Unsupervised Abnormality Detection with Normalizing Flow and Neural Network

In anomaly detection, the limited number of abnormal samples results in a significant long-tail distribution in the data. Additionally, labeling is heavily dependent on the experience and can be time-consuming. This problem is particularly prominent in the field of abnormality detection from medical images. To this end, we propose a novel end-to-end unsupervised method for detecting abnormalities using a flow model-based neural network. Specifically, This approach utilizes deep neural networks to obtain the features of normal images and model the distribution with the statistical method. It maps image features to tractable distributions and effectively integrates both local and global relationships to identify distinctive features. Comprehensive experiments on the MURA dataset demonstrate that our method outperforms recent methods by a considerable margin.