The role of noise in denoising models for anomaly detection in medical images

Antanas Kascenas, Pedro Sanchez, Patrick Schrempf, Chaoyang Wang, William Clackett, Shadia Mikhael, Jeremy P. Voisey, Keith A. Goatman, Alexander Weir, Nicolas Pugeault, Sotirios A. Tsaftaris, Alison Q. O'Neil

Abstract: Highlights•Denoising is an effective training task for anomaly detection in medical images.•Training denoising models with coarse noise improves anomaly detection.•Noise parameters generalize between models, image modalities and anomaly types.•A simple DAE trained with coarse noise gives state-of-the-art results.