Objective:To explore the clinical application value of artificial intelligence models in identifying bleeding events and hemorrhagic points during laparoscopic pancreatic surgery.Methods:This single-center retrospective cohort study collected surgical videos of 25 patients undergoing laparoscopic pancreatic surgery at the Department of General Surgery, Peking Union Medical College Hospital from January 2022 to December 2024. Videos within 5 seconds before and after representative bleeding events were captured at 30 frames/s, with 11 666 hemorrhagic-related video frames annotated. Two algorithm models were developed: a pigment-based model and a pigment+optical flow-based model for classification and target recognition of bleeding frames. The training and test sets for the pigment-based algorithm contained 4 692 hemorrhagic and 4 309 non-hemorrhagic frames, while those for the pigment+optical flow model included 1 339 hemorrhagic and 1 326 non-hemorrhagic frames. Performance evaluation was conducted using overlap thresholds, with accuracy and recall rates as key metrics.Results:The pigment-based model achieved 93.8% accuracy (134/143) and 43.3% recall (134/310) in hemorrhagic frame classification. At an overlap threshold of 0.3, the pigment-based model showed 84.1% accuracy (433/515) and 85.4% recall (433/507) in target recognition. When the threshold was increased to 0.5, the pigment+optical flow model demonstrated 88.1% accuracy (354/402) and 89.2% recall (354/397) in hemorrhagic target recognition.Conclusions:It is difficult to distinguish active bleeding from old bleeding completely by pigment information alone. The spatio-temporal features can be effectively extracted by combining pigment and optical flow information, and the bleeding can be accurately identified and located, which has potential clinical application value.

Objective:To explore the clinical application value of artificial intelligence models in identifying bleeding events and hemorrhagic points during laparoscopic pancreatic surgery.Methods:This single-center retrospective cohort study collected surgical videos of 25 patients undergoing laparoscopic pancreatic surgery at the Department of General Surgery, Peking Union Medical College Hospital from January 2022 to December 2024. Videos within 5 seconds before and after representative bleeding events were captured at 30 frames/s, with 11 666 hemorrhagic-related video frames annotated. Two algorithm models were developed: a pigment-based model and a pigment+optical flow-based model for classification and target recognition of bleeding frames. The training and test sets for the pigment-based algorithm contained 4 692 hemorrhagic and 4 309 non-hemorrhagic frames, while those for the pigment+optical flow model included 1 339 hemorrhagic and 1 326 non-hemorrhagic frames. Performance evaluation was conducted using overlap thresholds, with accuracy and recall rates as key metrics.Results:The pigment-based model achieved 93.8% accuracy (134/143) and 43.3% recall (134/310) in hemorrhagic frame classification. At an overlap threshold of 0.3, the pigment-based model showed 84.1% accuracy (433/515) and 85.4% recall (433/507) in target recognition. When the threshold was increased to 0.5, the pigment+optical flow model demonstrated 88.1% accuracy (354/402) and 89.2% recall (354/397) in hemorrhagic target recognition.Conclusions:It is difficult to distinguish active bleeding from old bleeding completely by pigment information alone. The spatio-temporal features can be effectively extracted by combining pigment and optical flow information, and the bleeding can be accurately identified and located, which has potential clinical application value.