Contrastive self-supervised learning for diabetic retinopathy early detection

This paper first uses contrastive self-learning algorithm to pre-train the encoder based on convolution network with unlabeled retinal images, then retrain the encoder with classifier on a small annotated training data to detect referable DR.