Abstract
Minimally invasive liver surgery (MILS) offers significant benefits but faces limited adoption due to its steep learning curve. This study explores the potential of artificial intelligence (AI) in assisting the performance of major MILS by providing intraoperative navigation through real-time segmentation of the safe plane for dissection. We developed and validated a deep learning model for segmenting vascular structures and the avascular plane during pure laparoscopic donor right hepatectomy (PLDRH). The study utilized 48 PLDRH videos from three institutions, with 40 videos used for five-fold cross-validation and 8 for external validation. The U-Net architecture with Mix Transformer encoder was employed for segmentation. Model performance was assessed using Dice similarity coefficient (DSC), precision, recall, and specificity. In internal validation, the model achieved mean DSC of 0.687 (SD 0.21) for vascular structures and 0.659 (SD 0.19) for the avascular plane. External validation showed comparable performance with DSC of 0.649 (SD 0.24) for vascular structures and 0.646 (SD 0.19) for the avascular plane. Visual assessment demonstrated accurate segmentation across different stages of right liver mobilization, despite lower quantitative metrics for vascular structures. This multicenter external validation study demonstrates the feasibility of AI-assisted intraoperative navigation for safe right liver mobilization in MILS. While promising, the study highlights the need for improved annotation strategies and further research to incorporate this technology into real operating theaters.