Dealing with Long-tail Issue in Diabetic Retinopathy and Diabetic Macular Edema Grading

Diabetic Retinopathy (DR) is a multiply occurring complication induced by prolonged course of diabetes. Diabetic Macular Edema (DME) is the most common complication of DR which is the major threat of vision loss. Hence, it is urgently needed to expand the early screening and diagnosis via computer-assisted therapy. However, prior works mainly focus on investigating DR and DME in isolation, largely ignoring their inherent relationships. Besides, the fundus data distribution is typically long-tailed, with tail classes concentrating on critical levels of DME. Motivated by the distinctive complexion above, this work presents a novel position-guided attention block (PGAB) as well as an innovative label-sensitive (LS) loss, which are respectively in charge of extracting position-sensitive features to exploit interactions between hard exudate and macular and encouraging the model to embrace tail classes to lift the accuracy on critical levels of DME. Comprehensive experiments on popular Messidor and IDRiD datasets well demonstrate the superiority of our approach in achieving competitive performance compared to state-of-the-arts.