Table 2.
Studies using AI to diagnose urothelial cancer.
| Study | Application of the Study | Type of Study | Size of the Sample Used | Features Used for Training | Algorithms Used | Accuracy, % | Sensitivity, % | Specificity, % | AUC |
|---|---|---|---|---|---|---|---|---|---|
| Xu et al., 2017 [14] | Differentiate bladder tumor and bladder wall tissue by MRI | Retrospective | 62 patients (62 cancerous regions and 62 bladder wall regions) | Radiomic MRI characteristics: 2D texture characteristics and 3D texture characteristics |
SVM (2D) | 70.16–78.23 | - | - | 0.72–0.83 |
| SVM (3D) | 71.77–85.48 | - | - | 0.77–0.89 | |||||
| RF (2D) | 70.16–79.84 | - | - | 0.72–0.82 | |||||
| RF (3D) | 68.56–85.48 | - | - | 0.73–0.87 | |||||
| SVM (RFE-selected optimal features) | 87.9 | 90.3 | 85.5 | 0.90 | |||||
| Garapati et al., 2017 [15] | Forecast the stage of the disease based on CT urography | Retrospective | 76 CT urography cases (84 bladder cancer lesions: 43 < T2; 41 ≥ T2) | Pathological stage, CT urography morphological features, and textural features | LDA (training set) | - | - | - | 0.91 |
| LDA (testing set) | 0.88 | ||||||||
| SVM (training set) | 0.91 | ||||||||
| SVM (testing set) | 0.89 | ||||||||
| RF (training set) | 0.89 | ||||||||
| RF (testing set) | 0.97 | ||||||||
| NN (training set | 0.89 | ||||||||
| NN (testing set) | 0.92 | ||||||||
| Shao et al., 2017 [16] | Forecast whether the disease is present or not | Prospective | 87 bladder cancer patients and 65 patients without bladder cancer | 6 urine metabolite markers (spectral ions) | DT: testing | 76.6 | 71.9 | 86.7 | - |
| DT: training (5-fold cross validation) | 84.8 | 81.8 | 88.0 | - | |||||
| Ikeda et al., 2019 [17] | Detect tumors | Retrospective | 422 cystoscopic images | Transfer learning using features extracted from gastroscopic images | CNN | - | 96.5 | 96.5 | - |
Computed Tomography (CT); convolutional neural network (CNN).