Normative Prior Network for Anomaly Segmentation in Retinal OCT Images

Abstract

Anomaly segmentation refers to leveraging only normal images for model training to detect pixel-level anomalies in abnormal images in test phase, which gets rid of the reliance on expensive manual labels and is able to detect rare anomalies that are hard to collect in training dataset. However, previous works performing anomaly segmentation in retinal OCT images heavily relied on complex post-processing methods and cannot detect anomalies in an end-to-end manner. To overcome that, in this paper, a novel anomaly segmentation method is proposed to segment retinopathy based on normative prior network. Specifically, a segmentation network and a variational autoencoder is constructed to extract retinal regions and learn normative prior of normal retinal structure, respectively. In inference phase, the anomalous region can be directly segmented by low structural similarity between abnormal retinal structure and the normal one reconstructed by variational autoencoder. Extensive experiments are conducted to evaluate the performance of the proposed method and the experimental results show that it achieves the state-of-the-art (SOTA) performance for anomaly segmentation in retinal OCT images.