Related Experiment Videos
A novel artificial intelligence anatomical model for landmark recognition in endolaparoscopic totally extraperitoneal
YuFu Jarrell Lai1, Rou Yi Soong1, Le'En Maegan Sim1
1Yong Loo Lin School of Medicine, National University of Singapore, 10 Medical Dr, Singapore, 117597, Singapore.
Background:
Totally extraperitoneal (TEP) inguinal hernia repair demands precise recognition of posterior groin anatomy, yet the learning curve remains steep. We developed and evaluated an artificial intelligence (AI) model to identify key landmarks of the myopectineal orifice during TEP repair to support training and intraoperative guidance.
Methods:
Thirty high-definition videos of unilateral elective TEP inguinal hernia repair were reviewed. From these, 1,946 image frames were extracted and annotated for six critical landmarks: pubic bone, inferior epigastric vessels, testicular vessels, vas deferens, triangle of pain, and triangle of doom. After cleaning, 1,420 frames remained (training/validation/evaluation split 70/15/15). Training images underwent augmentation (flips, rotations, colour adjustments) and oversampling to address class imbalance, yielding 14,961 training images from an initial 994. A YOLO11-M-seg object detection model was trained for 100 epochs and assessed using COCO metrics (precision, recall, mAP50, mAP50-95).
Results:
The final dataset comprised 15,387 images. The model achieved precision 0.805, recall 0.704, mAP50 0.774, and mAP50-95 0.500. Landmark performance excelled for triangle of pain (AP50 0.847; AP50-95 0.502) and testicular vessels (AP50 0.853; AP50-95 0.560), with strong results for the other landmarks. Inference speed reached 60 FPS, supporting real-time application.
Discussion:
High accuracies for distinct landmarks like triangle of pain and testicular vessels reflect robust feature learning despite challenges like occlusion and lighting variability. Augmentation simulated intraoperative conditions (angle changes, smoke, blood), enhancing generalisability. Real-time speed exceeds video capture rates, enabling AR overlays for navigation.
Conclusion:
This AI model shows promise for real-time TEP landmark recognition, potentially accelerating training and reducing complications in endo-laparoscopic hernia repair.