Reproduce “Beyond Classification: Structured Supervision Aligns Visual Evidence with Medical Semantics” 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

Hexiang Bai, Hanyang Xu, Xiaoxue Li, Xiaoliang Wu, Shangde Gao, Hongxia Xu, Ke Liu

The authors examine whether structured pretraining aligns a model’s strongest local image evidence with lesions. They compare graph self-supervision, segmentation co-training, and image-text alignment using class-specific prompts derived from labels, measuring Patch In Box (PIB). Fine-tuned BiomedCLIP averages 0.410 PIB versus 0.168 for standard ViT-B/16 across three datasets, but improvements are not uniform and do not establish clinical reasoning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.