Active Learning for Efficient Annotation of Surgical Videos with Weak Supervision

Developing computer-assisted surgical systems is often bottlenecked by the high cost and expert time required for dense, pixel-level annotation of surgical videos. This paper investigates whether a human-in-the-loop framework, which combines weak supervision with active learning, can reduce this annotation burden while maintaining competitive performance in surgical tool segmentation.

The authors propose a framework that leverages a DINOv3 foundation model to generate class activation maps (CAMs) from surgical videos. The training process uses a dual-loss optimization strategy: a weak supervision loss based on video-level tool presence labels and a mask loss based on human-corrected annotations. Instead of requiring dense annotations upfront, the model iteratively proposes pseudo-masks. Human annotators then refine these masks, which are subsequently integrated into the training process as ground truth. This cycle repeats, allowing the model to learn from increasingly accurate data with minimal expert input.

The framework demonstrates a significant reduction in annotation effort, achieving a 50% decrease in time compared to fully manual annotation. Experimental results on the Cholec80 dataset show that the active learning approach consistently matches or exceeds the performance of control models trained with all ground truth masks provided at the start. The method is particularly effective for minority tool classes, where it achieves higher localization scores than traditional semi-supervised approaches.

By eliminating the need for large, fully annotated datasets at the start of model development, this framework provides a scalable and practical strategy for surgical tool segmentation. It offers a deployable path for expanding AI-driven surgical analysis to more diverse datasets and real-world clinical settings, where expert time is a scarce and expensive resource.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.