Artificial intelligence system for anatomical landmark detection in rectovaginal fistula repair surgery: technical development and educational evaluation.

purposeRectovaginal fistula from perineal body injury during delivery is rare, with limited case volumes preventing standardized protocols. We developed an AI system for anatomical landmark detection and evaluated its educational potential during surgery.

methodsThis pilot feasibility study developed an AI system using HyperSeg semantic segmentation to detect anatomical landmarks (perineal body, vaginal wall, rectum) during rectovaginal fistula repair. Training data comprised 2000 annotated images from 10 cases; 100 images from 5 independent cases served as the test dataset. Dense sampling, reverse-chronological annotation, and multi-structure detection were employed to maximize performance with limited data. Surgical outcomes and intraoperative communication were compared between a pre-AI period and an AI evaluation period (5 cases each; 25 consecutive cases total). The trainee assistant-with no prior experience in this procedure-referenced the AI display while the supervisor guided without viewing it.

resultsThe AI model achieved Dice coefficients of 0.655 ± 0.185 (perineal body), 0.672 ± 0.167 (vaginal wall), and 0.705 ± 0.144 (rectum). The rectum met the predefined threshold of 0.7; the vaginal wall and perineal body approached but did not reach this threshold, reflecting the inherently ambiguous boundaries of these thin membranous structures. Temporal comparison between periods showed: 221% difference in dissection time efficiency (p = 0.008), 73% reduction in blood loss (p = 0.056), and 185% difference in repair time efficiency (p = 0.056). Despite shorter operative times, intraoperative communication increased substantially: Supervisor instructions increased 71% (dissection phase, p = 0.008) and 68% (repair phase, p = 0.012); assistant questions increased 67% and 100%, respectively (all p ≤ 0.012, Cliff's δ = 1.000).

conclusionThis pilot feasibility study demonstrated technical feasibility of AI landmark detection in rectovaginal fistula repair using only 10 training cases, with the rectum achieving the predefined Dice threshold. A temporal increase in intraoperative communication was observed following AI introduction, suggesting potential facilitation of supervisor-trainee interactions. Future controlled studies are required to validate these findings.