| Kitaguchi et al., 2020 [51] | Japan | 19 intraoperative videos | 9 stages and 3 actions, 5 tools | CNN | The overall accuracy of the full-phase classification model was 81.0%, and the mean intersection over union (mIoU) for tools T1 - T5 was 51.2%. | The CNN model can be used to build an automated video analysis and indexing system. |
Ichimasa et al., 2022 [52] | Japan | 511 patients with early T2 colorectal cancer (CRC) | 8 clinical and pathological variables (age, sex, tumor size, tumor location, lymphatic invasion, vascular invasion, histological differentiation, and CEA level) | RF algorithm | The AUC of the AI system in validation data was 0.93, while the AUC of the nomogram (based on multivariable logistic regression analysis) was 0.88. | The AI prediction model has the potential to reduce unnecessary surgeries for T2 CRC patients and enable minimally invasive treatment through endoscopic full-thickness resection (EFTR) for T2 CRC patients with lymph node metastasis. |