SUPRDAD: A Robust Feature Extractor Better Recognizes Low-Prevalent Retinal Diseases

Leveraging the large-scale data-driven deep learning, many researchers have attempted to automatically screen retinal abnormalities from fundus images. Most of them focus on a single disease that allows easy accessibility to clinical cases. In real clinical environment, however, patients can suffer from various retinal diseases with low prevalence and co-occurrence. Data distribution shift makes classification task more difficult. To tackle these issues, we propose a novel framework that boosts representation learning of the feature extractor using additional unlabeled fundus images, from which we can benefit effective fine-tuning for disease-specific classifiers. The feature extractor is trained based on the extended single label supervision, and generalized to unseen features via self-supervised semi-supervised learning under multi-task training scheme. Then we adapt the feature representation to a more robust space using domain-adaptive distillation. Experiments are conducted on three carefully prepared test datasets; in all metrics, every fine-tuned classifier out of five diseases demonstrates superior performance to the corresponding one-versus-rest supervised learning baseline and, in particular, by 10.4 percent in AUPR. The proposed method largely improves classification performance for low-prevalent retinal diseases and can be potentially extended to other diseases.