Self-Supervised Metric Learning for Gaussian Anomaly Detection in Fundus Images

Anomaly detection in fundus images receives increasing interest in automated occular diseases screening. Instead of detecting a fixed range of diseases, based on provided labels during training, the global health status of a patient's retina is infered by focusing only on healthy fundus images, considered as samples from a normal class. Though promising, these methods need to be significantly improved before reaching clinical applications. On the one hand, multi-disease datasets providing fully healthy fundus images are rather small in size, due to the complexity of the labeling process. On the other hand, single-disease or few -disease datasets are large, but the scope of their normal class is limited as fundus images labeled as normal in these datasets might contain signs of other pathologies that were not taken into account during their constitution. Therefore, we propose an embedding similarity-based method performing a one-class classification, via a Mahalanobis distance, on the latent distribution of healthy fundus images in the embedding space of a feature extractor. Our method is particularly well-suited for the detection of anomalies in fundus images by leveraging both types of datasets. As the feature extractor is trained in a self supervised fashion on large datasets, using data augmentation specially designed for fundus images, the resulting low-dimensional embedding space reduces the number of fully healthy fundus images required for the one-class classification. We significantly improve upon the state of the art on JSIEC1000 and set a base-line for RFMiD and RFMiD2.0, the largest publicly available multi-disease datasets of fundus images.