HernIA: Real-Time Anatomical Structure Segmentation in Video Laparoscopic Inguinal Hernioplasties With AI

Abstract

Background

Laparoscopic transabdominal preperitoneal (TAPP) hernioplasty, a minimally invasive procedure, reduces postoperative pain and recovery time but faces challenges like the “ping-pong effect” (alternating focus between operative field and monitors) and a 1%–2% error rate due to anatomical misidentification, risking complications like vascular injuries.

Objective

To develop and validate HernIA, an AI-based system for real-time segmentation of anatomical structures in TAPP, targeting an Intersection over Union (IoU) ≥85% and error reduction ≥50% compared to manual identification.

Methods

HernIA employs YOLOv11m-seg, trained on 21 443 annotated laparoscopic images from 45 TAPP procedures at Clinica Colón and Hospital de Campaña Escuela Hogar. Annotation by expert laparoscopists achieved high inter-rater reliability (Cohen’s kappa = 0.87). Validation used 5-fold cross-validation and a 10 800-frame dataset.

Results

HernIA achieved an IoU of 89.4% (±2.1%), Jaccard Index of 81.2%, mAP@50 of 92.3%, and F1 score of 0.94 (confidence threshold ∼0.45). It reduced identification errors by 62% in a simulated TAPP environment (10 800 frames, 24 FPS, 42 ms latency). Clinical validation was limited to one case of bilateral hernia repair.

Conclusion

HernIA enhances surgical precision and training in TAPP, with potential to reduce complications. Multi-center trials are needed to confirm generalizability.