Deep learning-assisted near-real-time identification of the inferior mesenteric artery and vein during laparoscopic rectal resection: a single-center proof-of-concept study
TL;DR
A near-real-time semantic segmentation model using standard white-light laparoscopic video to support intraoperative anatomical recognition and achieved high segmentation scores on selected internal test data and supported near-real-time visualization.
Abstract
Background Reliable identification of the inferior mesenteric artery (IMA) and inferior mesenteric vein (IMV) is essential for safe vascular control during laparoscopic rectal resection (LRR). We developed and internally evaluated a near-real-time semantic segmentation model using standard white-light laparoscopic video to support intraoperative anatomical recognition. Methods This single-center retrospective proof-of-concept study included operative videos from 16 patients who underwent LRR between February 2024 and May 2025. Frames were sampled at 1 frame/s, yielding 2,720 expert-annotated RGB images (1,758 IMA images and 962 IMV images). Data were divided at the patient level into a training/development cohort (13 patients, 2,260 frames) and an internal holdout test cohort (3 patients, 460 frames). A frame-wise two-dimensional U-Net was configured using nnU-Net v2. Multiclass segmentation was compared with vessel-specific binary segmentation. Twenty gastrointestinal surgeons who had not participated in model development assessed preliminary usability and acceptance using three 0–4 Likert items. Results Vessel-specific binary segmentation produced more balanced performance, with the clearest improvement for the IMV. In the internal holdout test cohort, the IMA Dice coefficient, precision, and recall were 0.940 +/− 0.025, 0.945 +/− 0.023, and 0.940 +/− 0.020, respectively; the corresponding IMV values were 0.980 +/− 0.010, 0.982 +/− 0.017, and 0.978 +/− 0.018. On an NVIDIA RTX 3090 GPU, inference reached 12.7 frames/s, with approximately 0.08 s of network processing per frame. Surgeons rated perceived recognition accuracy at 3.41 +/− 0.09 and future clinical potential at 3.39 +/− 0.09 on the 0–4 scale. Conclusion The nnU-Net-configured two-dimensional U-Net achieved high segmentation scores on selected internal test data and supported near-real-time visualization. These findings provide early evidence of technical feasibility rather than clinical effectiveness. Multicenter external validation, testing in difficult operative fields, objective human-factors experiments, and prospective clinical evaluation are required before clinical use.