 22 |
In Situ Classification of Cell Types in Human Kidney Tissue Using 3D Nuclear Staining |
|
| Andre Woloshuk, Suraj Khochare, Aljohara F. Almulhim, Andrew T. McNutt, Dawson Dean, Daria Barwinska, Michael J. Ferkowicz, Michael T. Eadon, Katherine J. Kelly, Kenneth W. Dunn, Mohammad A. Hasan, Tarek M. El-Achkar, Seth Winfree | | Cytometry Part A. 2021; 99(7): 707 | | [Pubmed] | [DOI] | | | 23 |
Quantitative neurotoxicology: Potential role of artificial intelligence/deep learning approach |
|
| Anshul Srivastava, Joseph P. Hanig | | Journal of Applied Toxicology. 2021; 41(7): 996 | | [Pubmed] | [DOI] | | | 24 |
Automated detection and delineation of hepatocellular carcinoma on multiphasic contrast-enhanced MRI using deep learning |
|
| Khaled Bousabarah, Brian Letzen, Jonathan Tefera, Lynn Savic, Isabel Schobert, Todd Schlachter, Lawrence H. Staib, Martin Kocher, Julius Chapiro, MingDe Lin | | Abdominal Radiology. 2021; 46(1): 216 | | [Pubmed] | [DOI] | | | 25 |
Deep learning system for lymph node quantification and metastatic cancer identification from whole-slide pathology images |
|
| Yajie Hu, Feng Su, Kun Dong, Xinyu Wang, Xinya Zhao, Yumeng Jiang, Jianming Li, Jiafu Ji, Yu Sun | | Gastric Cancer. 2021; 24(4): 868 | | [Pubmed] | [DOI] | | | 26 |
A comprehensive survey of image segmentation: clustering methods, performance parameters, and benchmark datasets |
|
| Himanshu Mittal, Avinash Chandra Pandey, Mukesh Saraswat, Sumit Kumar, Raju Pal, Garv Modwel | | Multimedia Tools and Applications. 2021; | | [Pubmed] | [DOI] | | | 27 |
Automated detection of COVID-19 using ensemble of transfer learning with deep convolutional neural network based on CT scans |
|
| Parisa gifani, Ahmad Shalbaf, Majid Vafaeezadeh | | International Journal of Computer Assisted Radiology and Surgery. 2021; 16(1): 115 | | [Pubmed] | [DOI] | | | 28 |
Breast Cancer Detection, Segmentation and Classification on Histopathology Images Analysis: A Systematic Review |
|
| R. Krithiga, P. Geetha | | Archives of Computational Methods in Engineering. 2021; 28(4): 2607 | | [Pubmed] | [DOI] | | | 29 |
A Comprehensive Review of Markov Random Field and Conditional Random Field Approaches in Pathology Image Analysis |
|
| Yixin Li, Chen Li, Xiaoyan Li, Kai Wang, Md Mamunur Rahaman, Changhao Sun, Hao Chen, Xinran Wu, Hong Zhang, Qian Wang | | Archives of Computational Methods in Engineering. 2021; | | [Pubmed] | [DOI] | | | 30 |
A review on deep learning in medical image analysis |
|
| S. Suganyadevi, V. Seethalakshmi, K. Balasamy | | International Journal of Multimedia Information Retrieval. 2021; | | [Pubmed] | [DOI] | | | 31 |
An optimal nuclei segmentation method based on enhanced multi-objective GWO |
|
| Ravi Sharma, Kapil Sharma | | Complex & Intelligent Systems. 2021; | | [Pubmed] | [DOI] | | | 32 |
Generative Deep Learning in Digital Pathology Workflows |
|
| David Morrison, David Harris-Birtill, Peter D. Caie | | The American Journal of Pathology. 2021; 191(10): 1717 | | [Pubmed] | [DOI] | | | 33 |
A deep learning approach for mitosis detection: Application in tumor proliferation prediction from whole slide images |
|
| Ramin Nateghi, Habibollah Danyali, Mohammad Sadegh Helfroush | | Artificial Intelligence in Medicine. 2021; 114: 102048 | | [Pubmed] | [DOI] | | | 34 |
A hybrid deep learning approach for gland segmentation in prostate histopathological images |
|
| Massimo Salvi, Martino Bosco, Luca Molinaro, Alessandro Gambella, Mauro Papotti, U. Rajendra Acharya, Filippo Molinari | | Artificial Intelligence in Medicine. 2021; 115: 102076 | | [Pubmed] | [DOI] | | | 35 |
A financial statement fraud model based on synthesized attribute selection and a dataset with missing values and imbalanced classes |
|
| Ching-Hsue Cheng, Yung-Fu Kao, Hsien-Ping Lin | | Applied Soft Computing. 2021; 108: 107487 | | [Pubmed] | [DOI] | | | 36 |
Closing the translation gap: AI applications in digital pathology |
|
| David F. Steiner, Po-Hsuan Cameron Chen, Craig H. Mermel | | Biochimica et Biophysica Acta (BBA) - Reviews on Cancer. 2021; 1875(1): 188452 | | [Pubmed] | [DOI] | | | 37 |
A dataset and a methodology for intraoperative computer-aided diagnosis of a metastatic colon cancer in a liver |
|
| Dario Sitnik, Gorana Aralica, Mirko Hadžija, Marijana Popovic Hadžija, Arijana Pacic, Marija Milkovic Periša, Luka Manojlovic, Karolina Krstanac, Andrija Plavetic, Ivica Kopriva | | Biomedical Signal Processing and Control. 2021; 66: 102402 | | [Pubmed] | [DOI] | | | 38 |
Digital pathology and artificial intelligence in translational medicine and clinical practice |
|
| Vipul Baxi, Robin Edwards, Michael Montalto, Saurabh Saha | | Modern Pathology. 2021; | | [Pubmed] | [DOI] | | | 39 |
Deep learning in cancer pathology: a new generation of clinical biomarkers |
|
| Amelie Echle, Niklas Timon Rindtorff, Titus Josef Brinker, Tom Luedde, Alexander Thomas Pearson, Jakob Nikolas Kather | | British Journal of Cancer. 2021; 124(4): 686 | | [Pubmed] | [DOI] | | | 40 |
A multi-phase deep CNN based mitosis detection framework for breast cancer histopathological images |
|
| Anabia Sohail, Asifullah Khan, Noorul Wahab, Aneela Zameer, Saranjam Khan | | Scientific Reports. 2021; 11(1) | | [Pubmed] | [DOI] | | | 41 |
A review and comparison of breast tumor cell nuclei segmentation performances using deep convolutional neural networks |
|
| Andrew Lagree, Majidreza Mohebpour, Nicholas Meti, Khadijeh Saednia, Fang-I. Lu, Elzbieta Slodkowska, Sonal Gandhi, Eileen Rakovitch, Alex Shenfield, Ali Sadeghi-Naini, William T. Tran | | Scientific Reports. 2021; 11(1) | | [Pubmed] | [DOI] | | | 42 |
A generalized deep learning framework for whole-slide image segmentation and analysis |
|
| Mahendra Khened, Avinash Kori, Haran Rajkumar, Ganapathy Krishnamurthi, Balaji Srinivasan | | Scientific Reports. 2021; 11(1) | | [Pubmed] | [DOI] | | | 43 |
CAD systems for colorectal cancer from WSI are still not ready for clinical acceptance |
|
| Sara P. Oliveira, Pedro C. Neto, Joăo Fraga, Diana Montezuma, Ana Monteiro, Joăo Monteiro, Liliana Ribeiro, Sofia Gonçalves, Isabel M. Pinto, Jaime S. Cardoso | | Scientific Reports. 2021; 11(1) | | [Pubmed] | [DOI] | | | 44 |
Sliding window based deep ensemble system for breast cancer classification |
|
| Amin Alqudah, Ali Mohammad Alqudah | | Journal of Medical Engineering & Technology. 2021; 45(4): 313 | | [Pubmed] | [DOI] | | | 45 |
An empirical analysis of machine learning frameworks for digital pathology in medical science |
|
| S.K.B. Sangeetha, R Dhaya, Dhruv T Shah, R Dharanidharan, K. Praneeth Sai Reddy | | Journal of Physics: Conference Series. 2021; 1767(1): 012031 | | [Pubmed] | [DOI] | | | 46 |
High-performance deep learning pipeline predicts individuals in mixtures of DNA using sequencing data |
|
| Nam Nhut Phan, Amrita Chattopadhyay, Tsui-Ting Lee, Hsiang-I Yin, Tzu-Pin Lu, Liang-Chuan Lai, Hsiao-Lin Hwa, Mong-Hsun Tsai, Eric Y Chuang | | Briefings in Bioinformatics. 2021; 22(6) | | [Pubmed] | [DOI] | | | 47 |
Artificial Intelligence and Mapping a New Direction in Laboratory Medicine: A Review |
|
| Daniel S Herman, Daniel D Rhoads, Wade L Schulz, Thomas J S Durant | | Clinical Chemistry. 2021; 67(11): 1466 | | [Pubmed] | [DOI] | | | 48 |
An automated computational image analysis pipeline for histological grading of cardiac allograft rejection |
|
| Eliot G Peyster, Sara Arabyarmohammadi, Andrew Janowczyk, Sepideh Azarianpour-Esfahani, Miroslav Sekulic, Clarissa Cassol, Luke Blower, Anil Parwani, Priti Lal, Michael D Feldman, Kenneth B Margulies, Anant Madabhushi | | European Heart Journal. 2021; 42(24): 2356 | | [Pubmed] | [DOI] | | | 49 |
Deep Learning-Based Image Classification in Differentiating Tufted Astrocytes, Astrocytic Plaques, and Neuritic Plaques |
|
| Shunsuke Koga, Nikhil B Ghayal, Dennis W Dickson | | Journal of Neuropathology & Experimental Neurology. 2021; 80(4): 306 | | [Pubmed] | [DOI] | | | 50 |
Artificial intelligence neuropathologist for glioma classification using deep learning on hematoxylin and eosin stained slide images and molecular markers |
|
| Lei Jin, Feng Shi, Qiuping Chun, Hong Chen, Yixin Ma, Shuai Wu, N U Farrukh Hameed, Chunming Mei, Junfeng Lu, Jun Zhang, Abudumijiti Aibaidula, Dinggang Shen, Jinsong Wu | | Neuro-Oncology. 2021; 23(1): 44 | | [Pubmed] | [DOI] | | | 51 |
An Advanced Automated Image Analysis Model for Scoring of ER, PR, HER-2 and Ki-67 in Breast Carcinoma |
|
| Min Feng, Jie Chen, Xuhui Xiang, Yang Deng, Yanyan Zhou, Zhang Zhang, Zhongxi Zheng, Ji Bao, Hong Bu | | IEEE Access. 2021; 9: 108441 | | [Pubmed] | [DOI] | | | 52 |
Automatic Detection of Invasive Ductal Carcinoma Based on the Fusion of Multi-Scale Residual Convolutional Neural Network and SVM |
|
| Jianfei Zhang, Xiaoyan Guo, Bo Wang, Wensheng Cui | | IEEE Access. 2021; 9: 40308 | | [Pubmed] | [DOI] | | | 53 |
Multi-Task Pre-Training of Deep Neural Networks for Digital Pathology |
|
| Romain Mormont, Pierre Geurts, Raphael Maree | | IEEE Journal of Biomedical and Health Informatics. 2021; 25(2): 412 | | [Pubmed] | [DOI] | | | 54 |
A Visually Interpretable Deep Learning Framework for Histopathological Image-Based Skin Cancer Diagnosis |
|
| Shancheng Jiang, Huichuan Li, Zhi Jin | | IEEE Journal of Biomedical and Health Informatics. 2021; 25(5): 1483 | | [Pubmed] | [DOI] | | | 55 |
A Review of Deep Learning in Medical Imaging: Imaging Traits, Technology Trends, Case Studies With Progress Highlights, and Future Promises |
|
| S. Kevin Zhou, Hayit Greenspan, Christos Davatzikos, James S. Duncan, Bram Van Ginneken, Anant Madabhushi, Jerry L. Prince, Daniel Rueckert, Ronald M. Summers | | Proceedings of the IEEE. 2021; 109(5): 820 | | [Pubmed] | [DOI] | | | 56 |
Small Blob Detector Using Bi-Threshold Constrained Adaptive Scales |
|
| Yanzhe Xu, Teresa Wu, Jennifer R. Charlton, Fei Gao, Kevin M. Bennett | | IEEE Transactions on Biomedical Engineering. 2021; 68(9): 2654 | | [Pubmed] | [DOI] | | | 57 |
Visual Analytics for Hypothesis-Driven Exploration in Computational Pathology |
|
| A. Corvo, H. S. Garcia Caballero, M. A. Westenberg, M. A. van Driel, J. J. van Wijk | | IEEE Transactions on Visualization and Computer Graphics. 2021; 27(10): 3851 | | [Pubmed] | [DOI] | | | 58 |
Recent technical advances in whole slide imaging instrumentation |
|
| Prateek Katare, Sai Siva Gorthi | | Journal of Microscopy. 2021; 284(2): 103 | | [Pubmed] | [DOI] | | | 59 |
Deep learning-based model for diagnosing Alzheimer's disease and tauopathies |
|
| Shunsuke Koga, Akihiro Ikeda, Dennis W. Dickson | | Neuropathology and Applied Neurobiology. 2021; | | [Pubmed] | [DOI] | | | 60 |
Current and future applications of artificial intelligence in pathology: a clinical perspective |
|
| Emad A Rakha, Michael Toss, Sho Shiino, Paul Gamble, Ronnachai Jaroensri, Craig H Mermel, Po-Hsuan Cameron Chen | | Journal of Clinical Pathology. 2021; 74(7): 409 | | [Pubmed] | [DOI] | | | 61 |
A comparative study on machine learning-based classification to find photothrombotic lesion in histological rabbit brain images |
|
| Sang Hee Jo, Yoonhee Kim, Yoon Bum Lee, Sung Suk Oh, Jong-ryul Choi | | Journal of Innovative Optical Health Sciences. 2021; 14(06) | | [Pubmed] | [DOI] | | | 62 |
A Calibrated Multiexit Neural Network for Detecting Urothelial Cancer Cells |
|
| L. Lilli, E. Giarnieri, S. Scardapane, Nadia A. Chuzhanova | | Computational and Mathematical Methods in Medicine. 2021; 2021: 1 | | [Pubmed] | [DOI] | | | 63 |
DICOM Format and Protocol Standardization—A Core Requirement for Digital Pathology Success |
|
| David A. Clunie | | Toxicologic Pathology. 2021; 49(4): 738 | | [Pubmed] | [DOI] | | | 64 |
Using Deep Learning Artificial Intelligence Algorithms to Verify N-Nitroso-N-Methylurea and Urethane Positive Control Proliferative Changes in Tg-RasH2 Mouse Carcinogenicity Studies |
|
| Daniel Rudmann, Jay Albretsen, Colin Doolan, Mark Gregson, Beth Dray, Aaron Sargeant, Donal O’Shea D, Jogile Kuklyte, Adam Power, Jenny Fitzgerald | | Toxicologic Pathology. 2021; 49(4): 938 | | [Pubmed] | [DOI] | | | 65 |
Deep Learning in Toxicologic Pathology: A New Approach to Evaluate Rodent Retinal Atrophy |
|
| Maria Cristina De Vera Mudry, Jim Martin, Vanessa Schumacher, Raghavan Venugopal | | Toxicologic Pathology. 2021; 49(4): 851 | | [Pubmed] | [DOI] | | | 66 |
HistoNet: A Deep Learning-Based Model of Normal Histology |
|
| Holger Hoefling, Tobias Sing, Imtiaz Hossain, Julie Boisclair, Arno Doelemeyer, Thierry Flandre, Alessandro Piaia, Vincent Romanet, Gianluca Santarossa, Chandrassegar Saravanan, Esther Sutter, Oliver Turner, Kuno Wuersch, Pierre Moulin | | Toxicologic Pathology. 2021; 49(4): 784 | | [Pubmed] | [DOI] | | | 67 |
Retrospective analysis and time series forecasting with automated machine learning of ascariasis, enterobiasis and cystic echinococcosis in Romania |
|
| Johannes Benecke, Cornelius Benecke, Marius Ciutan, Mihnea Dosius, Cristian Vladescu, Victor Olsavszky, Kate Zinszer | | PLOS Neglected Tropical Diseases. 2021; 15(11): e0009831 | | [Pubmed] | [DOI] | | | 68 |
Deep Learning Model for Cell Nuclei Segmentation and Lymphocyte Identification in Whole Slide Histology Images |
|
| Elzbieta Budginaite, Mindaugas Morkunas, Arvydas Laurinavicius, Povilas Treigys | | Informatica. 2021; : 23 | | [Pubmed] | [DOI] | | | 69 |
Artificial intelligence for melanoma diagnosis |
|
| Philipp TSCHANDL | | Italian Journal of Dermatology and Venereology. 2021; 156(3) | | [Pubmed] | [DOI] | | | 70 |
An efficient glomerular object locator for renal whole slide images using proposal-free network and dynamic scale evaluation method |
|
| Xueyu Liu, Ming Li, Yongfei Wu, Yilin Chen, Fang Hao, Daoxiang Zhou, Chen Wang, Chuanfeng Ma, Guangze Shi, Xiaoshuang Zhou | | AI Communications. 2021; : 1 | | [Pubmed] | [DOI] | | | 71 |
Classification of Invasive Ductal Carcinoma from histopathology breast cancer images using Stacked Generalized Ensemble |
|
| Deepika Kumar, Usha Batra | | Journal of Intelligent & Fuzzy Systems. 2021; 40(3): 4919 | | [Pubmed] | [DOI] | | | 72 |
Detecting breast cancer using artificial intelligence: Convolutional neural network |
|
| Avishek Choudhury, Sunanda Perumalla | | Technology and Health Care. 2021; 29(1): 33 | | [Pubmed] | [DOI] | | | 73 |
Improved Classification of Cancerous Histopathology Images using
Color Channel Separation and Deep Learning |
|
| Rachit Kumar Gupta, Jatinder Manhas | | Journal of Multimedia Information System. 2021; 8(3): 175 | | [Pubmed] | [DOI] | | | 74 |
Multi_Scale_Tools: A Python Library to Exploit Multi-Scale Whole Slide Images |
|
| Niccolň Marini, Sebastian Otálora, Damian Podareanu, Mart van Rijthoven, Jeroen van der Laak, Francesco Ciompi, Henning Müller, Manfredo Atzori | | Frontiers in Computer Science. 2021; 3 | | [Pubmed] | [DOI] | | | 75 |
Deep Learning in Head and Neck Tumor Multiomics Diagnosis and Analysis: Review of the Literature |
|
| Xi Wang, Bin-bin Li | | Frontiers in Genetics. 2021; 12 | | [Pubmed] | [DOI] | | | 76 |
Prediction of Breast Cancer Recurrence Using a Deep Convolutional Neural Network Without Region-of-Interest Labeling |
|
| Nam Nhut Phan, Chih-Yi Hsu, Chi-Cheng Huang, Ling-Ming Tseng, Eric Y. Chuang | | Frontiers in Oncology. 2021; 11 | | [Pubmed] | [DOI] | | | 77 |
Detection of Metastatic Tumor Cells in the Bone Marrow Aspirate Smears by Artificial Intelligence (AI)-Based Morphogo System |
|
| Pu Chen, Run Chen Xu, Nan Chen, Lan Zhang, Li Zhang, Jianfeng Zhu, Baishen Pan, Beili Wang, Wei Guo | | Frontiers in Oncology. 2021; 11 | | [Pubmed] | [DOI] | | | 78 |
Deep Learning of Histopathology Images at the Single Cell Level |
|
| Kyubum Lee, John H. Lockhart, Mengyu Xie, Ritu Chaudhary, Robbert J. C. Slebos, Elsa R. Flores, Christine H. Chung, Aik Choon Tan | | Frontiers in Artificial Intelligence. 2021; 4 | | [Pubmed] | [DOI] | | | 79 |
Histopathological Classification of Canine Cutaneous Round Cell Tumors Using Deep Learning: A Multi-Center Study |
|
| Massimo Salvi, Filippo Molinari, Selina Iussich, Luisa Vera Muscatello, Luca Pazzini, Silvia Benali, Barbara Banco, Francesca Abramo, Raffaella De Maria, Luca Aresu | | Frontiers in Veterinary Science. 2021; 8 | | [Pubmed] | [DOI] | | | 80 |
A Transfer Learning Architecture Based on a Support Vector Machine for Histopathology Image Classification |
|
| Jiayi Fan, JangHyeon Lee, YongKeun Lee | | Applied Sciences. 2021; 11(14): 6380 | | [Pubmed] | [DOI] | | | 81 |
Shifting Gears in Precision Oncology—Challenges and Opportunities of Integrative Data Analysis |
|
| Ka-Won Noh, Reinhard Buettner, Sebastian Klein | | Biomolecules. 2021; 11(9): 1310 | | [Pubmed] | [DOI] | | | 82 |
Klasifikasi Citra Histopatologi Kanker Payudara menggunakan Data Resampling Random dan Residual Network |
|
| Wahyudi Setiawan | | JURNAL SISTEM INFORMASI BISNIS. 2021; 11(1): 70 | | [Pubmed] | [DOI] | | | 83 |
Deep Learning for the Classification of Non-Hodgkin Lymphoma on Histopathological Images |
|
| Georg Steinbuss, Mark Kriegsmann, Christiane Zgorzelski, Alexander Brobeil, Benjamin Goeppert, Sascha Dietrich, Gunhild Mechtersheimer, Katharina Kriegsmann | | Cancers. 2021; 13(10): 2419 | | [Pubmed] | [DOI] | | | 84 |
A Review of Computer-Aided Expert Systems for Breast Cancer Diagnosis |
|
| Xin Yu Liew, Nazia Hameed, Jeremie Clos | | Cancers. 2021; 13(11): 2764 | | [Pubmed] | [DOI] | | | 85 |
A Deep Neural Network-Based Method for Prediction of Dementia Using Big Data |
|
| Jungyoon Kim, Jihye Lim | | International Journal of Environmental Research and Public Health. 2021; 18(10): 5386 | | [Pubmed] | [DOI] | | | 86 |
Deep Learning in Pancreatic Tissue: Identification of Anatomical Structures, Pancreatic Intraepithelial Neoplasia, and Ductal Adenocarcinoma |
|
| Mark Kriegsmann, Katharina Kriegsmann, Georg Steinbuss, Christiane Zgorzelski, Anne Kraft, Matthias M. Gaida | | International Journal of Molecular Sciences. 2021; 22(10): 5385 | | [Pubmed] | [DOI] | | | 87 |
Quantification of the Immune Content in Neuroblastoma: Deep Learning and Topological Data Analysis in Digital Pathology |
|
| Nicole Bussola, Bruno Papa, Ombretta Melaiu, Aurora Castellano, Doriana Fruci, Giuseppe Jurman | | International Journal of Molecular Sciences. 2021; 22(16): 8804 | | [Pubmed] | [DOI] | | | 88 |
Biomedical Image Processing and Classification |
|
| Luca Mesin | | Electronics. 2021; 10(1): 66 | | [Pubmed] | [DOI] | | | 89 |
Deep Learning Techniques for the Classification of Colorectal Cancer Tissue |
|
| Min-Jen Tsai, Yu-Han Tao | | Electronics. 2021; 10(14): 1662 | | [Pubmed] | [DOI] | | | 90 |
Deeply Supervised UNet for Semantic Segmentation to Assist Dermatopathological Assessment of Basal Cell Carcinoma |
|
| Jean Le’Clerc Arrastia, Nick Heilenkötter, Daniel Otero Baguer, Lena Hauberg-Lotte, Tobias Boskamp, Sonja Hetzer, Nicole Duschner, Jörg Schaller, Peter Maass | | Journal of Imaging. 2021; 7(4): 71 | | [Pubmed] | [DOI] | | | 91 |
A Bidirectional Long Short-Term Memory Model Algorithm for Predicting COVID-19 in Gulf Countries |
|
| Theyazn H. H. Aldhyani, Hasan Alkahtani | | Life. 2021; 11(11): 1118 | | [Pubmed] | [DOI] | | | 92 |
On the Scale Invariance in State of the Art CNNs Trained on ImageNet |
|
| Mara Graziani, Thomas Lompech, Henning Müller, Adrien Depeursinge, Vincent Andrearczyk | | Machine Learning and Knowledge Extraction. 2021; 3(2): 374 | | [Pubmed] | [DOI] | | | 93 |
Classification of Diffuse Glioma Subtype from Clinical-Grade Pathological Images Using Deep Transfer Learning |
|
| Sanghyuk Im, Jonghwan Hyeon, Eunyoung Rha, Janghyeon Lee, Ho-Jin Choi, Yuchae Jung, Tae-Jung Kim | | Sensors. 2021; 21(10): 3500 | | [Pubmed] | [DOI] | | | 94 |
TISSUE CLASSIFICATION FOR COLORECTAL CANCER UTILIZING TECHNIQUES OF DEEP LEARNING AND MACHINE LEARNING |
|
| Kasikrit Damkliang, Thakerng Wongsirichot, Paramee Thongsuksai | | Biomedical Engineering: Applications, Basis and Communications. 2021; 33(03): 2150022 | | [Pubmed] | [DOI] | | | 95 |
Automated cervical digitized histology whole-slide image analysis toolbox |
|
| Sudhir Sornapudi, Ravitej Addanki, RJoe Stanley, WilliamV Stoecker, Rodney Long, Rosemary Zuna, ShellaineR Frazier, Sameer Antani | | Journal of Pathology Informatics. 2021; 12(1): 26 | | [Pubmed] | [DOI] | | | 96 |
Liver Pathologic Changes After Direct-Acting Antiviral Agent Therapy and Sustained Virologic Response in the Setting of Chronic Hepatitis C Virus Infection |
|
| Romulo Celli, Saad Saffo, Saleem Kamili, Nicholas Wiese, Tonya Hayden, Tamar Taddei, Dhanpat Jain | | Archives of Pathology & Laboratory Medicine. 2021; 145(4): 419 | | [Pubmed] | [DOI] | | | 97 |
Pathologist Concordance for Ovarian Carcinoma Subtype Classification and Identification of Relevant Histologic Features Using Microscope and Whole Slide Imaging |
|
| Marios A. Gavrielides, Brigitte M. Ronnett, Russell Vang, Stephanie Barak, Elsie Lee, Paul N. Staats, Erik Jenson, Priya Skaria, Fahime Sheikhzadeh, Meghan Miller, Ian S. Hagemann, Nicholas Petrick, Jeffrey D. Seidman | | Archives of Pathology & Laboratory Medicine. 2021; 145(12): 1516 | | [Pubmed] | [DOI] | | | 98 |
Feasibility of fully automated classification of whole slide images based on deep learning |
|
| Kyung-Ok Cho, Sung Hak Lee, Hyun-Jong Jang | | The Korean Journal of Physiology & Pharmacology. 2020; 24(1): 89 | | [Pubmed] | [DOI] | | | 99 |
Classification of Molecular Biomarkers |
|
| Ankeet Shah, Dominic C Grimberg, Brant A Inman | | Société Internationale d’Urologie Journal. 2020; 1(1): 8 | | [Pubmed] | [DOI] | | | 100 |
Clinical Decision Support for Ovarian Carcinoma Subtype Classification: A Pilot Observer Study With Pathology Trainees |
|
| Marios A. Gavrielides, Meghan Miller, Ian S. Hagemann, Heba Abdelal, Zahra Alipour, Jie-Fu Chen, Behzad Salari, Lulu Sun, Huifang Zhou, Jeffrey D Seidman | | Archives of Pathology & Laboratory Medicine. 2020; 144(7): 869 | | [Pubmed] | [DOI] | | | 101 |
Counting Mitoses With Digital Pathology in Breast Phyllodes Tumors |
|
| Zi Long Chow, Aye Aye Thike, Hui Hua Li, Nur Diyana Md Nasir, Joe Poh Sheng Yeong, Puay Hoon Tan | | Archives of Pathology & Laboratory Medicine. 2020; 144(11): 1397 | | [Pubmed] | [DOI] | | | 102 |
Objective Diagnosis for Histopathological Images Based on Machine Learning Techniques: Classical Approaches and New Trends |
|
| Naira Elazab, Hassan Soliman, Shaker El-Sappagh, S. M. Riazul Islam, Mohammed Elmogy | | Mathematics. 2020; 8(11): 1863 | | [Pubmed] | [DOI] | | | 103 |
Relevant Applications of Generative Adversarial Networks in Drug Design and Discovery: Molecular De Novo Design, Dimensionality Reduction, and De Novo Peptide and Protein Design |
|
| Eugene Lin, Chieh-Hsin Lin, Hsien-Yuan Lane | | Molecules. 2020; 25(14): 3250 | | [Pubmed] | [DOI] | | | 104 |
Arctic Vision: Using Neural Networks for Ice Object Classification, and Controlling How They Fail |
|
| Ole-Magnus Pedersen, Ekaterina Kim | | Journal of Marine Science and Engineering. 2020; 8(10): 770 | | [Pubmed] | [DOI] | | | 105 |
Time Series Analysis and Forecasting with Automated Machine Learning on a National ICD-10 Database |
|
| Victor Olsavszky, Mihnea Dosius, Cristian Vladescu, Johannes Benecke | | International Journal of Environmental Research and Public Health. 2020; 17(14): 4979 | | [Pubmed] | [DOI] | | | 106 |
Artificial Intelligence Tools for Refining Lung Cancer Screening |
|
| J. Luis Espinoza, Le Thanh Dong | | Journal of Clinical Medicine. 2020; 9(12): 3860 | | [Pubmed] | [DOI] | | | 107 |
The Utility of Deep Learning in Breast Ultrasonic Imaging: A Review |
|
| Tomoyuki Fujioka, Mio Mori, Kazunori Kubota, Jun Oyama, Emi Yamaga, Yuka Yashima, Leona Katsuta, Kyoko Nomura, Miyako Nara, Goshi Oda, Tsuyoshi Nakagawa, Yoshio Kitazume, Ukihide Tateishi | | Diagnostics. 2020; 10(12): 1055 | | [Pubmed] | [DOI] | | | 108 |
Application of Big Data Technology for COVID-19 Prevention and Control in China: Lessons and Recommendations |
|
| Jun Wu, Jian Wang, Stephen Nicholas, Elizabeth Maitland, Qiuyan Fan | | Journal of Medical Internet Research. 2020; 22(10): e21980 | | [Pubmed] | [DOI] | | | 109 |
Automated histologic diagnosis of CNS tumors with machine learning |
|
| Siri Sahib S Khalsa, Todd C Hollon, Arjun Adapa, Esteban Urias, Sudharsan Srinivasan, Neil Jairath, Julianne Szczepanski, Peter Ouillette, Sandra Camelo-Piragua, Daniel A Orringer | | CNS Oncology. 2020; 9(2): CNS56 | | [Pubmed] | [DOI] | | | 110 |
Glioma Grading via Analysis of Digital Pathology Images Using Machine Learning |
|
| Saima Rathore, Tamim Niazi, Muhammad Aksam Iftikhar, Ahmad Chaddad | | Cancers. 2020; 12(3): 578 | | [Pubmed] | [DOI] | | | 111 |
Artificial Intelligence and Digital Microscopy Applications in Diagnostic Hematopathology |
|
| Hanadi El El Achi, Joseph D. Khoury | | Cancers. 2020; 12(4): 797 | | [Pubmed] | [DOI] | | | 112 |
Integrative Data Augmentation with U-Net Segmentation Masks Improves Detection of Lymph Node Metastases in Breast Cancer Patients |
|
| Yong Won Jin, Shuo Jia, Ahmed Bilal Ashraf, Pingzhao Hu | | Cancers. 2020; 12(10): 2934 | | [Pubmed] | [DOI] | | | 113 |
Enhancing Multi-tissue and Multi-scale Cell Nuclei Segmentation with Deep Metric Learning |
|
| Tomas Iesmantas, Agne Paulauskaite-Taraseviciene, Kristina Sutiene | | Applied Sciences. 2020; 10(2): 615 | | [Pubmed] | [DOI] | | | 114 |
An Empirical Evaluation of Nuclei Segmentation from H&E Images in a Real Application Scenario |
|
| Lorenzo Putzu, Giorgio Fumera | | Applied Sciences. 2020; 10(22): 7982 | | [Pubmed] | [DOI] | | | 115 |
Diagnosing Epidermal basal Squamous Cell Carcinoma in High-resolution, and Poorly Labeled Histopathological Imaging |
|
| Mani Manavalan | | Engineering International. 2020; 8(2): 139 | | [Pubmed] | [DOI] | | | 116 |
Computationally Derived Image Signature of Stromal Morphology Is Prognostic of Prostate Cancer Recurrence Following Prostatectomy in African American Patients |
|
| Hersh K. Bhargava, Patrick Leo, Robin Elliott, Andrew Janowczyk, Jon Whitney, Sanjay Gupta, Pingfu Fu, Kosj Yamoah, Francesca Khani, Brian D. Robinson, Timothy R. Rebbeck, Michael Feldman, Priti Lal, Anant Madabhushi | | Clinical Cancer Research. 2020; 26(8): 1915 | | [Pubmed] | [DOI] | | | 117 |
Deep Learning Algorithms for Corneal Amyloid Deposition Quantitation in Familial Amyloidosis |
|
| Klaus Kessel, Jaakko Mattila, Nina Linder, Tero Kivelä, Johan Lundin | | Ocular Oncology and Pathology. 2020; 6(1): 58 | | [Pubmed] | [DOI] | | | 118 |
A deep learning image-based intrinsic molecular subtype classifier of breast tumors reveals tumor heterogeneity that may affect survival |
|
| Mustafa I. Jaber, Bing Song, Clive Taylor, Charles J. Vaske, Stephen C. Benz, Shahrooz Rabizadeh, Patrick Soon-Shiong, Christopher W. Szeto | | Breast Cancer Research. 2020; 22(1) | | [Pubmed] | [DOI] | | | 119 |
Validation of machine learning models to detect amyloid pathologies across institutions |
|
| Juan C. Vizcarra, Marla Gearing, Michael J. Keiser, Jonathan D. Glass, Brittany N. Dugger, David A. Gutman | | Acta Neuropathologica Communications. 2020; 8(1) | | [Pubmed] | [DOI] | | | 120 |
Quantitative Assessment of the Effects of Compression on Deep Learning in Digital Pathology Image Analysis |
|
| Yijiang Chen, Andrew Janowczyk, Anant Madabhushi | | JCO Clinical Cancer Informatics. 2020; (4): 221 | | [Pubmed] | [DOI] | | | 121 |
Deep-Learning–Based Characterization of Tumor-Infiltrating Lymphocytes in Breast Cancers From Histopathology Images and Multiomics Data |
|
| Zixiao Lu, Siwen Xu, Wei Shao, Yi Wu, Jie Zhang, Zhi Han, Qianjin Feng, Kun Huang | | JCO Clinical Cancer Informatics. 2020; (4): 480 | | [Pubmed] | [DOI] | | | 122 |
Reimagining T Staging Through Artificial Intelligence and Machine Learning Image Processing Approaches in Digital Pathology |
|
| Kaustav Bera, Ian Katz, Anant Madabhushi | | JCO Clinical Cancer Informatics. 2020; (4): 1039 | | [Pubmed] | [DOI] | | | 123 |
Hybrid autofluorescence and photoacoustic label-free microscopy for the investigation and identification of malignancies in ocular biopsies |
|
| George J. Tserevelakis, Kostas G. Mavrakis, Danai Pantazopoulou, Eleni Lagoudaki, Efstathios Detorakis, Giannis Zacharakis | | Optics Letters. 2020; 45(20): 5748 | | [Pubmed] | [DOI] | | | 124 |
Bioinformatics analysis of whole slide images reveals significant neighborhood preferences of tumor cells in Hodgkin lymphoma |
|
| Jennifer Hannig, Hendrik Schäfer, Jörg Ackermann, Marie Hebel, Tim Schäfer, Claudia Döring, Sylvia Hartmann, Martin-Leo Hansmann, Ina Koch, Jason A. Papin | | PLOS Computational Biology. 2020; 16(1): e1007516 | | [Pubmed] | [DOI] | | | 125 |
Review of the current state of digital image analysis in breast pathology |
|
| Martin C. Chang, Miralem Mrkonjic | | The Breast Journal. 2020; 26(6): 1208 | | [Pubmed] | [DOI] | | | 126 |
Artificial intelligence as the next step towards precision pathology |
|
| B. Acs, M. Rantalainen, J. Hartman | | Journal of Internal Medicine. 2020; 288(1): 62 | | [Pubmed] | [DOI] | | | 127 |
Impact of image analysis and artificial intelligence in thyroid pathology, with particular reference to cytological aspects |
|
| Ilaria Girolami, Stefano Marletta, Liron Pantanowitz, Evelin Torresani, Claudio Ghimenton, Mattia Barbareschi, Aldo Scarpa, Matteo Brunelli, Valeria Barresi, Pierpaolo Trimboli, Albino Eccher | | Cytopathology. 2020; 31(5): 432 | | [Pubmed] | [DOI] | | | 128 |
Same same but different: A Web-based deep learning application revealed classifying features for the histopathologic distinction of cortical malformations |
|
| Joshua Kubach, Angelika Muhlebner-Fahrngruber, Figen Soylemezoglu, Hajime Miyata, Pitt Niehusmann, Mrinalini Honavar, Fabio Rogerio, Se-Hoon Kim, Eleonora Aronica, Rita Garbelli, Samuel Vilz, Alexander Popp, Stefan Walcher, Christoph Neuner, Michael Scholz, Stefanie Kuerten, Verena Schropp, Sebastian Roeder, Philip Eichhorn, Markus Eckstein, Axel Brehmer, Katja Kobow, Roland Coras, Ingmar Blumcke, Samir Jabari | | Epilepsia. 2020; 61(3): 421 | | [Pubmed] | [DOI] | | | 129 |
A Multi-Organ Nucleus Segmentation Challenge |
|
| Neeraj Kumar, Ruchika Verma, Deepak Anand, Yanning Zhou, Omer Fahri Onder, Efstratios Tsougenis, Hao Chen, Pheng-Ann Heng, Jiahui Li, Zhiqiang Hu, Yunzhi Wang, Navid Alemi Koohbanani, Mostafa Jahanifar, Neda Zamani Tajeddin, Ali Gooya, Nasir Rajpoot, Xuhua Ren, Sihang Zhou, Qian Wang, Dinggang Shen, Cheng-Kun Yang, Chi-Hung Weng, Wei-Hsiang Yu, Chao-Yuan Yeh, Shuang Yang, Shuoyu Xu, Pak Hei Yeung, Peng Sun, Amirreza Mahbod, Gerald Schaefer, Isabella Ellinger, Rupert Ecker, Orjan Smedby, Chunliang Wang, Benjamin Chidester, That-Vinh Ton, Minh-Triet Tran, Jian Ma, Minh N. Do, Simon Graham, Quoc Dang Vu, Jin Tae Kwak, Akshaykumar Gunda, Raviteja Chunduri, Corey Hu, Xiaoyang Zhou, Dariush Lotfi, Reza Safdari, Antanas Kascenas, Alison O'Neil, Dennis Eschweiler, Johannes Stegmaier, Yanping Cui, Baocai Yin, Kailin Chen, Xinmei Tian, Philipp Gruening, Erhardt Barth, Elad Arbel, Itay Remer, Amir Ben-Dor, Ekaterina Sirazitdinova, Matthias Kohl, Stefan Braunewell, Yuexiang Li, Xinpeng Xie, Linlin | | IEEE Transactions on Medical Imaging. 2020; 39(5): 1380 | | [Pubmed] | [DOI] | | | 130 |
Guided Soft Attention Network for Classification of Breast Cancer Histopathology Images |
|
| Heechan Yang, Ji-Ye Kim, Hyongsuk Kim, Shyam P. Adhikari | | IEEE Transactions on Medical Imaging. 2020; 39(5): 1306 | | [Pubmed] | [DOI] | | | 131 |
Multiplex Cellular Communities in Multi-Gigapixel Colorectal Cancer Histology Images for Tissue Phenotyping |
|
| Sajid Javed, Arif Mahmood, Naoufel Werghi, Ksenija Benes, Nasir Rajpoot | | IEEE Transactions on Image Processing. 2020; 29: 9204 | | [Pubmed] | [DOI] | | | 132 |
Improved small blob detection in 3D images using jointly constrained deep learning and Hessian analysis |
|
| Yanzhe Xu, Teresa Wu, Fei Gao, Jennifer R. Charlton, Kevin M. Bennett | | Scientific Reports. 2020; 10(1) | | [Pubmed] | [DOI] | | | 133 |
Tailored for Real-World: A Whole Slide Image Classification System Validated on Uncurated Multi-Site Data Emulating the Prospective Pathology Workload |
|
| Julianna D. Ianni, Rajath E. Soans, Sivaramakrishnan Sankarapandian, Ramachandra Vikas Chamarthi, Devi Ayyagari, Thomas G. Olsen, Michael J. Bonham, Coleman C. Stavish, Kiran Motaparthi, Clay J. Cockerell, Theresa A. Feeser, Jason B. Lee | | Scientific Reports. 2020; 10(1) | | [Pubmed] | [DOI] | | | 134 |
Automated Classification for Visual-Only Postmortem Inspection of Porcine Pathology |
|
| Stephen McKenna, Telmo Amaral, Ilias Kyriazakis | | IEEE Transactions on Automation Science and Engineering. 2020; 17(2): 1005 | | [Pubmed] | [DOI] | | | 135 |
Computer-Aided Diagnosis in Histopathological Images of the Endometrium Using a Convolutional Neural Network and Attention Mechanisms |
|
| Hao Sun, Xianxu Zeng, Tao Xu, Gang Peng, Yutao Ma | | IEEE Journal of Biomedical and Health Informatics. 2020; 24(6): 1664 | | [Pubmed] | [DOI] | | | 136 |
AI in Medical Imaging Informatics: Current Challenges and Future Directions |
|
| Andreas S. Panayides, Amir Amini, Nenad D. Filipovic, Ashish Sharma, Sotirios A. Tsaftaris, Alistair Young, David Foran, Nhan Do, Spyretta Golemati, Tahsin Kurc, Kun Huang, Konstantina S. Nikita, Ben P. Veasey, Michalis Zervakis, Joel H. Saltz, Constantinos S. Pattichis | | IEEE Journal of Biomedical and Health Informatics. 2020; 24(7): 1837 | | [Pubmed] | [DOI] | | | 137 |
Artificial intelligence driven next-generation renal histomorphometry |
|
| Briana A. Santo, Avi Z. Rosenberg, Pinaki Sarder | | Current Opinion in Nephrology and Hypertension. 2020; 29(3): 265 | | [Pubmed] | [DOI] | | | 138 |
MitosisNet: End-to-End Mitotic Cell Detection by Multi-Task Learning |
|
| Md Zahangir Alom, Theus Aspiras, Tarek M. Taha, T.J. Bowen, Vijayan K. Asari | | IEEE Access. 2020; 8: 68695 | | [Pubmed] | [DOI] | | | 139 |
An Accurate and Fast Cardio-Views Classification System Based on Fused Deep Features and LSTM |
|
| Ahmed I. Shahin, Sultan Almotairi | | IEEE Access. 2020; 8: 135184 | | [Pubmed] | [DOI] | | | 140 |
Proteomic investigations into resistance in colorectal cancer |
|
| David I. Cantor, Harish R. Cheruku, Jack Westacott, Joo-Shik Shin, Abidali Mohamedali, Seong Boem Ahn | | Expert Review of Proteomics. 2020; 17(1): 49 | | [Pubmed] | [DOI] | | | 141 |
Supervised machine learning tools: a tutorial for clinicians |
|
| Lucas Lo Vercio, Kimberly Amador, Jordan J Bannister, Sebastian Crites, Alejandro Gutierrez, M. Ethan MacDonald, Jasmine Moore, Pauline Mouches, Deepthi Rajashekar, Serena Schimert, Nagesh Subbanna, Anup Tuladhar, Nanjia Wang, Matthias Wilms, Anthony Winder, Nils D Forkert | | Journal of Neural Engineering. 2020; 17(6): 062001 | | [Pubmed] | [DOI] | | | 142 |
Pan-cancer diagnostic consensus through searching archival histopathology images using artificial intelligence |
|
| Shivam Kalra, H. R. Tizhoosh, Sultaan Shah, Charles Choi, Savvas Damaskinos, Amir Safarpoor, Sobhan Shafiei, Morteza Babaie, Phedias Diamandis, Clinton J. V. Campbell, Liron Pantanowitz | | npj Digital Medicine. 2020; 3(1) | | [Pubmed] | [DOI] | | | 143 |
Accurate diagnosis of lymphoma on whole-slide histopathology images using deep learning |
|
| Charlotte Syrykh, Arnaud Abreu, Nadia Amara, Aurore Siegfried, Véronique Maisongrosse, François X. Frenois, Laurent Martin, Cédric Rossi, Camille Laurent, Pierre Brousset | | npj Digital Medicine. 2020; 3(1) | | [Pubmed] | [DOI] | | | 144 |
Non-disruptive collagen characterization in clinical histopathology using cross-modality image synthesis |
|
| Adib Keikhosravi, Bin Li, Yuming Liu, Matthew W. Conklin, Agnes G. Loeffler, Kevin W. Eliceiri | | Communications Biology. 2020; 3(1) | | [Pubmed] | [DOI] | | | 145 |
Deep learning shows the capability of high-level computer-aided diagnosis in malignant lymphoma |
|
| Hiroaki Miyoshi, Kensaku Sato, Yoshinori Kabeya, Sho Yonezawa, Hiroki Nakano, Yusuke Takeuchi, Issei Ozawa, Shoichi Higo, Eriko Yanagida, Kyohei Yamada, Kei Kohno, Takuya Furuta, Hiroko Muta, Mai Takeuchi, Yuya Sasaki, Takuro Yoshimura, Kotaro Matsuda, Reiji Muto, Mayuko Moritsubo, Kanako Inoue, Takaharu Suzuki, Hiroaki Sekinaga, Koichi Ohshima | | Laboratory Investigation. 2020; 100(10): 1300 | | [Pubmed] | [DOI] | | | 146 |
Automated detection algorithm for C4d immunostaining showed comparable diagnostic performance to pathologists in renal allograft biopsy |
|
| Gyuheon Choi, Young-Gon Kim, Haeyon Cho, Namkug Kim, Hyunna Lee, Kyung Chul Moon, Heounjeong Go | | Modern Pathology. 2020; 33(8): 1626 | | [Pubmed] | [DOI] | | | 147 |
Cellular community detection for tissue phenotyping in colorectal cancer histology images |
|
| Sajid Javed, Arif Mahmood, Muhammad Moazam Fraz, Navid Alemi Koohbanani, Ksenija Benes, Yee-Wah Tsang, Katherine Hewitt, David Epstein, David Snead, Nasir Rajpoot | | Medical Image Analysis. 2020; 63: 101696 | | [Pubmed] | [DOI] | | | 148 |
Clinically applicable histopathological diagnosis system for gastric cancer detection using deep learning |
|
| Zhigang Song, Shuangmei Zou, Weixun Zhou, Yong Huang, Liwei Shao, Jing Yuan, Xiangnan Gou, Wei Jin, Zhanbo Wang, Xin Chen, Xiaohui Ding, Jinhong Liu, Chunkai Yu, Calvin Ku, Cancheng Liu, Zhuo Sun, Gang Xu, Yuefeng Wang, Xiaoqing Zhang, Dandan Wang, Shuhao Wang, Wei Xu, Richard C. Davis, Huaiyin Shi | | Nature Communications. 2020; 11(1) | | [Pubmed] | [DOI] | | | 149 |
Report on computational assessment of Tumor Infiltrating Lymphocytes from the International Immuno-Oncology Biomarker Working Group |
|
| Mohamed Amgad, Elisabeth Specht Stovgaard, Eva Balslev, Jeppe Thagaard, Weijie Chen, Sarah Dudgeon, Ashish Sharma, Jennifer K. Kerner, Carsten Denkert, Yinyin Yuan, Khalid AbdulJabbar, Stephan Wienert, Peter Savas, Leonie Voorwerk, Andrew H. Beck, Anant Madabhushi, Johan Hartman, Manu M. Sebastian, Hugo M. Horlings, Jan Hudecek, Francesco Ciompi, David A. Moore, Rajendra Singh, Elvire Roblin, Marcelo Luiz Balancin, Marie-Christine Mathieu, Jochen K. Lennerz, Pawan Kirtani, I-Chun Chen, Jeremy P. Braybrooke, Giancarlo Pruneri, Sandra Demaria, Sylvia Adams, Stuart J. Schnitt, Sunil R. Lakhani, Federico Rojo, Laura Comerma, Sunil S. Badve, Mehrnoush Khojasteh, W. Fraser Symmans, Christos Sotiriou, Paula Gonzalez-Ericsson, Katherine L. Pogue-Geile, Rim S. Kim, David L. Rimm, Giuseppe Viale, Stephen M. Hewitt, John M. S. Bartlett, Frédérique Penault-Llorca, Shom Goel, Huang-Chun Lien, Sibylle Loibl, Zuzana Kos, Sherene Loi, Matthew G. Hanna, Stefan Michiels, Marleen Kok, Torsten O. Nielsen, | | npj Breast Cancer. 2020; 6(1) | | [Pubmed] | [DOI] | | | 150 |
Dataset of segmented nuclei in hematoxylin and eosin stained histopathology images of ten cancer types |
|
| Le Hou, Rajarsi Gupta, John S. Van Arnam, Yuwei Zhang, Kaustubh Sivalenka, Dimitris Samaras, Tahsin M. Kurc, Joel H. Saltz | | Scientific Data. 2020; 7(1) | | [Pubmed] | [DOI] | | | 151 |
New methodologies in ageing research |
|
| Brenna Osborne, Daniela Bakula, Michael Ben Ezra, Charlotte Dresen, Esben Hartmann, Stella M. Kristensen, Garik V. Mkrtchyan, Malte H. Nielsen, Michael A. Petr, Morten Scheibye-Knudsen | | Ageing Research Reviews. 2020; 62: 101094 | | [Pubmed] | [DOI] | | | 152 |
Policy Implications of Artificial Intelligence and Machine Learning in Diabetes Management |
|
| David T. Broome, C. Beau Hilton, Neil Mehta | | Current Diabetes Reports. 2020; 20(2) | | [Pubmed] | [DOI] | | | 153 |
Novel convolutional neural network architecture for improved pulmonary nodule classification on computed tomography |
|
| Yi Wang, Hao Zhang, Kum Ju Chae, Younhee Choi, Gong Yong Jin, Seok-Bum Ko | | Multidimensional Systems and Signal Processing. 2020; 31(3): 1163 | | [Pubmed] | [DOI] | | | 154 |
Cell Image Classification: A Comparative Overview |
|
| Mohammad Shifat-E-Rabbi, Xuwang Yin, Cailey E. Fitzgerald, Gustavo K. Rohde | | Cytometry Part A. 2020; 97(4): 347 | | [Pubmed] | [DOI] | | | 155 |
Computer-aided diagnosis in the era of deep learning |
|
| Heang-Ping Chan, Lubomir M. Hadjiiski, Ravi K. Samala | | Medical Physics. 2020; 47(5) | | [Pubmed] | [DOI] | | | 156 |
A computer-aided diagnosis system for differentiation and delineation of malignant regions on whole-slide prostate histopathology image using spatial statistics and multidimensional DenseNet |
|
| Chiao-Min Chen, Yao-Sian Huang, Pei-Wei Fang, Cher-Wei Liang, Ruey-Feng Chang | | Medical Physics. 2020; 47(3): 1021 | | [Pubmed] | [DOI] | | | 157 |
Classification of digital pathological images of non-Hodgkin's lymphoma subtypes based on the fusion of transfer learning and principal component analysis |
|
| Jianfei Zhang, Wensheng Cui, Xiaoyan Guo, Bo Wang, Zhen Wang | | Medical Physics. 2020; 47(9): 4241 | | [Pubmed] | [DOI] | | | 158 |
A comparative study of breast cancer tumor classification by classical machine learning methods and deep learning method |
|
| Yadavendra, Satish Chand | | Machine Vision and Applications. 2020; 31(6) | | [Pubmed] | [DOI] | | | 159 |
Culture codes of scientific concepts in global scientific online discourse |
|
| Dina I. Spicheva, Ekaterina V. Polyanskaya | | AI & SOCIETY. 2020; 35(3): 699 | | [Pubmed] | [DOI] | | | 160 |
Deep learning a boon for biophotonics? |
|
| Pranita Pradhan, Shuxia Guo, Oleg Ryabchykov, Juergen Popp, Thomas W. Bocklitz | | Journal of Biophotonics. 2020; 13(6) | | [Pubmed] | [DOI] | | | 161 |
Autofocusing technologies for whole slide imaging and automated microscopy |
|
| Zichao Bian, Chengfei Guo, Shaowei Jiang, Jiakai Zhu, Ruihai Wang, Pengming Song, Zibang Zhang, Kazunori Hoshino, Guoan Zheng | | Journal of Biophotonics. 2020; 13(12) | | [Pubmed] | [DOI] | | | 162 |
Immune contexture analysis in immuno-oncology: applications and challenges of multiplex fluorescent immunohistochemistry |
|
| Reshma Shakya, Tam Hong Nguyen, Nigel Waterhouse, Rajiv Khanna | | Clinical & Translational Immunology. 2020; 9(10) | | [Pubmed] | [DOI] | | | 163 |
Classification of glomerular pathological findings using deep learning and nephrologist–AI collective intelligence approach |
|
| Eiichiro Uchino, Kanata Suzuki, Noriaki Sato, Ryosuke Kojima, Yoshinori Tamada, Shusuke Hiragi, Hideki Yokoi, Nobuhiro Yugami, Sachiko Minamiguchi, Hironori Haga, Motoko Yanagita, Yasushi Okuno | | International Journal of Medical Informatics. 2020; 141: 104231 | | [Pubmed] | [DOI] | | | 164 |
A CNN-based active learning framework to identify mycobacteria in digitized Ziehl-Neelsen stained human tissues |
|
| Mu Yang, Karolina Nurzynska, Ann E. Walts, Arkadiusz Gertych | | Computerized Medical Imaging and Graphics. 2020; 84: 101752 | | [Pubmed] | [DOI] | | | 165 |
A machine learning model for detecting invasive ductal carcinoma with Google Cloud AutoML Vision |
|
| Yan Zeng, Jinmiao Zhang | | Computers in Biology and Medicine. 2020; 122: 103861 | | [Pubmed] | [DOI] | | | 166 |
Computer-aided classification of hepatocellular ballooning in liver biopsies from patients with NASH using persistent homology |
|
| Takashi Teramoto, Toshiya Shinohara, Akihiro Takiyama | | Computer Methods and Programs in Biomedicine. 2020; 195: 105614 | | [Pubmed] | [DOI] | | | 167 |
Gated multimodal networks |
|
| John Arevalo, Thamar Solorio, Manuel Montes-y-Gómez, Fabio A. González | | Neural Computing and Applications. 2020; 32(14): 10209 | | [Pubmed] | [DOI] | | | 168 |
A bird’s-eye view of deep learning in bioimage analysis |
|
| Erik Meijering | | Computational and Structural Biotechnology Journal. 2020; 18: 2312 | | [Pubmed] | [DOI] | | | 169 |
Learning from irregularly sampled data for endomicroscopy super-resolution: a comparative study of sparse and dense approaches |
|
| Agnieszka Barbara Szczotka, Dzhoshkun Ismail Shakir, Daniele Ravě, Matthew J. Clarkson, Stephen P. Pereira, Tom Vercauteren | | International Journal of Computer Assisted Radiology and Surgery. 2020; 15(7): 1167 | | [Pubmed] | [DOI] | | | 170 |
Inconsistent Performance of Deep Learning Models on Mammogram Classification |
|
| Xiaoqin Wang, Gongbo Liang, Yu Zhang, Hunter Blanton, Zachary Bessinger, Nathan Jacobs | | Journal of the American College of Radiology. 2020; 17(6): 796 | | [Pubmed] | [DOI] | | | 171 |
Deep-Hipo: Multi-scale receptive field deep learning for histopathological image analysis |
|
| Sai Chandra Kosaraju, Jie Hao, Hyun Min Koh, Mingon Kang | | Methods. 2020; 179: 3 | | [Pubmed] | [DOI] | | | 172 |
Deep-learning approaches for Gleason grading of prostate biopsies |
|
| Anant Madabhushi, Michael D Feldman, Patrick Leo | | The Lancet Oncology. 2020; 21(2): 187 | | [Pubmed] | [DOI] | | | 173 |
Clinical deployment of AI for prostate cancer diagnosis |
|
| Andrew Janowczyk, Patrick Leo, Mark A Rubin | | The Lancet Digital Health. 2020; 2(8): e383 | | [Pubmed] | [DOI] | | | 174 |
Deep feature transfer learning for trusted and automated malware signature generation in private cloud environments |
|
| Daniel Nahmias, Aviad Cohen, Nir Nissim, Yuval Elovici | | Neural Networks. 2020; 124: 243 | | [Pubmed] | [DOI] | | | 175 |
An experimental study on classification of thyroid histopathology images using transfer learning |
|
| Vijaya Gajanan Buddhavarapu, Angel Arul Jothi J | | Pattern Recognition Letters. 2020; 140: 1 | | [Pubmed] | [DOI] | | | 176 |
Robust nuclei segmentation in histopathology using ASPPU-Net and boundary refinement |
|
| Tao Wan, Lei Zhao, Hongxiang Feng, Deyu Li, Chao Tong, Zengchang Qin | | Neurocomputing. 2020; 408: 144 | | [Pubmed] | [DOI] | | | 177 |
Trace, Machine Learning of Signal Images for Trace-Sensitive Mass Spectrometry: A Case Study from Single-Cell Metabolomics |
|
| Zhichao Liu, Erika P. Portero, Yiren Jian, Yunjie Zhao, Rosemary M. Onjiko, Chen Zeng, Peter Nemes | | Analytical Chemistry. 2019; 91(9): 5768 | | [Pubmed] | [DOI] | | | 178 |
Automating the Paris System for urine cytopathology—A hybrid deep-learning and morphometric approach |
|
| Louis J. Vaickus, Arief A. Suriawinata, Jason W. Wei, Xiaoying Liu | | Cancer Cytopathology. 2019; 127(2): 98 | | [Pubmed] | [DOI] | | | 179 |
Performance of an artificial intelligence algorithm for reporting urine cytopathology |
|
| Adit B. Sanghvi, Erastus Z. Allen, Keith M. Callenberg, Liron Pantanowitz | | Cancer Cytopathology. 2019; 127(10): 658 | | [Pubmed] | [DOI] | | | 180 |
Classification of Benign and Malignant Breast Cancer using Supervised Machine Learning Algorithms Based on Image and Numeric Datasets |
|
| Ratula Ray, Azian Azamimi Abdullah, Debasish Kumar Mallick, Satya Ranjan Dash | | Journal of Physics: Conference Series. 2019; 1372(1): 012062 | | [Pubmed] | [DOI] | | | 181 |
Spatial Architecture and Arrangement of Tumor-Infiltrating Lymphocytes for Predicting Likelihood of Recurrence in Early-Stage Non–Small Cell Lung Cancer |
|
| Germán Corredor, Xiangxue Wang, Yu Zhou, Cheng Lu, Pingfu Fu, Konstantinos Syrigos, David L. Rimm, Michael Yang, Eduardo Romero, Kurt A. Schalper, Vamsidhar Velcheti, Anant Madabhushi | | Clinical Cancer Research. 2019; 25(5): 1526 | | [Pubmed] | [DOI] | | | 182 |
The Impact of Artificial Intelligence on the Labor Market |
|
| Michael Webb | | SSRN Electronic Journal. 2019; | | [Pubmed] | [DOI] | | | 183 |
Deep learning techniques for detecting preneoplastic and neoplastic lesions in human colorectal histological images |
|
| Paola Sena, Rita Fioresi, Francesco Faglioni, Lorena Losi, Giovanni Faglioni, Luca Roncucci | | Oncology Letters. 2019; | | [Pubmed] | [DOI] | | | 184 |
Systems biology primer: the basic methods and approaches |
|
| Iman Tavassoly, Joseph Goldfarb, Ravi Iyengar | | Essays in Biochemistry. 2018; 62(4): 487 | | [Pubmed] | [DOI] | | | 185 |
Application of deep learning to the classification of images from colposcopy |
|
| Masakazu Sato, Koji Horie, Aki Hara, Yuichiro Miyamoto, Kazuko Kurihara, Kensuke Tomio, Harushige Yokota | | Oncology Letters. 2018; | | [Pubmed] | [DOI] | | | 186 |
Automatic labeling of molecular biomarkers of immunohistochemistry images using fully convolutional networks |
|
| Fahime Sheikhzadeh,Rabab K. Ward,Dirk van Niekerk,Martial Guillaud,Christophe Egles | | PLOS ONE. 2018; 13(1): e0190783 | | [Pubmed] | [DOI] | | | 187 |
Glomerulus Classification and Detection Based on Convolutional Neural Networks |
|
| Jaime Gallego,Anibal Pedraza,Samuel Lopez,Georg Steiner,Lucia Gonzalez,Arvydas Laurinavicius,Gloria Bueno | | Journal of Imaging. 2018; 4(1): 20 | | [Pubmed] | [DOI] | | | 188 |
Deep learning based tissue analysis predicts outcome in colorectal cancer |
|
| Dmitrii Bychkov,Nina Linder,Riku Turkki,Stig Nordling,Panu E. Kovanen,Clare Verrill,Margarita Walliander,Mikael Lundin,Caj Haglund,Johan Lundin | | Scientific Reports. 2018; 8(1) | | [Pubmed] | [DOI] | | | 189 |
Cell detection in pathology and microscopy images with multi-scale fully convolutional neural networks |
|
| Xipeng Pan,Dengxian Yang,Lingqiao Li,Zhenbing Liu,Huihua Yang,Zhiwei Cao,Yubei He,Zhen Ma,Yiyi Chen | | World Wide Web. 2018; | | [Pubmed] | [DOI] | | | 190 |
COUNTERPOINT: Is International Statistical Classification of Diseases and Related Health Problems, 10th Revision Diagnosis Coding Important in the Era of Big Data? No |
|
| David M. Liebovitz,John Fahrenbach | | Chest. 2018; | | [Pubmed] | [DOI] | | | 191 |
Association of Pathological Fibrosis With Renal Survival Using Deep Neural Networks |
|
| Vijaya B. Kolachalama,Priyamvada Singh,Christopher Q. Lin,Dan Mun,Mostafa E. Belghasem,Joel M. Henderson,Jean M. Francis,David J. Salant,Vipul C. Chitalia | | Kidney International Reports. 2018; | | [Pubmed] | [DOI] | | | 192 |
Machine Learning Methods for Histopathological Image Analysis |
|
| Daisuke Komura,Shumpei Ishikawa | | Computational and Structural Biotechnology Journal. 2018; 16: 34 | | [Pubmed] | [DOI] | | | 193 |
Enabling Precision Cardiology Through Multiscale Biology and Systems Medicine |
|
| Kipp W. Johnson,Khader Shameer,Benjamin S. Glicksberg,Ben Readhead,Partho P. Sengupta,Johan L.M. Björkegren,Jason C. Kovacic,Joel T. Dudley | | JACC: Basic to Translational Science. 2017; 2(3): 311 | | [Pubmed] | [DOI] | | | 194 |
A survey on deep learning in medical image analysis |
|
| Geert Litjens,Thijs Kooi,Babak Ehteshami Bejnordi,Arnaud Arindra Adiyoso Setio,Francesco Ciompi,Mohsen Ghafoorian,Jeroen A.W.M. van der Laak,Bram van Ginneken,Clara I. Sánchez | | Medical Image Analysis. 2017; 42: 60 | | [Pubmed] | [DOI] | | | 195 |
Prediction of multi-drug resistant TB from CT pulmonary Images based on deep learning techniques |
|
| Xiaohong Gao,Yu Qian | | Molecular Pharmaceutics. 2017; | | [Pubmed] | [DOI] | | | 196 |
A deep learning based strategy for identifying and associating mitotic activity with gene expression derived risk categories in estrogen receptor positive breast cancers |
|
| David Romo-Bucheli,Andrew Janowczyk,Hannah Gilmore,Eduardo Romero,Anant Madabhushi | | Cytometry Part A. 2017; | | [Pubmed] | [DOI] | | | 197 |
Automatic cellularity assessment from post-treated breast surgical specimens |
|
| Mohammad Peikari,Sherine Salama,Sharon Nofech-Mozes,Anne L. Martel | | Cytometry Part A. 2017; | | [Pubmed] | [DOI] | | | 198 |
A deep learning method for classifying mammographic breast density categories |
|
| Aly A. Mohamed,Wendie A. Berg,Hong Peng,Yahong Luo,Rachel C. Jankowitz,Shandong Wu | | Medical Physics. 2017; | | [Pubmed] | [DOI] | | | 199 |
Bringing 3D tumor models to the clinic - predictive value for personalized medicine |
|
| Kathrin Halfter,Barbara Mayer | | Biotechnology Journal. 2017; : 1600295 | | [Pubmed] | [DOI] | | | 200 |
Development of CD3 cell quantitation algorithms for renal allograft biopsy rejection assessment utilizing open source image analysis software |
|
| Andres Moon,Geoffrey H. Smith,Jun Kong,Thomas E. Rogers,Carla L. Ellis,Alton B. “Brad” Farris | | Virchows Archiv. 2017; | | [Pubmed] | [DOI] | | | 201 |
Precision histology: how deep learning is poised to revitalize histomorphology for personalized cancer care |
|
| Ugljesa Djuric,Gelareh Zadeh,Kenneth Aldape,Phedias Diamandis | | npj Precision Oncology. 2017; 1(1) | | [Pubmed] | [DOI] | | | 202 |
A novel machine learning approach reveals latent vascular phenotypes predictive of renal cancer outcome |
|
| Nathan Ing,Fangjin Huang,Andrew Conley,Sungyong You,Zhaoxuan Ma,Sergey Klimov,Chisato Ohe,Xiaopu Yuan,Mahul B. Amin,Robert Figlin,Arkadiusz Gertych,Beatrice S. Knudsen | | Scientific Reports. 2017; 7(1) | | [Pubmed] | [DOI] | | | 203 |
Prediction of recurrence in early stage non-small cell lung cancer using computer extracted nuclear features from digital H&E images |
|
| Xiangxue Wang,Andrew Janowczyk,Yu Zhou,Rajat Thawani,Pingfu Fu,Kurt Schalper,Vamsidhar Velcheti,Anant Madabhushi | | Scientific Reports. 2017; 7(1) | | [Pubmed] | [DOI] | | | 204 |
Glandular Morphometrics for Objective Grading of Colorectal Adenocarcinoma Histology Images |
|
| Ruqayya Awan,Korsuk Sirinukunwattana,David Epstein,Samuel Jefferyes,Uvais Qidwai,Zia Aftab,Imaad Mujeeb,David Snead,Nasir Rajpoot | | Scientific Reports. 2017; 7(1) | | [Pubmed] | [DOI] | | | 205 |
Relevance of deep learning to facilitate the diagnosis of HER2 status in breast cancer |
|
| Michel E. Vandenberghe,Marietta L. J. Scott,Paul W. Scorer,Magnus Söderberg,Denis Balcerzak,Craig Barker | | Scientific Reports. 2017; 7: 45938 | | [Pubmed] | [DOI] | | | 206 |
Automatic Nuclear Segmentation Using Multiscale Radial Line Scanning With Dynamic Programming |
|
| Hongming Xu,Cheng Lu,Richard Berendt,Naresh Jha,Mrinal Mandal | | IEEE Transactions on Biomedical Engineering. 2017; 64(10): 2475 | | [Pubmed] | [DOI] | | | 207 |
A Dataset and a Technique for Generalized Nuclear Segmentation for Computational Pathology |
|
| Neeraj Kumar,Ruchika Verma,Sanuj Sharma,Surabhi Bhargava,Abhishek Vahadane,Amit Sethi | | IEEE Transactions on Medical Imaging. 2017; 36(7): 1550 | | [Pubmed] | [DOI] | | | 208 |
Retrieval From and Understanding of Large-Scale Multi-modal Medical Datasets: A Review |
|
| Henning Muller,Devrim Unay | | IEEE Transactions on Multimedia. 2017; 19(9): 2093 | | [Pubmed] | [DOI] | | | 209 |
Very Deep Convolutional Neural Networks for Morphologic Classification of Erythrocytes |
|
| Thomas J S Durant, Eben M Olson, Wade L Schulz, Richard Torres | | Clinical Chemistry. 2017; 63(12): 1847 | | [Pubmed] | [DOI] | | | 210 |
Deep Learning Makes Its Way to the Clinical Laboratory |
|
| Ronald Jackups | | Clinical Chemistry. 2017; 63(12): 1790 | | [Pubmed] | [DOI] | |
|
| |
 |
|
|
|
