Self-supervised Anomaly Detection with Random-shape Pseudo-outliers.

Abstract

Anomaly detection in a medical image is a challenging yet essential task. It relies on learning patterns/distributions from health data only, and no abnormal samples are available during training. This study proposes a novel self-supervised learning method to precisely detect and localize anomalies in MRI medical images. We synthesize abnormal images by overlaying random pseudo-outliers onto normal samples and propose a discriminative model for anomaly detection. Unlike prior arts that generate abnormalities with pre-determined regular geometric shapes, we introduce a new outlier synthesis strategy capable of generating random-shape anomalies. By learning the disentanglement of pseudo-outliers and normal regions in the synthesized images, our model can capture natural anomalies in images at both the pixel level and sample level. We present our empirical experimentation on two publicly accessible datasets and demonstrate the proposed method's superiority over SOTA solutions on MRIs.