Diabetic Retinopathy Detection with Weighted Cross-Entropy Loss

Diabetic Retinopathy (DR) has received increasing attention due to the growing number of DR patients. Statistics show that the number of diabetic patients worldwide is about 425 million at present and is estimated to rise to 592 million in 2035. Computer-aided diagnosis techniques have also been used for DR grading. The small lesion features and the overall similarity of the pictures bring great challenges to computer-aided diagnosis. In the paper, we add Squeeze and Excitation (SE) blocks to the standard ResNet network, expecting that the network can pay more attention to smaller features. Meanwhile, we design a distance-weighted cross-entropy loss function to reduce the occurrence of large classification differences and improve the quadratic weighted kappa value. Finally, we validate our results on the Messidor datasets and achieve SOTA on the binary classification problem.