Colored Kimia Path24 Dataset: Configurations and Benchmarks with Deep Embeddings
CColored Kimia Path24
Dataset:Configurations and Benchmarks with Deep Embeddings
Sobhan Shafiei, Morteza Babaie, Shivam Kalra, H.R.TizhooshKimia Lab, University of Waterloo, Canada https://kimia.uwaterloo.ca/
Abstract
The Kimia Path24 dataset has been introduced as a clas-sification and retrieval dataset for digital pathology. Al-though it provides multi-class data, the color informationhas been neglected in the process of extracting patches.The staining information plays a major role in the recogni-tion of tissue patterns. To address this drawback, we intro-duce the color version of Kimia Path24 by recreating sam-ple patches from all 24 scans to propose
Kimia Path24C .We run extensive experiments to determine the best con-figuration for selected patches. To provide preliminary re-sults for setting a benchmark for the new dataset, we uti-lize VGG16, InceptionV3 and DenseNet-121 model as fea-ture extractors. Then, we use these feature vectors to re-trieve test patches. The accuracy of image retrieval usingDenseNet was 95.92% while the highest accuracy using In-ceptionV3 and VGG16 reached 92.45% and 92%, respec-tively. We also experimented with “deep barcodes” and es-tablished that with a small loss in accuracy (e.g., 93.43%for binarized features for DenseNet instead of 95.92% whenthe features themselves are used), the search operations canbe significantly accelerated.
1. Introduction
Diagnosis of biopsy samples is performed by a patholo-gist examining the stained specimen on the glass slide us-ing a microscope. In recent years, attempts have been madeto capture the entire slide with a scanner and save it as adigital image or a whole slide image (WSI) [24]. Digi-tal pathology opens new horizons for biomedical researchand clinical routine but also creates several challenges forcomputer vision research [21, 37]. Laboratories can scan alarge number of slides per day, each being a gigapixel imagecontaining different types of tissues with different stainingtechniques. Quantity, variability, and size of WSIs requireefficient and versatile computer vision methods. As well, la-beled data seems to be a rarity in digital pathology; manual delineation of regions of interest is not part of the clinicalworkflow, a fact that hinders the usage of many supervisedlearning algorithms.The most actively researched task in digital pathologyimage analysis is the computer-assisted diagnosis (CAD),where the machine is trying to imitate the pathologist’s task.The diagnostic process maps a WSI or multiple WSIs to oneof the disease categories. This is essentially a supervisedlearning task. Since the errors made by a computer systemreportedly differ from those made by a human pathologist[40], classification accuracy could be improved using theCAD system. The CAD may also lead to the reduced vari-ability in interpretations and prevent overlooking regions ofinterest by investigating all pixels within WSIs.Another research tasks in the digital pathology isContent-Based Image Retrieval (CBIR). A CBIR system re-trieves similar images when a query image is provided. Indigital pathology, CBIR systems are useful in many situ-ations, particularly in diagnostics, education, and research[3, 33]. For instance, CBIR systems can be used for ed-ucational purposes by students and pathologist residents toretrieve relevant cases or histopathology images. CBIR sys-tems can help pathologists to diagnose rare cases if largearchives are available.Probably the biggest problem in pathology image analy-sis using machine learning is that only a limited number oflabeled data is available. In the field of digital pathology,there are some public datasets that contain hand-labelledhistopathology images [7, 11, 17, 43, 6, 29, 12, 5]. Theycould be useful if the purpose of the analysis, staining pro-cedure, and magnification levels and resolutions are simi-lar. However, because each of these datasets focuses on aspecific disease, cell type, or staining, they are not genericenough to cover most necessary CAD tasks. There arealso several large-scale histopathology image databases thatcontain high-resolution WSIs. The Cancer Genome At-las (TCGA) [41] contains over 30,000 WSIs from variouscancer types, and the Genotype-Tissue Expression dataset(GTEx) [20] contains over 20,000 WSIs from various tis-sues. These datasets may serve as potential training data1 a r X i v : . [ ee ss . I V ] F e b or various tasks. Furthermore, both TCGA and GTEx alsoprovide genomic profiles, which could be used to investi-gate the relationships between genotype and morphology.One of pathology dataset is the Kimia Path24 dataset.This dataset includes 24 WSIs from different tissue tex-tures and different staining procedures [1]. This dataset isstructured to mimic retrieval tasks in clinical practice. Thepatch-to-scan classification may appear trivial but has rele-vant clinical applications such as “ floater detection ” wherewe need to find the origin of a foreign tissue [26, 18, 25].To generate the training and test datasets, Babaie et al.[1] slide a window with no overlapping over each WSIto crop patches of size 1000 × Kimia Path24C ”. To provide preliminary re-sults and set a benchmark for the proposed dataset, we usedthree pre-trained deep networks for feature extraction. Sub-sequently, we use these feature vectors for image retrieval.We show that adding color to the dataset combined with us-ing deep features provide high retrieval accuracy. This dataset is publicly available on the website of Kimia Lab .
2. Related Works
A convolutional neural network (CNN) is a type of deep,feed-forward artificial neural network which receives rawinput images and extracts complex features from them tooutput class assignments [30]. The CNNs use the convo-lution of spatial information in the images to share weightsacross units.Most of the image recognition techniques have beenreplaced by deep learning algorithms, after the outstand-ing results of deep learning algorithms in the analysis oflarge-scale databases [15]. More recently, researchers havebeen working on applying deep learning to pathology im-age analysis. For example, Spanhol et al. [32] present aset of experiments using a deep learning approach to avoidhand-crafted features. They have shown that an existingCNN architecture designed for classifying color images ofobjects can be used for the classification of histopathol-ogy images. Xu et al. [42] proposed a deep convolu-tional neural network based on feature learning to automat-ically segment or classify epithelium and stroma regions https://kimia.uwaterloo.ca/ from digitized tumor tissue microarrays. Automated breastcancer multi-classification using histopathological imagesis of great clinical significance. Han et al. [8] proposeda breast cancer multi-classification method using a deeplearning model that achieves remarkable performance (av-erage 93.2% accuracy) on a large dataset. Mormont et al.[23] study transfer learning as a way of overcoming ob-ject recognition challenges encountered in the field of dig-ital pathology. Their experiments on eight classificationdatasets show that densely connected and residual networksconsistently yield best performances across all topologies.Khosravi et al. [13] introduced a classification pipelineincluding a basic CNN architecture, Google’s Inceptionswith three training strategies, and an ensemble of Inceptionand ResNet to effectively classify different histopathologyimages for different types of cancer. They demonstrated theaccuracy of deep learning approaches for identifying cancersubtypes, and the robustness of Google’s Inceptions evenin presence of extensive tumor heterogeneity. On average,their pipeline achieves accuracy values of 100%, 92%, 95%,and 69% for discrimination of various cancer tissues, sub-types, biomarkers, and scores, respectively. For more in-formation about deep neural network approaches and theirapplication in medical image analysis see Litjens et al. [19]. Segmentation is a common task in both natural and med-ical image analysis. K-means clustering and Gaussian Mix-ture Model (GMM) appear to be popular choices for theunsupervised segmentation tasks. The K-means clusteringis an iterative distance-based clustering algorithm that as-signs n observations to exactly one of the k clusters de-fined by centroids, where k is chosen before the algorithmstarts [9]. The Gaussian mixture model is a well-knownmethod used in most applications such as data mining, pat-tern recognition, machine learning, and statistical analysis.GMM assumes the data to consist of several Gaussian dis-tributions, which can be discovered through the parameterlearning process. The major advantage of the GMM ap-proach is that its mathematical equation is easy to evaluate.As well, it requires a small number of parameters for learn-ing. For more information on GMM see McLachlan andPeel [22].Image segmentation is one of the main components in afully automated cell image analysis. Segmentation mainlyfocuses on the separation of the cells from the backgroundas well as separation of the nucleus from the cytoplasmwithin the cell regions for extracting cellular features fromimages. For example, Bamford and Lovell [2] segmentedthe nuclei in a pap smear image using an active con-tour model that was estimated by using dynamic program-ming to find the boundary with the minimum cost within abounded space around the darkest point in the image. Yang-2ao et al. [44] applies automatic thresholding to the imagegradient to identify the edge pixels corresponding to nucleiand cytoplasm boundaries in cervical cell images. Tsai etal. [39] replace the thresholding step with K-means clus-tering into two partitions. Ragothaman et al. [27] used theGMM to segment cell regions to identify cellular featuressuch as the nucleus and cytoplasm. Segmentation is a cru-cial task for WSI representation. In this paper, we utilizeboth K-means and GMM to localize the tissue region andextract patches from the informative part of the WSI andavoid background pixels.
3. A new dataset: Kimia Path24C
In this paper, we assemble, test and propose a new ver-sion of the Kimia Path24 dataset introduced by Babaie etal. [1]. It contains 24 WSIs with several staining tech-niques namely immunohistochemical (IHC), Hematoxylinand Eosin (H&E) and Masson’s trichrome staining, selectedfrom 350 WSIs depicting diverse body parts so that the im-ages clearly represent different texture pattern [1]. Babaieet al. [1] have visually selected 1,325 patches of size1000 × × magnification as test dataset that rep-resent all dominant tissue textures in each WSI. Indeed, testdataset contains multiple texture patterns.Considering Babaie et al. [1] saved the extracted patchesas grayscale images, the staining information that is veryimportant in the analysis of histopathological images waslost. We create a new version of the dataset that not onlycontains the staining information but also contains less ir-relevant patches (i.e., patches with background pixels anddebris). Besides, we would like to fix the training patchessuch that benchmarking becomes more consistent. In the new version of the Kimia Path24 dataset, the lo-cations and number of patches of the test dataset have notchanged and only the extracted patches have been savedas RGB images instead of grayscale. For patch extractionand creating an informative training dataset, we first iden-tify tissue within each WSI and exclude background pixels(i.e., largely white/bright space). Background detection isneeded to reduce the computation time and find informa-tive regions that contain tissue patterns suitable for analy-sis. To achieve this, we use two popular segmentation al-gorithms, namely the K-means and GMM, to automaticallydetect the background pixels. In particular, all thumbnailswere first extracted in × magnification and artifacts suchas air bubbles under the cover slip or dust were manuallyeliminated. Then each thumbnail is segmented into fiveclusters, namely the background, fat, cell nuclei, connectivetissues, and blood. After segmentation, a window is movedover each label matrix to crop patches of size 50 ×
50 pix-els in × magnification. The label of background pixels assigned by the segmentation algorithm is found using thepatch in the top right corner of the label matrix. If back-ground share of a patch is less than a predefined value, weconsider that patch as a suitable patch (containing enoughtissue sample) otherwise we ignore it. Finally, using the co-ordinates of selected patches, we extract the correspondingpatches of size 1000 × × magnification level.Note that to construct a training dataset, we first remove(whiten) locations of test patches in each WSI so that thesepatches cannot be mistakenly used for training. Extracting suitable features is a critical step for success-fully implementing an image search or classification algo-rithm. The literature in recent years suggests that the pre-trained deep networks can be used as fine feature extractors[14]. A significant benefit of the pre-trained deep networksis that we do not need to train a deep network again and ex-tracted features can be directly applied to the existing imageanalysis pipelines [19].In this study, we used three pre-trained deep networks,namely
VGG16 [31],
InceptionV3 [35], and
DenseNet-121 [10], as feature extractors without any fine-tuning. Thesenetworks have been trained on more than a million imagesfrom the
ImageNet database [28] and can classify imagesinto 1000 object categories, such as pencil, keyboard, andmany objects and animals. Simonyan et al. [31] proposed
VGG16 architecture with layers being deeper than theother existing architectures at that time. They also utilized × kernels with a stride of pixel in order to decrease thenumber of parameters. InceptionV3 is a 42-layer networkintroduced by Szegedy et al. [35]. They believed that theuse of multiple convolutional layers with a small filter sizeinstead of a single convolutional layer with a large filter sizecan reduce the number of parameters and the computationalcost, and can still achieve a similar level of model expres-siveness. By the introduction of the
DenseNet by Huanget al. [10], it was shown that deep networks can achievethe highest accuracy values while they stay efficient if theyhave shorter connections between layers either close to theinput or the output of the network. Unlike the original neu-ral network configurations where an L -layer network has L connections, in a DenseNet the feature maps of all preced-ing layers are fed into each layer. This solution not only di-minishes the vanishing gradient problem but also enhancesfeature propagation and feature reuse, while it decreases thenumber of parameters.We run each experiment using the architecture for the
VGG16 , Inceptionv3 and
DenseNet networks as provided inthe Keras Python package [4]. The final average poolinglayer right before the last fully connected layer of these net-works prior to classification was extracted as a feature vec-tor. The length of the feature vector from
VGG16 , Incep- ionv3 and DenseNet is 512, 2048 and 1024, respectively.
4. Results and discussion
To determine the optimal size of the training dataset nec-essary to achieve high accuracy, we consider seven differentcandidate values for the background. In addition, for eachof the 24 WSIs, patches are extracted with different overlapratios, without any overlap, and with 20% overlap. Over-all, we extracted 28 different training datasets using thesestrategies. The number of extracted patches for the trainingdataset via different strategies used for performance evalu-ation is reported in Table 1.Table 1: The number of extracted patches via differentstrategies.
K-means GMM
Back. %Overlap % 0 20 0 2010 11429 18795 13425 2314420 13293 21900 15636 2615530 14804 24412 17750 2912540 16313 26987 19479 3192450 17888 29558 20752 3420960 19437 32136 21811 3594470 20927 34549 22591 37444
From Table 1, we conclude that the size of the trainingdataset extracted based on the GMM is larger than K-meansalgorithm. The segmentation outputs of four WSIs obtainedby two techniques and extracted patches with at most 50%background are shown in Fig. 1. As Fig. 1 shows, clus-ter assignment is much more flexible in GMM than in K-means. The latter loses many patches containing fat tissueby mistaking them with background pixels.For the image search, one can compare the correspond-ing feature vectors of two images using a distance metricto measure the similarity. For three selected networks, wehave used similarity metrics to find the most similar im-age in the training dataset for each query image in the testdataset based on the minimum distance between the deepfeatures of the query and training images. In this study, Wehave utilized city block distance to measure the similaritybetween two feature vectors.To evaluate the different strategies, we utilize differentaccuracy measures [1, 14]. The test dataset consists of n =1325 patches that belong to 24 WSIs. Suppose that Γ s i = { P js i ; j = 1 , . . . , n i } is the set of all patches for slide i inthe test dataset in which n i is the number of test patches forslide s i and i = 0 , , . . . , . The patch-to-scan and whole- scan accuracy are defined as η p = (cid:80) i =0 | T s i ∩ Γ s i | n , and η w = 124 (cid:88) i =0 | T s i ∩ Γ s i | n i , respectively, where set T s i contains the correspondingpatches of slide s i in the training dataset and | A | denotesthe cardinality of set A . In addition, the total accuracy isgiven by η tot = η p × η w , which can be used to take into ac-count both patch-to-scan and whole-scan accuracy values.For comparison of all strategies and selecting the besttraining dataset, the accuracy for each strategy are summa-rized in Table 2. As shown in Table 2, the DenseNet yieldsthe best results. Accuracy values obtained with the
VGG16 network are rather low. In addition, the accuracy of ex-tracted training dataset without any overlap is higher thanthe training dataset with 20% overlap. Also, the GMM strat-egy is superior to K-means segmentation. From the resultsshown in Table 2, it is apparent that the highest accuracyvalues belong to training data with 22,591 patches whichwere extracted using GMM without any overlap and at most70% background. We considered this set as the training setfor the
Kimia Path24C dataset and make it publicly avail-able.Concerning the file size in archives of histopathology im-ages, low storage demand and high search speed are crucialfor feasible CBIR systems. We used the concept of deepbarcodes [36, 16] to compare the performance of deep fea-tures with deep barcodes. As Table 3 shows deep barcodesdo suffer from a slight drop in accuracy value. However,the practical benefits of using binary information may over-weigh the small accuracy loss.Fig. 2 shows some retrieval examples of our image re-trieval experiments based on the selected training set andthe
DenseNet model as a feature extractor. The top threerelevant images are listed for each query image. The out-standing patch retrieval accuracy has applications in digitalpathology such as floater detection when we are interestedto find the same tissue pattern originating from the samepatient.
In this paper, we have presented a study on determin-ing a suitable training dataset of RGB patches necessary toachieve high accuracy for the Kimia Path24 dataset. Foreach of the 24 slides, patches were extracted with sev-eral different values for the foreground/background andthe patch overlap ratios. Next, we selected a total of22,591 RGB patches as the training dataset based on theaccuracy of image retrieval. Although the number of ex-tracted patches is less than the grayscale training datasetproposed by [1] but the accuracy of image search has beenincreased. The highest accuracy of image retrieval using4igure 1: Four examples of our segmentation results and patch extraction: the first row is the thumbnail of the slide, theoutcome of the − class K-means marked in red and the outcome of the − class GMM marked in black. The locations of testpatches are marked with black squares.proposed training data is 95.92% while the highest accu-racy of grayscale training dataset based on the Local Bi-nary Patterns [1] and Fine-tuning InceptinV3 [14] reached66.11% and 74.87%, respectively. The new dataset, called
Kimia Path24C , is publicly available at http://kimia.uwaterloo.ca/ . References [1] M. Babaie, S. Kalra, A. Sriram, C. Mitcheltree, S. Zhu,A. Khatami, S. Rahnamayan, and H. R. Tizhoosh. Classifi- cation and retrieval of digital pathology scans: a new dataset.In
Cvmi workshop@ cvpr , 2017.[2] P. Bamford and B. Lovell. Unsupervised cell nucleussegmentation with active contours.
Signal processing ,71(2):203–213, 1998.[3] J. C. Caicedo, F. A. Gonz´alez, and E. Romero. Content-based histopathology image retrieval using a kernel-basedsemantic annotation framework.
Journal of biomedical in-formatics , 44(4):519–528, 2011.[4] F. Chollet et al. Keras. https://keras.io , 2015.[5] K. Faust, Q. Xie, D. Han, K. Goyle, Z. Volynskaya,U. Djuric, and P. Diamandis. Visualizing histopathologic
Measure Network Without Overlap With OverlapBackground Percentage Background Percentage
10 20 30 40 50 60 70 10 20 30 40 50 60 70 η p Dense 76.53 † ‡ η w Dense 74.67 † ‡ η tot Dense 57.14 † ‡ † and ‡ stand for the K-means and GMM, respectively. Table 3: Comparison of deep features and deep barcodes.
Measure Network
VGG16 InceptionV3 DenseNet-121 η p † η p ‡ η w η w η tot η tot † and ‡ stand for deep features and barcodes, respectively. deep learning classification and anomaly detection usingnonlinear feature space dimensionality reduction. BMCbioinformatics , 19(1):1–15, 2018.[6] J. Gamper, N. A. Koohbanani, K. Benet, A. Khuram, andN. Rajpoot. Pannuke: an open pan-cancer histology datasetfor nuclei instance segmentation and classification. In
Euro-pean Congress on Digital Pathology , pages 11–19. Springer,2019.[7] E. D. Gelasca, J. Byun, B. Obara, and B. Manjunath. Evalu-ation and benchmark for biological image segmentation. In
Image Processing, 2008. ICIP 2008. 15th IEEE InternationalConference on , pages 1816–1819. IEEE, 2008.[8] Z. Han, B. Wei, Y. Zheng, Y. Yin, K. Li, and S. Li.Breast cancer multi-classification from histopathological im-ages with structured deep learning model.
Scientific reports ,7(1):4172, 2017.[9] J. A. Hartigan and M. A. Wong. Algorithm as 136: A k- means clustering algorithm.
Journal of the Royal StatisticalSociety. Series C (Applied Statistics) , 28(1):100–108, 1979.[10] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Wein-berger. Densely connected convolutional networks. In
Pro-ceedings of the IEEE conference on computer vision and pat-tern recognition , pages 4700–4708, 2017.[11] A. Janowczyk and A. Madabhushi. Deep learning for digitalpathology image analysis: A comprehensive tutorial with se-lected use cases.
Journal of pathology informatics , 7, 2016.[12] S. Javed, A. Mahmood, M. M. Fraz, N. A. Koohbanani,K. Benes, Y.-W. Tsang, K. Hewitt, D. Epstein, D. Snead, andN. Rajpoot. Cellular community detection for tissue pheno-typing in colorectal cancer histology images.
Medical imageanalysis , 63:101696, 2020.[13] P. Khosravi, E. Kazemi, M. Imielinski, O. Elemento, andI. Hajirasouliha. Deep convolutional neural networks en-able discrimination of heterogeneous digital pathology im-ages.
EBioMedicine , 27:317–328, 2018.[14] B. Kieffer, M. Babaie, S. Kalra, and H. R. Tizhoosh. Con-volutional neural networks for histopathology image classi-fication: Training vs. using pre-trained networks. In , pages 1–6. IEEE, 2017.[15] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenetclassification with deep convolutional neural networks. In
Advances in neural information processing systems , pages1097–1105, 2012.[16] M. D. Kumar, M. Babaie, and H. R. Tizhoosh. Deep bar-codes for fast retrieval of histopathology scans. In ,pages 1–8. IEEE, 2018.[17] M. D. Kumar, M. Babaie, S. Zhu, S. Kalra, and H. R.Tizhoosh. A comparative study of cnn, bovw and lbp for lassification of histopathological images. In ,pages 1–7. IEEE, 2017.[18] L. J. Layfield, B. L. Witt, K. G. Metzger, and G. M. Ander-son. Extraneous tissue: a potential source for diagnostic errorin surgical pathology. American journal of clinical pathol-ogy , 136(5):767–772, 2011.[19] G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi,M. Ghafoorian, J. A. Van Der Laak, B. Van Ginneken, andC. I. S´anchez. A survey on deep learning in medical imageanalysis.
Medical image analysis , 42:60–88, 2017.[20] J. Lonsdale, J. Thomas, M. Salvatore, R. Phillips, E. Lo,S. Shad, R. Hasz, G. Walters, F. Garcia, N. Young, et al. Thegenotype-tissue expression (gtex) project.
Nature genetics ,45(6):580, 2013.[21] M. T. McCann, J. A. Ozolek, C. A. Castro, B. Parvin, andJ. Kovacevic. Automated histology analysis: Opportunitiesfor signal processing.
IEEE Signal Processing Magazine ,32(1):78–87, 2015.[22] G. J. McLachlan and D. Peel.
Finite Mixture Models . JohnWiley & Sons, 2004.[23] R. Mormont, P. Geurts, and R. Mar´ee. Comparison of deeptransfer learning strategies for digital pathology. In
Proceed-ings of the IEEE Conference on Computer Vision and PatternRecognition Workshops , pages 2262–2271, 2018.[24] L. Pantanowitz. Digital images and the future of digitalpathology.
Journal of pathology informatics , 1, 2010.[25] L. Pantanowitz, P. Michelow, S. Hazelhurst, S. Kalra,C. Choi, S. Shah, M. Babaie, and H. R. Tizhoosh. A digitalpathology solution to resolve the tissue floater conundrum.
Archives of Pathology & Laboratory Medicine , 2020.[26] E. Platt, P. Sommer, L. McDonald, A. Bennett, and J. Hunt.Tissue floaters and contaminants in the histology laboratory.
Archives of pathology & laboratory medicine , 133(6):973–978, 2009.[27] S. Ragothaman, S. Narasimhan, M. G. Basavaraj, and R. De-war. Unsupervised segmentation of cervical cell images us-ing gaussian mixture model. In
Proceedings of the IEEEConference on Computer Vision and Pattern RecognitionWorkshops , pages 70–75, 2016.[28] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,et al. Imagenet large scale visual recognition challenge.
International Journal of Computer Vision , 115(3):211–252,2015.[29] M. Shaban, R. Awan, M. M. Fraz, A. Azam, Y.-W. Tsang,D. Snead, and N. M. Rajpoot. Context-aware convolutionalneural network for grading of colorectal cancer histology im-ages.
IEEE transactions on medical imaging , 39(7):2395–2405, 2020.[30] P. Y. Simard, D. Steinkraus, and J. C. Platt. Best practices forconvolutional neural networks applied to visual documentanalysis. In null , page 958. IEEE, 2003.[31] K. Simonyan and A. Zisserman. Very deep convolutionalnetworks for large-scale image recognition. arXiv preprintarXiv:1409.1556 , 2014. [32] F. A. Spanhol, L. S. Oliveira, C. Petitjean, and L. Heutte.Breast cancer histopathological image classification usingconvolutional neural networks. In
Neural Networks (IJCNN),2016 International Joint Conference on , pages 2560–2567.IEEE, 2016.[33] A. Sridhar, S. Doyle, and A. Madabhushi. Content-basedimage retrieval of digitized histopathology in boosted spec-trally embedded spaces.
Journal of pathology informatics , 6,2015.[34] Z. Swiderska-Chadaj, T. de Bel, L. Blanchet, A. Baidoshvili,D. Vossen, J. van der Laak, and G. Litjens. Impact ofrescanning and normalization on convolutional neural net-work performance in multi-center, whole-slide classificationof prostate cancer.
Scientific RepoRtS , 10(1):1–14, 2020.[35] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna.Rethinking the inception architecture for computer vision. In
Proceedings of the IEEE conference on computer vision andpattern recognition , pages 2818–2826, 2016.[36] H. R. Tizhoosh, C. Mitcheltree, S. Zhu, and S. Dutta. Bar-codes for medical image retrieval using autoencoded radontransform. In , pages 3150–3155. IEEE, 2016.[37] H. R. Tizhoosh, L. Pantanowitz, et al. Artificial intelligenceand digital pathology: Challenges and opportunities.
Journalof Pathology Informatics , 9(1):38, 2018.[38] T. A. A. Tosta, P. R. de Faria, L. A. Neves, and M. Z.do Nascimento. Computational normalization of h&e-stained histological images: Progress, challenges and futurepotential.
Artificial intelligence in medicine , 95:118–132,2019.[39] M.-H. Tsai, Y.-K. Chan, Z.-Z. Lin, S.-F. Yang-Mao, and P.-C. Huang. Nucleus and cytoplast contour detector of cervicalsmear image.
Pattern Recognition Letters , 29(9):1441–1453,2008.[40] D. Wang, A. Khosla, R. Gargeya, H. Irshad, and A. H. Beck.Deep learning for identifying metastatic breast cancer. arXivpreprint arXiv:1606.05718 , 2016.[41] J. N. Weinstein, E. A. Collisson, G. B. Mills, K. R. M. Shaw,B. A. Ozenberger, K. Ellrott, I. Shmulevich, C. Sander, J. M.Stuart, C. G. A. R. Network, et al. The cancer genome atlaspan-cancer analysis project.
Nature genetics , 45(10):1113,2013.[42] J. Xu, X. Luo, G. Wang, H. Gilmore, and A. Madabhushi. Adeep convolutional neural network for segmenting and clas-sifying epithelial and stromal regions in histopathologicalimages.
Neurocomputing , 191:214–223, 2016.[43] Y. Xu, Z. Jia, L.-B. Wang, Y. Ai, F. Zhang, M. Lai, I. Eric,and C. Chang. Large scale tissue histopathology image clas-sification, segmentation, and visualization via deep convolu-tional activation features.
BMC bioinformatics , 18(1):281,2017.[44] S.-F. Yang-Mao, Y.-K. Chan, and Y.-P. Chu. Edge en-hancement nucleus and cytoplast contour detector of cervicalsmear images.
IEEE Transactions on Systems, Man, and Cy-bernetics, Part B (Cybernetics) , 38(2):353–366, 2008., 38(2):353–366, 2008.