Table I.
Deep learning models for diagnosis and classification of colorectal cancer
| Author | Year | Country | Study type | Aim of model | Architecture | Task | Dataset used | Evaluation metrics |
|---|---|---|---|---|---|---|---|---|
| Kainz [34] | 2017 | Austria | Journal article | Classification of glands | Custom CNN | 4-class (benign, benign background, malignant, malignant background) | 2015 MICCAI Gland Segmentation Challenge, Training 85 Images (37 benign and 48 malignant) Testing 80 (37/43) |
Detection results (MICCAI Glas): F1 score = (0.68 + 0.61)/2, DICE index (0.75 + 0.65)/2, Hausdorff (103.49 + 187.76)/2 |
| Xu [35] | 2017 | China | Journal article | Gland Classification | Custom CNN | Binary (bening/malignant) | 2015 MICCAI Gland Segmentation Challenge Training 85 Images Testing 80 Images |
Detection results (MICCAI Glas): F1 score (0.893 + 0.843)/2, DICE index (0.908 + 0.833)/2, Hausdorff (44.129 + 116.821)/2 |
| Awan [36] | 2017 | UK | Journal article | Grading of CRC | UNET | (A) Binary (normal/cancer) (B) 3-class: normal/low grade/high grade | 38 WSIs, extracted 139 parts (71 normal, 33 low grade, 35 high grade) | (A) Binary ACC: 97% (B) 3-vlass ACC: 91% |
| Chen [37] | 2017 | China | Journal article | Nuclei Classification | Custom CNN | Binary (benign/malignant) | 2015 MICCAI Gland Segmentation Challenge and 2015 MICCAI Nuclei Segmentation Challenge | Detection results (MICCAI Glas): F1 score = 0.887, DICE index 0.868 Hausdorff = 74.731 Segmentation results: D1 and D2 metrics from Challenge |
| Xu [38] | 2017 | China | Journal article | Tissue Classification | Alexnet – SVM | (1) Binary (cancer/not cancer) | 2014 MICCAI Brain Tumour (incl. colorectal metastases) Digital Pathology Challenge | ACC: (1) Binary: 98% |
| Haj-Hassan [39] | 2017 | France | Journal article | Tissue Classification | Custom CNN | Triple: benign hyperplasia (BH), intraepithelial neoplasia (IN), and carcinoma (Ca) | Biopsy images of 30 CRC patients | ACC: 99.17% |
| Chaofeng | 2017 | China | Journal article | Tissue Classification | Custom Bilinear CNN | Eight types of tissue, namely tumour epithelium, simple stroma, complex stroma, immune cells, debris, normal mucosal glands, adipose tissue, and background (no tissue) | H&E-stained colorectal cancer histopathological image dataset from the University Medical Centre Mannheim | AUC value of 0.985 |
| Van Eycke [40] | 2018 | Belgium | Journal article | Diagnosis | VGG | Binary (benign/malignant) | 2015 MICCAI Gland Segmentation Challenge Training 85 Images Testing 80 (37/43) |
Detection results (MICCAI Glas): F1 score = (0.895 + 0.788)/2, DICE index (0.902 + 0.841)/2, Hausdorff (42.943 + 105.926)/2 |
| Graham [41] | 2019 | UK | Journal article | Diagnosis | Custom Architecture MILD-net | Binary (benign/malignant) | (1) MICCAI Gland Segmentation Challenge (2) 38 WSIs, extracted 139 parts (71 normal, 33 low grade, 35 high grade) | (1) F1 score: (0.914 + 0.844)/2, Dice: (0.913 + 0.836)/2, Hausdorff (41.54 + 105.89)/2 (2) F1 score: 0.825, Dice: 0.875, Hausdorff: 160.14 |
| Yoon [42] | 2019 | South Korea | Journal article | Diagnosis | Modified-VGG | Binary (benign/malignant) | Centre for CRC, National Cancer Centre, Korea, 57 WSIs, 10,280 patches | ACC: 93.48%, SP: 92.76%, SE: 95.1% |
| Qaiser [9] | 2019 | UK | Journal article | Diagnosis | Custom CNN | Binary (benign/malignant) | (1) Warwick-UHCW 75 H/E WSIs (112,500 patches), (2) Warwick-Osaka 50 H/E WSIs (75,000 patches) | (A) PHP/CNN: F1 score 0.9243, Precision 0.9267 (B) PHP/CNN: F1 score 0.8273, Precision 0.8311 |
| Sari [43] | 2019 | Turkey | Journal article | Grading of CRC | Feature Extraction from Deep Belief Network and classification employing linear SVM, Comparison with Alexnet, GoogleNet, Inceptionv3, and autoencoders | (1) 3-class: normal (N), Low Grade (LG), High Grade (HG) (2) 5-class: Normal, Low (1), Low (1–2), Low (2), High | (1) 3236 images 1001 N, 1703 LG, 532 (HG) (2) 1468 images | (1) mean ACC: 96.13 (2) mean ACC: 79.28 |
| Rączkowski [44] | 2019 | Poland | Journal article | Tissue Classification | Custom CNN | (A) Binary: tumour/stroma (B) 8-class: tumour epithelium, simple stroma, complex stroma, immune cells, debris, normal mucosal glands, adipose tissue, background | 5000 patches | (1) AUC 0.998 ACC: 99.11 ±0.97% (2) AUC 0.995 ACC: 92.44 ±0.81% |
| Sena [26] | 2019 | Italy | Journal article | Tissue Classification | Custom CNN | Normal mucosa, preneoplastic lesion, adenoma, cancer | 393 WSIs | ACC: 81.7 |
| Xu [35] | 2020 | Canada | Journal article | Diagnosis | Custom CNN | Binary (benign/malignant) | St. Paul’s Hospital, 307 H/E images | ACC: 99.9% (normal slides), ACC: 94.8% (cancer slides) Independent dataset: median ACC: 88.1%, AUROC: 0.99 |
| Song [45] | 2020 | China | Journal article | Diagnosis | Custom CNN | Binary (benign/malignant) | 579 slides | ACC: 90.4, AUC: 0.92 |
| Shaban [18] | 2020 | UK | Journal article | Grading of CRC | Custom CNN | 3-Class: normal, low grade, high grade | 38 WSIs, extracted 139 parts (71 normal, 33 low grade, 35 high grade) | ACC: 95.70 |
| Iizuka | 2020 | Japan | Journal article | Tissue Classification | (1) Inception v3, (2)RNN | 3-class: adenocarcinoma/adenoma/non-neoplastic | 4.036 WSIs | (1) AUC: (ADC: 0.967, adenoma: 0.99), (2) AUC: (ADC: 0.963, adenoma: 0.992) |
| Masud [46] | 2021 | Saudi Arabia | Journal article | Diagnosis | Custom CNN | Binary (benign/malignant) | LC25000 dataset, James A. Haley Veterans’ Hospital, 5000 images of Colon ADC, 5000 images of Colon benign tissue | ACC: 96.33% F-measure score 96.38% for colon and lung cancer identification |
| Wang [6] | 2021 | China | Journal article | Diagnosis | Modified Inception v3 | Binary (bening/malignant) | 14,234 CRC WSIs and 170.099 patches | ACC: 98.11%, AUC: 99.83%, SP: 99.22%, SE: 96.99% |
| Gupta [14] | 2021 | Switzerland | Journal article | Diagnosis | (a) VGG, ResNet, Inception, and IR-v2 for transfer learning, (b) Five types of customized architectures based on Inception-ResNet-v | Binary (benign/malignant) | 215 H/E WSIs, 1.303.012 patches | (a) IR-v2 performed better than the others: AUC: 0.97, F-score: 0.97 (b) IR-v2 Type 5: AUC: 0.99, F-score: 0.99 |
| Yu [10] | 2021 | China | Journal article | Diagnosis | SSL | Binary (benign/malignant) | 13.111 WSIs, 62,919 patches | Patch-level diagnosis AUC: 0.980 ±0.014 Patient-level diagnosis AUC: 0.974 ±0.013 |
| Toğaçar [47] | 2021 | Turkey | Journal article | Diagnosis | YOLO-based DarkNet-19 | Binary (benign/malignant) | 10,000 images | Overall ACC: 99.69% |
| Terradillos [48] | 2021 | Spain | Journal article | Diagnosis | Custom CNN based on Xception | Binary (benign/malignant) | 14,712 images | SE: 0.8228, SP: 0.9114 |
| Paladini [28] | 2021 | Italy | Journal article | Tissue Classification | 2 × Ensemble approach ResNet-101, ResNeXt-50, Inception-v3 and DensNet-161. (1) Mean-Ensemble-CNN approach, the predicted class of each image is assigned using the average of the predicted probabilities of 4 trained models. (2) In the NN-Ensemble-CNN approach, the deep features corresponding to the last FC layer are extracted from the 4 trained models | 7-class, 8-class, respectively | WIS-Kather-CRC-2016 Database (5000 CRC images) and CRC-TP Database (280,000 CRC images) | Kather-CRC-2016 Database: Mean-Ensemble-CNN mean ACC: 96.16% NN-Ensemble-CNN mean ACC: 96.14% CRC-TP Database: Mean-Ensemble-CNN ACC: 86.97% Mean-Ensemble-CNN F1-Score: 86.99% NN-Ensemble-CNN ACC: 87.26% NN-Ensemble-CNN F1-Score: 87.27% |
| Ben Hamida [24] | 2021 | France | Journal article | Tissue Classification | (1) Comparison of 4 different architectures Alexnet, VGG-16, ResNet, DenseNet, Inceptionv3, with transfer learning strategy (2) Comparison of SegNet and U-Net for semantic Segmentation | (A) 8-class: tumour, stroma, tissue, necrosis, immune, fat, background, trash (B) Binary (tumour/no-tumour) | (1) AiCOLO (396 H/E slides), (2) NCT Biobank, University Medical Centre Mannheim (100,000 H/E patches), (3) CRC-5000 dataset (5000 images), (4) Warwick (16 H/E) | (1) ResNet On AiCOLO-8: overall ACC: 96.98% On CRC-5000: ACC: 96.77% On NCT-CRC-HE-100: ACC: 99.76% On merged: ACC: 99.98% (2) On AiCOLO-2 UNet: ACC: 76.18%, SegNet: ACC: 81.22% |
| Zhou [49] | 2021 | China | Journal article | Tissue Classification | Custom CNN with Res-Net | Binary (benign/malignant) | TCGA 1346 H/E WSIs | ACC: 0.946 Precision: 0.9636 Recall: 0.9815 F1 score: 0.9725 |
| Riasatian [30] | 2021 | Canada | Journal article | Tissue Classification | Custom CNN based on DenseNEt | 8-class: tumour epithelium, simple stroma, complex stroma, immune cells, debris, normal mucosal glands, adipose tissue, background | The Cancer Genome Atlas’ 5000 patches | ACC: 96.38% (KN-I) and 96.80% (KN-IV) |
| Tsuneki [25] | 2021 | Switzerland | Journal article | Tissue Classification | EfficientNetB1 model | 4-class: poorly differentiated ADC, well-to-moderately differentiated ADC, adenoma, non-neoplastic) | 1799 H/E WSIs | AUC: 0.95 |
| Jiao [29] | 2021 | China | Journal article | Tissue Classification | DELR based | 8-class: normal mucosa, adipose, debris, lymphocytes, mucus, smooth muscle, stroma, tumour epithelium | 180,082 patches | AUC: > 0.95 |
| Kim [27] | 2021 | South Korea | Journal article | Tissue Classification | Custom CNN | 5-class: ADC, high-grade adenoma with dysplasia, low-grade adenoma with dysplasia, carcinoid, hyperplastic polyp | 390 WSIs | ACC: 0.957 ±0.025 Jac: 0.690 ±0.174 Dice: 0.804 ±0.125 |
| Yan [19] | 2022 | China | Journal article | CRC Grading | Custom Divide-and-Attention Network (DANet) + Majority Voting | 3-class (normal, low high grade) | 15,303 patches as proposed by Awan et al. | ACC: 95.3%, AUC: 0.94 |
| Dabass [20] | 2022 | India | Journal article | CRC Grading | Custom CNN based on Enhanced Convolutional Learning Modules (ECLMs), multi-level Attention Learning Module (ALM), and Transitional Modules (TMs) | 3-class (normal, low high grade) | Gland Segmentation challenge (GlaS), Lung Colon(LC)-25000, Kather_Colorectal_Cancer_Texture_Images (Kather-5k), NCT_HE_CRC_100K(NCT-100k) and a private dataset Hospital Colon (HosC) | GlaS (Accuracy (97.5%), Precision (97.67%), F1-Score (97.67%), and Recall (97.67%)), LC-25000 (Accuracy (100%), Precision (100%), F1-Score (100%), and Recall (100%)), and HosC (Accuracy (99.45%), Precision (100%), F1-Score (99.65%), and Recall (99.31%)), and while for the tissue structure classification, it achieves results for Kather-5k (Accuracy (98.83%), Precision (98.86%), F1-Score (98.85%), and Recall (98.85%)) and NCT-100k (Accuracy (97.7%), Precision (97.69%), F1-Score (97.71%), and Recall (97.73%)) |
| Kassani [13] | 2022 | USA | Journal article | Diagnosis | Comparison of ResNet, MobileNet, VGG, Inceptionv3, InceptionResnetv2, ResNeXt, SE-ResNet, SE-ResNeXt | Binary (Healthy/Cancer) | DigestPath, 250 H/E WSIs, 1.746 patches | Dice: 82.74% ACC: 87.07% F1 score: 82.79% |
| Shen [50] | 2022 | China | Journal article | Diagnosis | Custom CNN based on DenseNEt | 3-class: loose non-tumour tissue, dense non-tumour tissue, gastrointestinal cancer tissues | 1063 TCGA samples | DP-FTD: AUC 0.779, DCRF-FTD: AUC: 0.786 |
| Chehade [15] | 2022 | France | Journal a rticle | Diagnosis | XGBoost, SVM, RF, LDA, MLP and LightGBM | Binary (benign/malignant) for CRC | LC25000 dataset | ACC of 99% and a F1-score of 98.8% for XGBoost |
| Collins [16] | 2022 | Italy | Prospective study | Diagnosis | Custom CNN | 3-class (normal, T1-2, T3-4) | Images from 34 patients | 15-fold cross-validation (Se: 87% and Sp: 90%, respectively), ROC-AUC: 0.95. T1-2 group Se: 89%, Sp: 90%, T3-4 group, Se: 81%, Sp: 93% |
| Ho [12] | 2022 | Singapore | Journal article | Diagnosis | Faster-RCC Architecture | Binary (low risk/high risk) | 66,191 image tiles extracted from 39 WSIs, Evaluation 150 WSI biopsies | AUC of 91.7% |
| Reis [51] | 2022 | Turkey | Journal article | Nuclei Classification | Custom method (DenseNet169 + SVM, DenseNet169 + GRU) | 10-class | 10-class MedCLNet visual dataset consisting of the NCT-CRC-HE-100 K dataset, LC25000 dataset, and GlaS dataset | 95% accuracy was obtained in the DenseNet169 model after pre-train |
| Albashish | 2022 | Jordan | Journal article | Tissue Classification | (1) E-CNN (product rule), (2) E-CNN (majority voting) | 4-class and 7 class | Stoean (357 images) and Kather colorectal histology (5000 images) | (1) ACC: 97.20%, (2) 91.28% |
| Li [31] | 2022 | Hong Jong | Journal article | Tissue Classification | Pretrained ImageNet | 9 class | 100,000 annotated H&E image patches | ACC: 98.4% |