Discovering Abnormal Patches and Transformations of Diabetics Retinopathy in Big Fundus Collections

Diabetic retinopathy (DR) is one of the retinal diseases due to long-term effect of diabetes. Early detection for diabetic retinopathy is crucial since timely treatment can prevent progressive loss of vision. The most common diagnosis technique of diabetic retinopathy is to screen abnormalities through retinal fundus images by clinicians. However, limited number of well-trained clinicians increase the possibilities of misdiagnosing. In this work, we propose a big-data-driven automatic computer-aided diagnosing (CAD) system for diabetic retinopathy severity regression based on transfer learning, which starts from a deep convolutional neural network pre-trained on generic images, and adapts it to large-scale DR datasets. From images in the training set, we also automatically segment the abnormal patches with an occlusion test, and model the transformations and deterioration process of DR. Our results can be widely used for fast diagnosis of DR, medical education and public-level healthcare propagation.