Table 2 Deep learning for computer-aided cystoscopy and ureteroscopy datasets: target disease, method, dataset and outcome summaries of selected comprehensive studies.
From: Where do we stand in AI for endoscopic image analysis? Deciphering gaps and future directions
Type proc. | Organ | Mod. | Target disease | Dataset | Method | Outcome | Similar studies |
|---|---|---|---|---|---|---|---|
Cyst. | Bladder | WL/BL | Tumour | Train: 95 patients 2335 frames (benign) 417 frames (cancer) Test: 54 patients | Detection67—tumour vs normal (CystoNet) | Sensitivity: 90.9%, specificity: 98.6% | Hashemi et al.123 (VGG16) Ikeda et al.66 (CNN) |
Cyst. | Bladder | BL | Tumour | Train: 10 patients, 196 frames Val: 10 frames Test: 10 frames (total: 216) | Classification68 Tumour vs normal (T1) Tumour invasiveness (T2) Grade classification (T3) (Ensemble) | (T1) sensitivity: 95.7%, specificity: 87.84% (T2) sensitivity: 88%, specificity: 96.56% (T3) sensitivity: 92.07%, specificity: 96.04% | NA |
Uter. | Ureter | WL | Stone | Train: 127 frames (2 per stone) leave-one-out | Classification69—Composition (ResNet101) | Sensitivity (mean): 83.34%, Specificity (mean): 96.5% | Lopez et al.70 (Inception) |
Uter. | Ureter | WL | Stone | Train: 92 frames Val: 32 frames Test: 30 frames (in vivo human) | Segmentation71—Stone and laser (MI-HybridResUNet) | Dice coeff.: 83.47% (stone) 86.58% (laser) | Zachary et al.124 (UNet) |
Other | Nasopharynx | WL | Tumour | Train: 19,576 frames Val: 2690 frames Test: 5270 frames | Accuracy (mean): 88.7% Dice coeff.: 78% (retrospective), 75% (prospective) | Xu et al.104 (Siamese) (WL/NBI) |
