Table 2.
Summary of AI models for CRC diagnosis and staging session. (CNN, convolutional neural network; AUC, area under the curve; SVM, support vector machine; PNN, probabilistic neural network; NL, normal mucosa; AD, adenoma; ADC, adenocarcinoma; WSI, whole slide images; RNN, recurrent neural network; TCGA, The Cancer Genome Atlas; ResNet, residual network architecture; HP, hyperplastic polyp; VGG, visual geometry group; RF, random forest; PET-CT, positron emission tomography or computed tomography; LR, logistic regression; NN, neural network; XGBoost, extreme gradient boosting; CT, computed tomography; MRI, magnetic resonance imaging; Faster R-CNN, faster region-based CNN; CAD, computer aided diagnosis).
| Topic | Task | Dataset | Model | Performance | Year | Ref. |
|---|---|---|---|---|---|---|
| Pathological diagnosis |
Tumor mutational burden-high prediction | 278 HE slides | CNN | AUC: 0.934 | 2021 | [41] |
| Low/high-grade classification | Immunohistochemically stained biopsy of 67 patients | hDL-system (VGG16, SVM) | hDL-system accuracy: 99.1%; sML-system accuracy: 92.5% | 2021 | [42] | |
| NL/AD/ADC classification | 4036 WSI | CNN, RNN | AUC: 0.96 for ADC; 0.99 for AD | 2020 | [43] | |
| Tumor immune microenvironment analysis | 404 CRC and 20 adjacent non-tumorous tissues | CIBERSORT | C-index: stage I-II 0.69; stage III-IV 0.71; AUC: 0.67 | 2019 | [44] | |
| NL/Tumor classification | 94 WSI, 370 TCGA-KR, 378 TCGA-DX | ResNet18 | AUC > 0.99 | 2019 | [45] | |
| NL/HP/AD/ADC classification |
393 WSI (12,565 patches) | CNN | Accuracy: 80% | 2019 | [46] | |
| NL/Tumor classification | 57 WSI (10,280 patches) |
VGG | Accuracy: 93.5%, Sensitivity: 95.1% |
2018 | [47] | |
| NL/AD/ADC classification | 27 WSI (13,500 patches) |
VGG16 | Accuracy: 96%, Specificity: 92.8% | 2018 | [48] | |
| NL/AD/ADC classification | 30 multispectral image patches | CNN | Accuracy: 99.2% | 2017 | [49] | |
| Cancer subtypes classification | 717 patches | AlexNet | Accuracy: 97.5% | 2017 | [50] | |
| Polyp subtypes classification | 2074 patches 936 WSI | ResNet | Accuracy: 93.0% | 2017 | [51] | |
| Radiological diagnosis |
Metastatic CRC prediction | MRI from 55 stage VI patients with known hepatic metastasis | RF | AUC: 0.94 (Add imaging-based heterogeneity features) | 2021 | [52] |
| Metastatic lymph node prediction | PET-CT scan images from 199 CRC patients | LR, SVM, RF, NN, and XGBoost | AUC of LR: 0.866; AUC of XGBoost: 0.903 | 2021 | [53] | |
| Colorectal liver metastasis prediction | 103 metastasis samples and 80 non-cancer tissues | Probe electrospray ionization-mass spectrometry, and LR | Accuracy: 99.5%, AUC: 0.9999 |
2021 | [54] | |
| Colorectal liver metastasis prediction | CT scan images from 91 patients | Bayesian-optimized RF with wrapper feature selection | AUC of radiomics features model: 86%; AUC of clinical features model: 71%; AUC of combination: 86% |
2021 | [55] | |
| KRAS mutations detection | CT scan images from 47 patients | Haralick texture analysis, SVM, LightGBM, NN, and RF | Accuracy: 83%, kappa: 64.7% | 2020 | [56] | |
| Classification of T2 and T3 | 290 MRI images from 133 patients | CNN | Accuracy: 0.94 | 2019 | [57] | |
| Metastatic lymph node prediction | MRI images from 414 patients | Faster R-CNN | r-radiologist-Faster R-CNN 0.912 | 2019 | [58] | |
| Polyp detection | 825 CT scan images | CNN | Accuracy: 0.87, Sensitivity: 0.8877, Specificity: 0.8735 |
2017 | [59] | |
| Polyp detection | 154 CT scan images | CNN | Accuracy: 0.971 | 2017 | [60] | |
| Polyp classification | 1035 endomicroscopy images | Mathworks “NAVICAD” system | Accuracy: 84.5% | 2016 | [61] | |
| Polyp detection and classification | 148 CT scan images | Haralick texture analysis, SVM | ROC: 0.85 | 2014 | [62] | |
| CAD system for polyp detection | 24 T1 stage patients’ CT scan images | Coloncad API 4.0, Medicsight plc | True positives rate >96.1% | 2008 | [63] |