Deep learning-based automatic identification model of main pancreatic duct in intraoperative ultrasonography.
Authors
Affiliations (10)
Affiliations (10)
- Faculty of Hepato-Biliary-Pancreatic Surgery, The First Medical Center of Chinese People's Liberation Army (PLA) General Hospital, 28 Fuxing Road, Beijing, 100853, P.R. China.
- Department of Anesthesia, First Medical Centre of Chinese, PLA General Hospital, Beijing, 100853, P.R. China.
- National Clinical Research Center for Geriatric Diseases Chinese, PLA General Hospital, Beijing, 100853, P.R. China.
- Nankai University School of Medicine, Tianjin, 300071, P.R. China.
- Department of Ultrasound, Second Medical Center of Chinese, PLA General Hospital, Beijing, 100853, P.R. China.
- Department of Organ Transplant, Third Medical Center of Chinese, PLA General Hospital, Beijing, 100853, P.R. China.
- School of Computer and Communication Engineering, University of Science and Technology Beijing, Beijing, 100853, P.R. China.
- China Academy of Engineering Physics, Beijing, 100853, P.R. China.
- Faculty of Hepato-Biliary-Pancreatic Surgery, The First Medical Center of Chinese People's Liberation Army (PLA) General Hospital, 28 Fuxing Road, Beijing, 100853, P.R. China. [email protected].
- Nankai University School of Medicine, Tianjin, 300071, P.R. China. [email protected].
Abstract
Pancreatic enucleation (PE) preserves parenchyma in benign/low-grade malignant pancreatic tumors but risks postoperative pancreatic fistula (POPF) if the main pancreatic duct (MPD) is injured. Intraoperative ultrasound (IOUS) guides MPD localization, yet its efficacy is limited by operator dependency and expertise gaps. This study developed a deep learning (DL) model for automated MPD identification during IOUS to enhance surgical precision and reduce POPF risk. A retrospective dataset included IOUS images/videos from 166 patients undergoing minimally invasive PE (2018-2023). Images were manually annotated with MPD bounding boxes by experienced surgeons, categorized as positive (MPD present) or negative. Data augmentation balanced positive-to-negative ratios (1:1), yielding training (44,702 images), internal validation (447 images), and test sets. External validation used 408 images and 16 video clips, including 10 videos from the other institution. The Fast R-CNN architecture, pretrained on the COCO dataset, was finetuned using stochastic gradient descent optimization. Performance was evaluated via recall rate and accuracy (Intersection over Union threshold: 50%), assessed by video annotation speed (frames per second, fps). The model achieved high recall rates: 98.7% (internal), 97.8% (external images), 96.5% (video validation) and 94.8% (additional external video validation), with accuracies of 93.9, 84.2, 83.1, and 82.0%, respectively. Annotation averaged 3.8 fps, enabling seamless integration into surgical workflows. False positives were predominant (73.3% with confidence > 0.8 in internal validation), often involving multiple detections per frame (80%). This DL-based IOUS system enables accurate MPD identification, mitigating operator variability and supporting surgical training. While false positives require refinement, the model enhances safety in PE by preventing MPD injury. Future integration into clinical ultrasound platforms and optimization for postoperative imaging contexts are warranted.