New AI framework cuts surgical video annotation effort by 50%

A team of researchers led by Manasa Dendukuri has introduced a novel active learning framework for efficient annotation of surgical videos, accepted at IPCAI 2026. The method addresses the bottleneck of precise spatial-temporal annotation of laparoscopic videos, which is both time-consuming and requires expert knowledge. The framework combines weak supervision—using video-level tool presence labels—with human-in-the-loop corrections on pseudo-masks proposed by the model. A dual-loss optimization trains a foundation model to generate temporally consistent class activation maps (CAMs), leveraging both weak and strong supervision signals.

By iteratively presenting the expert with pseudo-masks rather than requiring dense pixel-level annotations upfront, the system reduces overall annotation effort by 50% by the end of training compared to fully manual annotation. This approach eliminates the need for large fully annotated datasets from the start, making it practical for scaling surgical tool segmentation models to larger, more diverse datasets and real-world clinical settings. The framework's iterative refinement process ensures efficient knowledge acquisition with minimal expert input, offering a deployable strategy for medical AI development.