Application of deep learning for surgical decision support during single-incision laparoscopic cholecystectomy

OBJECTIVE: This study aimed to develop an Artificial intelligence (AI) model based on Mask2Former to identify safe and hazard zones during Single-incision laparoscopic cholecystectomy (SILC). METHODS: 186 videos of SILC were included in the study, among which, 127 videos were used for training and the other 59 videos were used for testing the AI models. The model's performance was evaluated using Intersection over Union (IOU), Dice Score, F1 Score, Accuracy, Sensitivity, and Positive Predictive Value (PPV). RESULTS: A total of 186 videos with 9521 frames performed by two surgeons from two institutions were annotated simultaneously by an expert panel of 3 surgeons. The model achieved a high level of F1 score (0.966), IOU (0.805), Sensitivity (0.968) and PPV (0.942) for safe zones. For hazard zones, the model exhibited an exceptionally high Sensitivity (0.970), while the pixel-level overlap metrics for hazard zones (Dice = 0.825, IOU = 0.770) were lower than the object-level F1 score (0.923). CONCLUSIONS: The proposed model could be used to identify safe and hazard zones of dissection during SILC, with high sensitivity and PPV for safe zones and high sensitivity for hazard zones. This technology might eventually be used to provide real-time guidance intraoperatively to minimize the risk of major bile duct injury during SILC.