Semantic segmentation deep learning model boosts surgeons' organ recognition in minimally invasive hysterectomy - a...

Injury to the ureter and bladder during minimally invasive hysterectomy remains a serious complication, often resulting from insufficient intraoperative organ recognition. Artificial intelligence (AI)-based support systems may improve anatomical recognition and reduce such risks. We aimed to evaluate whether an AI-based anatomical recognition system for the ureter and bladder improves surgeons' organ recognition abilities across different experience levels without compromising specificity.

A deep learning model was trained on 13 934 ureter and 4940 bladder images sourced from 41 institutions. Model performance was evaluated using the Dice coefficient. The performance of 16 surgeons from eight facilities was evaluated for identifying the ureter and bladder in pre-recorded surgical videos with and without AI support. Sensitivity and specificity were compared between AI-assisted and non-assisted conditions, stratified by surgeon experience.

The AI model achieved a Dice coefficient of 0.66 for ureter segmentation and 0.62 for bladder segmentation. AI assistance significantly improved sensitivity for ureter detection (43.5-58.1%, P < 0.001) and bladder detection (54.2-70.0%, P < 0.001) without reducing specificity. The additive effect of AI was greater in less experienced surgeons, with sensitivity improvements of 27.3% for ureter and 26.8% for bladder recognition. Ureter recognition was stable across surgical phases, while bladder recognition demonstrated moderate variability.

The AI-based system enhanced surgeons' ability to recognize the ureter and bladder on surgical video, particularly among less experienced surgeons, without increasing the risk of misidentification. AI may help improve intraoperative anatomical identification and potentially contribute to safer, more standardized surgical performance.