Reproduce “Enjoying Information Dividend: Gaze Track-based Medical Weakly Supervised Segmentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhisong Wang, Yiwen Ye, Ziyang Chen, Yong Xia
GradTrack, a framework developed at Northwestern Polytechnical University, introduces a novel method for weakly supervised medical image segmentation by uniquely leveraging the temporal order of physicians' gaze tracks. It achieves state-of-the-art performance among gaze-based methods, improving Dice scores by over 3% on Kvasir-SEG and NCI-ISBI datasets compared to previous approaches, and significantly narrows the performance gap with fully supervised models.
Key insights
- The temporal sequence and progression of a physician's gaze fixations are a rich, untapped source of supervisory information for segmentation, reflecting expert diagnostic reasoning more accurately than static attention maps.
- Employing a 'reverse truncation' strategy for gaze tracks allows for gradual supervision, where the model first learns from more conservative, reliable gaze information and progressively incorporates more complete data.
- Integrating learned gaze priors within the decoder path of the segmentation network, rather than the encoder, is more effective for feature refinement and overall performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.