Explore anomaly detection and localization methods for medical imaging data

Abstract

Due to the limited availability of expert-labeled anomalous samples in medical imaging applications such as chest X-rays, most existing unsupervised anomaly detection methods rely on only normal imaging data for training. Recent studies in medical imaging literature have extensively explored generative adversarial network (GAN)-based approaches. These methods aim to reconstruct normal images during training and identify anomalies at test time based on their poor reconstructions. However, these reconstructions are not explicitly constrained to match the distribution of normal data. Consequently, anomalies may also be well-reconstructed, leading to degraded performance. Motivated by this, we introduce Contrastive Learning guided GAN (CLGAN), which leverages unlabeled mixed data and contrastive learning to guide the generation process. CLGAN regulates the generator by aligning generated images with the normal distribution while pushing them away from potential anomalies present in the unlabeled images. During training, the generator receives feedback regarding both the realness and normality of generated images, in the meantime, the discriminator learns better feature embeddings with the contrastive loss imposed. These contrastive features are subsequently utilized to indicate abnormality. Experimental evaluations on three public chest X-ray datasets demonstrate that CLGAN outperforms state-of-the-art (SOTA) anomaly detection methods significantly across various experimental settings. Ablation studies further confirm the effectiveness of each component of our proposed approach.