Table 1.
The summary of the application of AI in the diagnosis of CRC.
| Theme | Year | Subject | Model | Sample | Result | Ref |
|---|---|---|---|---|---|---|
| Endoscopic diagnosis | 2018 | Polyp detection | CNN | 155 videos | AUC = 0.87 | (40) |
| 2019 | Neoplasia detection | CNN | 685 subjects | ADR = 54.8% | (41) | |
| 2020 | Adenoma detection | DL | 386 patients | AMR = 13.89% | (42) | |
| 2022 | Neoplasia detection | CNN | 230 subjects | AMR = 15.5% | (43) | |
| 2020 | Polyp and adenoma detection | CNN | 308 patients | ADR = 0.289;0.367(polyp; adenoma) | (44) | |
| 2020 | Polyp detection | YOLO | 150 patients | PDR = 38.7% | (45) | |
| 2022 | Polyp and adenoma detection | CNN | 1,434 patients | PDR = 40.8% ADR = 20.1% | (46) | |
| 2021 | Adenoma detection | CADe | 1,076 patients | ADR = 21.27% | (47) | |
| 2021 | Polyp | DL | 2,352 patients detection | PDR = 38.8% | (48) | |
| 2019 | Polyp and adenoma detection | DL | 522 patients | ADR = 29.1% PDR = 64.93% | (49) | |
| 2021 | Adenoma detection | CNN | 358 patients | AMR = 13.8% | (50) | |
| Non-invasive screening | 2022 | Cancer diagnosis and stage | RF/SVM/DT | 521 samples | Average accuracy = 99.81% F1 value = 0.9968 accuracy = 99.88%recall = 99.5% | (51) |
| 2021 | Biomarkers screening | SVM/LR/RF/kNN/NB | 1,164 electronic medical records | AUC = 0.849 | (52) | |
| 2019 | Cancer detection | LR/SVM | 817 plasma samples | Mean AUC = 0.92 mean sensitivity = 85% specificity = 85% | (53) | |
| 2020 | Cancer screening | ML | 289 healthy individuals and 983 patients | Specificity = 0.89 sensitivity = 0.72 | (54) | |
| 2019 | Mutation detection | CP-ANN | 312 tissue samples | Sensitivity = 100% specificity = 87.5% accuracy = 93.8% | (55) | |
| 2020 | Gene detection | DL | 8,836 samples | Mean AUROC = 0.92 AUPRC = 0.63 | (56) | |
| 2020 | Gene identification | LASSO | 480 CRC and 41 normal tissues | AUC = 0.6923 (training set; 3-year) AUC = 0.7328 (training set; 5-year) AUC = 0.6803 (testing set; 3-year) AUC = 0.7035 (testing set; 5-year) | (57) | |
| 2021 | Gene detection | CGANs | 256 patients (training cohort 1) 1457 patients (training cohort 2) | AUROC = 0.742 (training cohort 1) AUROC = 0.757 (training cohort 2) AUROC = 0.743 (synthetic data) AUROC = 0.777 (mixed data) | (58) | |
| Histopathologic diagnosis | 2021 | Image learning | DELR | 500–3,000 samples | AUC > 0.95 | (59) |
| 2022 | Histopathologic segmentation | CNN/TL | 25 WSIs | DSI = 82.74% ± 1.77 accuracy = 87.07% ± 1.56 f1-score value = 82.79% ± 1.79 | (60) | |
| 2021 | Distinguish CRLM | DL/CNN/ICC | 93 CRLM patients | AUC = 0.69 | (61) | |
| 2020 | Histopathologic classification | CNN/RNN | 4,036 WSIs | AUC = 0.96; 0.99(adenocarcinoma; adenoma) | (62) | |
| 2021 | Histopathologic segmentation | PCA/DWT | 351 specimens | Dice = 0.804 ± 0.125 | (63) | |
| 2021 | Image classification | ANN/SVM | 5,000 histopathology image tiles | Performance accuracy = 95.3% | (64) | |
| 2022 | Histopathologic screening | DL/ML | 294 WSIs | AUC = 0.917 sensitivity = 97.4% | (65) | |
| 2022 | Histopathologic classification | DL/CNN | 1,865 pathological images | AUC = 0.995; 0.998 | (66) | |
| 2022 | Image grading | CNN/HCCANet | 630 images | Overall accuracy = 87.3% average AUC = 0.9 | (67) | |
| 2019 | Survival prediction | TL/CNN | 862 HE images | Accuracy>94% | (68) | |
| 2019 | Cancer diagnosis | CNN/RE/kNN/LR/NB/SVM | 357 images | Accuracy = 87–95% | (69) | |
| Radiologic diagnosis | 2019 | Preoperative Assessment | MLP/LR/SCM/DT/RF/KNN | 3T-MRI imaging from 152 patients | AUC = 0.809; 0.746 sensitivity = 76.2%; 79.3% specificity = 74.1%; 72.2% (MLP; RF) | (70) |
| 2021 | Cancer response prediction | ML | MRI scanning from 72 patients | AUC = 0.793 | (71) | |
| 2021 | Image segmentation | U-Net | T2WI segmentation from 300 LARC patients | Mean DSC = 0.675 median DSC = 0.702 | (72) | |
| 2020 | RC Circumferential Evaluation | Faster R-CNN | detect 12,258 T2WIs | Accuracy = 0.932 sensitivity = 0.838 specificity = 0.956 | (73) | |
| 2022 | CRCLM early diagnosis | FM | CT scan from 30 patients | Precision = 100% overall accuracy = 93.3% recall = 77.8% | (74) | |
| 2021 | Tissue assessment | ResNet | OCT differentiate from 43, 968 cancer and 41, 639 norm ROIs | AUC = 0.975 | (75) | |
| 2021 | Diagnosis detection | DLLD | 4,386 CT images from 502 patients | Sensitivity = 81.82% false positives = 1.330 | (76) | |
| 2021 | Metastasis prediction | DLRS | Collect and predict from 235 nCRT patients | AUC = 0.894 | (77) | |
| 2019 | Accurate segmentation | LAGAN | CT scan and segment from 223 CRC patients | DSC = 90.82%; 91.54% (FCN32; U-Net) | (78) | |
| 2020 | Metastasis prediction | ResNet | CT scan from 192 CRLM patients | AUC = 0.903 | (79) | |
| 2022 | Cancer segmentation | U-Net/CNN | Analysis 201 MRI images | DSC = 0.727; 0.930; 0.917 (tumor; rectum; mesorectum) | (80) | |
| 2020 | Predict response | DL | T2W MRI predict 383 participants | AUC = 0.99 | (81) | |
| 2021 | Detect differentiation | RF | 169 CT images segmentation from 63 patients | AUC = 0.91 sensitivity = 82% specificity = 85% | (82) | |
| 2021 | Improve prognostication | RF | MRI identifies 94 lesions from 55 patients | AUC = 0.94 | (83) | |
| 2020 | Real-time diagnosis | DL | 26,000 OCT images | AUC = 0.998 | (84) |
ADR, adenoma detection rate; AUC, area under the curve; AMR, adenoma miss rate; AUPRC, area under the precision-recall curve; AUROC, area under the receiver operating characteristic; ANN, artificial neural network; CADe, computer-aided detection; CRCLM, colorectal cancer liver metastasis; CNN, convolutional neural network; CGANs, conditional generative adversarial networks; CRC, colorectal cancer; CP-ANN, counter propagation artificial neural network; DL, deep learning; DSI, dice similarity index; DWT, discrete wavelet transform; DELR, deep embedded-based logical regression; DLLD, deep learning based lesion detection algorithm; DLRS, deep leaning radiomic signature; DSC, dices similarity coefficient; DT, decision trees; FCN, fully conventional network; FM, Formal Methods; ICC, intra-class correlation coefficient; kNN, nearest neighbors; LR, logical regression; LARC, local advanced rectal cancer; LASSO, least absolute shrinkage and section operator; LAGAN, label assignment generative adverbial network; ML, machine learning; MSI, microsatellite instability; NB, naive baves; OCT, optical coherence tomography; PCA, principal component analysis; PDR, polyp detection rate; RNN, recurrent neural networks; RC, rectal cancer; RF, random forest; ResNet, residual network; SVM, support vector machine; TL, transfer leaning; TWI, T2-weighted images; U-net, U-shaped neutral network; WSI, whole slide images.