We discovered that performance bias between demographic subgroups in unsupervised anomaly detection follows certain patterns, making it highly predictable.

So far the largest comparison of anomaly detection models with various paradigms on four medical imaging datasets.

RoDeO is an easy to use object detection metric. It evaluates three sources of errors (misclassification, faulty localization, and shape mismatch) separately and combines them to one score.

All current anomaly detection models are outperformed via simple thresholding on the commonly used modality of FLAIR MRI.