Artificial intelligence-based expert trajectory guidance in an ex vivo robot-assisted renal wound suturing training model

Engineering notes

In the prospective training study, novice trainees who received expert trajectory guidance significantly outperformed the unguided control group across six of eight performance measures by the final assessment, and maintained their advantage during unguided evaluation sessions.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Introduction Renorrhaphy is one of the most technically demanding steps in robot-assisted partial nephrectomy, requiring expert-level suturing to ensure adequate hemostasis and long-term renal function preservation, yet acquiring such proficiency requires extensive practice under expert supervision. Although artificial intelligence has been increasingly applied to perioperative surgical care, its potential to learn expert operative patterns from standard surgical video and translate them into real-time visual guidance for surgical training remains largely unexplored. Here we developed and evaluated an artificial intelligence framework that learns expert suturing trajectories from standard endoscopic video and provides intraoperative visual guidance for renal wound suturing training. Methods A multicenter expert trajectory dataset was constructed from robot-assisted partial nephrectomy procedures performed at two medical centers, supported by a standardized data pipeline comprising 28-class scene annotation, temporal phase labeling of complete suturing actions, and sliding-window trajectory sampling. Building on these data, we developed a Scene-Aware Transformer that integrates instrument motion with surgical scene context to predict future trajectories, and prospectively evaluated the resulting guidance system in a pilot ex vivo porcine kidney feasibility training study involving 24 novice trainees. Results The dataset comprised 18,515 annotated frames, 806 complete suturing actions, and 24,897 valid trajectory samples. On the independent held-out institutional test set, the model achieved an average displacement error of 34.25 pixels and a final displacement error of 52.54 pixels, with an end-to-end inference latency of 32.7 ms under laboratory computational conditions. In the prospective training study, novice trainees who received expert trajectory guidance significantly outperformed the unguided control group across six of eight performance measures by the final assessment, and maintained their advantage during unguided evaluation sessions. Discussion These preliminary findings suggest that artificial intelligence-derived expert trajectory guidance may support short-term skill acquisition for renal wound suturing. As the prospective training component was a single-institution feasibility study without long-term retention or clinical transfer assessment, larger multicenter randomized trials are warranted before broader integration into surgical training curricula.

Links and sources

• PDF from original source

Comments