Generalizability of surgical video analysis across unseen clinical environments

Establish robust algorithmic approaches for surgical phase recognition, surgical instrument instance segmentation, and surgical instrument keypoint estimation on laparoscopic cholecystectomy videos that generalize across unseen clinical environments and medical centers, maintaining comparable performance on hospitals with differing equipment, workflows, and data distributions.

Background

The paper evaluates three tasks—surgical phase recognition, instrument instance segmentation, and instrument keypoint estimation—on a multi-center dataset of laparoscopic cholecystectomy videos. Despite strong performances on videos from the hospital most represented in the training data, methods consistently underperformed on videos from the other two hospitals, indicating a domain shift across clinical environments.

This cross-hospital performance bias underscores a fundamental limitation in current approaches: the lack of generalizability across different clinical centers, equipment setups, and procedural styles. The authors explicitly note that this challenge remains unsolved, highlighting the need for methods that can sustain accuracy and robustness across diverse, previously unseen environments.