A Novel Multi-Scale Network based on Class Attention for Diabetes Retinopathy

Diabetes Retinopathy (DR) is a common eye disease, which brings irreversible blindness risk to patients in severe cases. Due to the scarcity of professional ophthalmologists, developing computer-aided diagnostic systems to participate in DR grading diagnosis has become increasingly important. However, the current mainstream deep learning methods still cannot accurately classify the severity of DR, and their unreliable results are difficult to serve as a reference for clinicians. To tackle this problem, we propose two novel modules to improve the accuracy of DR classification. Specifically, we designed a multi-scale feature extraction module (MFEM) to capture tiny lesions in fundus images and differentiate similar lesions simultaneously. In addition, we also created a class attention module (CAM) to alleviate the adverse impact of intra-class similarity on DR grading. Experiment on the APTOS2019 blind detection dataset show that our proposed two modules have made significant improvements to the designed model, achieving state-of-the-art performance with 95.98% for ACC and 97.12% for QWK, respectively.