Classification for diabetic retinopathy by using staged convolutional neural network

Diabetic retinopathy (DR) is the leading cause of permanent blindness in the working-age population, which is one of the common complications of diabetes. DR grading is crucial in determining the relevant treatment to reduce vision loss. Automatic grading approaches of DR are very significant for helping ophthalmologists design adequate treatment to patients. However, DR grading is challenging due to the facts of intra-class variations and inter-class similarities. The key point of solving DR grading is to find abundant discriminative lesions corresponding to subtle visual differences, such as microaneurysms, soft exudates and hemorrhages. To solve the problem, we proposed a two-stage classification process to firstly classify the presence or absence of DR based on the characteristics of fundus images of DR patients. Then for fundus images with DR, we proposed a novel lesion attention module to perceive and capture lesion features for fine-grained classification. Comprehensive experiments are conducted on DDR dataset to evaluate the effectiveness of the proposed DR grading method. Our method achieves the state-of-the-art results on DDR dataset.