Contra-Discriminative GAN-based Anomaly Detection Framework for Medical Images

Automated anomaly detection is of great interest for disease screening, image-based triage and quality control applications. As large collections of expert-labeled anomalous samples are typically unavailable, many studies have advanced label-efficient anomaly detection methods. These approaches often leverage generative adversarial networks (GANs) to reconstruct known normal images. However, as the generated outputs may not always follow the distribution of normal images, the ability to discriminate normal and anomalous samples could be limited. In this paper, we introduce a novel Contra-Discriminative GAN (CD-GAN) that utilizes unlabeled images to guide the generation process. Specifically, CD-GAN aligns the distribution of generated images with the distribution of normal images through a novel contrastive learning module that is seamlessly integrated into the GAN training process. In extensive experiments on four public and one real-world clinical medical image datasets, we demonstrate that CD-GAN significantly outperforms state-of-the-art anomaly detection methods across various experimental settings. As such, CD-GAN could offer competitive performance for diverse medical image anomaly detection applications.