Fairness in Unsupervised Visual Anomaly Detection: Evaluating Demographic Bias Across Medical Imaging Domains

This thesis investigates the fairness of unsupervised anomaly detection models in medical imaging. We evaluate three architecturally distinct methods — PatchCore, PaDiM, and STFPM — on two clinical datasets: NIH ChestX-ray14 (chest radiography) and Fitzpatrick17k (dermatology). We measure demographic fairness gaps under two real-world bias sources: training set underrepresentation and label noise disproportionately affecting minority groups. Results show all methods develop measurable fairness gaps under demographic skew, with magnitude varying by model architecture and backbone choice — demonstrating that fairness is jointly determined by data, feature extractor, and detection framework.