Self-supervised 3D Out-of-Distribution Detection via Pseudoanomaly Generation

Abstract

Recent studies on anomaly detection have achieved great success in data analysis, yet the application of out-of-distribution detection in medical imaging remains an underdeveloped area of study. In this paper, we propose a 3D fully self-supervised learning method for volumetric medical image data. Inspired by recent advancements in representation learning for out-of-distribution detection, we propose a training method for pseudoanomaly generation with copy-paste. The training uses contrasts of the normal image with the pseudoanomaly image that is generated from the normal image. Through this scheme, a representation is learned to detect an abnormal image and to localize the anomaly area. In addition, we use a 3D patch as an input to provide the spatial information of the third dimension from volumetric image data. The proposed approach was tested in the 2021 MICCAI MOOD challenge, and it ranked the first place in both sample-level and pixel-level tasks.