Table 1.
Summary of the studies in all aspects analyzed in this review.
| Author | Year | Sample | Number of classes | Diagnosisa | Dataseta | Training setb | Test set | External test set | Pre-processinga | Model (Patch level)a | Model (Slide level)a | Transfer learning | Training approach | Resultsa | Results of the external test seta |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Lucas et al.33 | 2019 | Prostate | 4 | Cancer | Private | 268 000 patches | 89 000 patches | – | Data Augmentation | InceptionV3 + SVM | Percentages of GPs used for final Gleason grade | No | Supervised | Kappa: 0.70 | - |
| Pantanowitz et al.40 | 2020 | Prostate | 18 | Cancer | Private | 549 WSIs | 2501 WSIs | 1627 WSIs | Tissue segmentation and data augmentation | InceptionV1, InceptionV3 and ResNet101 | Maximum score | Yes | Supervised | AUC: 0.997g | AUC: 0.991, 0.941, 0.971, and 0.957h |
| Ström et al.49 | 2020 | Prostate | 2 and 4a | Cancer | Private | 1069 WSIs | 246 WSIs | 73 WSIs | Tissue segmentation and data augmentation | 30 InceptionV3 models | Boosted tree | Yes | Supervised | Kappa: 0.83 | Kappa: 0.70 |
| BenTaieb et al.6 | 2017 | Ovary | 5 | Cancer | Public | 68 WSIs | 65 WSIs | – | – | K-means | LSVM | Yes | Weakly Supervised | Kappa: 0.89 | – |
| Barker et al.4 | 2016 | Central nervous system | 2 | Cancer | Public | 302 WSIs | 45 WSIs | 302 WSIs | Tissue segmentation, color deconvolution and nuclei segmentation | – | Feature Extraction + Elastic Net (Regression) | No | Weakly supervised | Accuracy: 1.0 | Accuracy: 0.93 |
| Xu et al.60 | 2017 | Central nervous system | 2 | Cancer | Public | 55 WSIs | 40 WSIs | - | Tissue segmentation, resize and data augmentation | Customized AlexNet | Feature Pooling + SVM | Yes | Supervised | Accuracy: 0.975 | – |
| Bulten et al.7 | 2020 | Prostate | 7 | Cancer | Private | 933 WSIS | 210 WSIs | – | Tissue segmentation and data augmentation | Own CNN to detect tumor and U-Net to final label | Normalized percentage of the volume of each class | No | Supervised (with a semi-automatic annotation) | Kappa: 0.819 on Gleason score | – |
| Gecer et al.17 | 2018 | Breast | 5 | Cancer | Private | 180 WSIs | 60 WSIs | – | Color Normalization | RoI detector and an own proposed CNN | Majority voting | No | Weakly supervised | Accuracy: 0.55 | – |
| Silva-Rodríguez et al.46 | 2020 | Prostate | 4 and 1a | Cancer | Public | 155 WSIs | 2122 patches | - | Tissue segmentation and data augmentation | Own CNN | MLP | No and yesa | Supervised | Kappa: 0.732 | – |
| Tokunaga et al.53 | 2019 | Gastric | 4 | Cancer | – | 29 WSIs | – | – | Data augmentation | AWMF-CNN | Aggregating CNN | No | Supervised | IoU (Mean): 0.536 | – |
| Sali et al.43 | 2019 | Small intestine | 4 | Celiac disease | Private | 336 WSIs | 120 WSIs | - | Tissue segmentation, color normalization, resize and data augmentation | Customized Resnet50 | Sum of all labels and majority | No | Weakly Supervised | Accuracy: 1.0 | - |
| Xu et al.61 | 2020 | Prostate | 3 | Cancer | Public | 312 WSIs | 49,883 patches | – | Grayscale and tissue segmentation | Feature extractor | PCA and SVM | No | Weakly Supervised | Accuracy: 0.771 | – |
| Mercan et al.34 | 2018 | Breast | 14 | Cancer | Private | 240 WSIs | 60 WSIs | – | – | Feature extractor + Linear classifier | PCA and SVM | No | Weakly supervised | Average precision: 0.737 | – |
| Adnan et al.1 | 2020 | Lung | 2 | Cancer | Public | 1026 WSIs | – | – | RoI selection | Feature extractor | GCN | No and yesa | Weakly supervised | 0.89 AUCf | – |
| van Zon et al.56 | 2020 | Skin | 3 | Cancer | Private | 232 WSIs | 331 WSIs | - | Tissue segmentation and data augmentation | U-Net | Own CNN | No | Supervised | 0.954 Accuracyd | – |
| Wang et al.57 | 2019 | Lung | 4 | Cancer | Private | 754 WSIs | 185 WSIs | - | Tissue segmentation, resize and data augmentation | ScanNet | Aggregation of patch preditcions values + Random foresta | No | Weakly supervised | Accuracy: 0.973 | – |
| Syrykh et al.51 | 2020 | Lymph node | 2 | Cancer | Private | 75% of 378 WSIs | 25% of 378 WSIs | 48 Cases | Tissue segmentation | CNNa | Average of patch inferences | – | Weakly supervised | AUC: 0.99 | AUC: 0.69 |
| Wei et al.58 | 2019 | Small intestine | 3 | Celiac disease | Private | 1,018 WSIs | 212 WSIs | - | Data augmentation and color normalization | ResNet50 | Threshold to discard low confidence + Most frequent predicted class | Yes | – | Average F1 score: 0.872 | – |
| Korbar et al.28 | 2017 | Small intestine | 6 | Colorectal polyps | Private | 458 WSIs | 239 WSIs | - | Data augmentation, color normalization and resize | ResNet-D | At least 5 positive class patches with 70% of confidence | No | Supervised | Overall F1 score: 0.888 | – |
| Nagpal et al.37 | 2019 | Prostate | 4 | Cancer | Public and private | 1,226 WSIs | 331 WSIs | - | Data augmentation | Customized inception V3 | K-nearest neighbor model from patch prediction | No | Supervised | Gleason Score Accuracy: 0.70 | – |
| Olsen et al.39 | 2018 | Skin | 3 models with 2 classes | Cancer | Private | Study 1: 300 WSIs | Study 1: 126 WSIs | – | Tissue segmentation | Derivative VGG + Rule-based discriminator | Classification model trained with the segmented areasa | No | Supervised | Study 1 Accuracy: 0.9945 | – |
| Study 2: 225 WSIs | Study 2: 114 WSIs | Study 2 Accuracy: 0.994 | |||||||||||||
| Study 3: 225 WSIs | Study 3: 123 WSIs | Study 3 Accuracy: 1.0 | |||||||||||||
| Wei et al.59 | 2019 | Lung | 6 | Cancer | Private | RoIs from 279 WSIs | 143 WSIs | – | Tissue segmentation, data augmentation and color normalization | ResNet18 | Threshold to discard low confidence + Most frequent predicted class | Yes | Supervised | Kappa Score: 0.525 | - |
| Ianni et al.20 | 2020 | Skin | 4 | Cancer | Private | 85% of 5070 WSIs | 15% of 5,070 WSIs | 13,537 WSIs | – | Own Enconder-Decoder CNN + U-Net | Own CNN | No | Supervised (Patch) and Weakly Supervised (Slide) | – | Accuracy: 0.98 |
| Iizuka et al.21 | 2020 | Stomach & Small intestine | 2 models with 3 classes | Cancer | Private | Stomach: 3,628 WSIs | Stomach & Colon: 500 WSIs | Stomach & Colon: 500 WSIs | Tissue segmentation and data augmentation | Customized Inception V3 | RNN using the last but one layer from the previous model as input | No | Supervised | AUC e: | AUC e: |
| Stomach: 0.97 and 0.99 | Stomach: 0.98 and 0.93 | ||||||||||||||
| Colon: 3,536 WSIs | Colon: 0.96 and 0.99 | Colon: 0.97 and 0.96 | |||||||||||||
| Campanella et al.8 | 2019 | Skin | 2 | Cancer | Private | 8387 WSIsc | 1575 WSIsc | – | – | ResNet34 | RNN using the last but one layer from the previous model as input | No | Weakly supervised | AUC: 0.994 | – |
| Chuang et al.9 | 2020 | Larynx, lip and oral cavity, esophagus, pharynx | 3 | Cancer | Private | 626 Cases | 100 Cases | – | – | ResNetXt | ResNet using the probability map as input | Yes | Supervised | AUC: 0.985 | – |
Captions – Not mentioned or not performed
a
Details can be found in the Supplementary Table
b
Training and validation set used during training was considered as training set in this column
c
Not clearly specified, only the test set size and the whole dataset size, this number was estimated with these 2 information
d
No metrics were performed by the authors in terms of final diagnosis, we calculated this metric using the table of misclassifcation comparison
e
AUC of adenocarcinoma and adenoma compared to benign, respectively
f
This study used the same model in 2 different tasks of lung carcinoma, one in a private set with 4 classes, and another in the TCGA differentiating 2 classes. We considered the most complex task.
g
Authors performed only the Benign vs. Cancer AUC in the internal test set.
h
Metrics representing: Benign vs Cancer, Gleason score 6 or ASAP vs Gleason score 7–10, ASAP or Gleason pattern 3 or 4 vs Gleason pattern 5, Cancer without vs with perineural invasion, respectively