Reproduce “MedConcept: Unsupervised Concept Discovery for Interpretability in Medical VLMs” 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
Md Rakibul Haque, KM Arefeen Sultan, Tushar Kataria, Shireen Elhabian
MedConcept addresses interpretability in 3D medical vision-language models by splitting frozen Merlin embeddings into sparse autoencoder features, assigning medical-language labels from an abdomen vocabulary, and checking them against radiology reports with MedGemma. At top K = 25, MerlinPlus has median Aligned scores around 0.30–0.35, while AbdomenAtlas 3.0 is below 0.1; these are report-consistency measures, not clinical accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.