Reproduce “Auditable CT Phenotyping Through Report-derived Radiological Observations” 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
Riga Wu, Walter Witschey, Yicheng Li, Felix Barajas Ordonez, Keno K. Bressem, Lisa C. Adams, Gary E. Weissman, Li Shen and 4 more
The Auditable CT phenotyping (ACT) framework integrates a native 3D volume-report model with a comprehensive observation bank derived from radiology reports to enhance the transparency and clinical utility of medical AI models. This system enables precise image annotation, robust EHR phenotyping, and a mechanism to audit and restrict models from relying on non-disease-specific shortcut observations, preserving accuracy while improving interpretability.
Key insights
- A shared, report-derived radiological observation vocabulary can effectively link and unify medical image annotation, EHR phenotyping, model auditing, and intervention within a single framework.
- Medical AI models trained for EHR phenotyping frequently learn and rely on non-specific 'shortcut' observations (e.g., vascular calcification) that correlate with multiple clinically unrelated conditions.
- Systematically auditing and clinically restricting the model's observation space can successfully redirect its focus towards relevant radiological evidence, maintaining or even improving classification accuracy while enhancing clinical interpretability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.