You can detect subtle distribution shifts in medical images by measuring how differently a diffusion model's prior and posterior distributions behave—no need for labeled anomaly examples or calibration data.
This paper introduces KLIP, a method for detecting when images deviate from expected distributions in medical imaging and other inverse problems. It uses diffusion models to spot both whole-image anomalies and localized abnormalities (like tumors in CT scans) without needing examples of the shifted distribution beforehand.