计算机视觉人工智能模型在胃袖状切除联合胃底折叠及同期行食管裂孔疝修补术中应用研究OA

Objective To construct a computer vision artificial intelligence model for the integrated surgical scenario of laparoscopic gastric sleeve resection combined with fundoplication(LSGFD)plus concomitant hiatal hernia repair(HHR)based on the YOLOv11m deep learning framework,and to conduct independent verification,for providing visual module support for the future development of surgical navigation and robotic semi-automated surgical systems.Methods The surgical video data of patients who underwent LSGFD combined with HHR at the Department of Minimally Invasive,Hernia and Abdominal Wall Surgery,People's Hospital of Xinjiang Uygur Autonomous Region and the Department of Hepatobiliary Surgery,Hotan Regional People's Hospital were retrospectively analyzed.After screening and processing,a total of 3180 sample images were obtained and annotated via instance segmentation.Stratified random sampling method was adopted for dataset allocation:first,10％of the total samples were randomly selected as the independent test set;subsequently,20％of the remaining 90％samples were further randomly extracted as the validation set;the rest were all assigned as the training set.The weight files of the YOLOv11m deep learning model were loaded for transfer training and validation.Results After stratified random partitioning,the training set,validation set,and independent test set contained 16 239,4022,and 2196 targets,respectively.The model showed a continuous and stable downward trend in bounding box loss,segmentation loss,classification loss,and distribution focal loss;the loss curve of the validation set was consistent with that of the training set,suggesting no overfitting and good generalization ability of the model.On the independent test set,the model demonstrated excellent overall detection and segmentation performance:the mean average precision at IoU=0.5(mAP50)for both the bounding box(Box)and mask(Mask)was 0.908;the precision(P)and recall(R)for Box were 0.848 and 0.884,respectively,and for Mask were 0.846 and 0.881,respectively.Categories with a Box mAP50＞0.90 included gastric graspers,liver retractors,liver,stomach,gauze,ultrasonic scalpels,intestinal clamps,needle holders,dissecting forceps,staplers,spleen,and appliers.Categories with a Box mAP50 between 0.80 and 0.90 included biological clips,needles,folded flaps,and esophagus(with a Mask mAP50＞0.9).Categories with a Box mAP50＜0.80 were diaphragmatic crus and diaphragm(with a Mask mAP50＞0.8).Conclusion Computer vision artificial intelligence model technology can efficiently and accurately detect and segment the key anatomical structures and surgical instruments involved in the combined LSGFD and HHR procedure,thereby providing technical support for the subsequent expanded application of this surgical approach across multiple scenarios.