Auditable CT Phenotyping Through Report-derived Radiological Observations

TL;DR

This paper introduces Auditable CT phenotyping, a method that predicts clinical phenotypes from computed tomography scans using report-derived radiological observations. The approach outperforms vision-language baselines and provides insights into the observations used for prediction. It achieves high accuracy while relying on a small set of observations.

Why It Matters

The method has practical significance as it enables accurate and transparent CT-based phenotyping, which can improve disease diagnosis and treatment. It also highlights the potential for AI models to rely on shortcuts rather than disease-specific findings, which can impact their robustness and generalizability. This work can lead to more reliable and trustworthy AI-assisted medical diagnosis.

Key Findings

• Auditable CT phenotyping outperforms vision-language baselines
• A small set of observations occupies the top rank-1 positions
• Restricting the observation bank to clinician-specified evidence improves probe accuracy