Multitask Learning for Joint Grading of Diabetic Retinopathy and Macular Edema Using Transformer

Abstract

AbstractDiabetic Retinopathy (DR) and Diabetic Macular Edema (DME) are the two major reasons of vision loss. These diseases occur due to high levels of sugar in the blood, which leads to damage to the retina. Recently, some computer-aided detection methods for DR and DME have been proposed to help ophthalmologists with early detection and better treatment for patients. However, earlier research either graded DR or DME neglected the relation between DR and DME. In this paper, two modules have been proposed. The individual disease grading module for learning specific features of each disease. Moreover, the multitask learning module for joint grading DR and DME, to inspect the intrinsic relation between DR and DME, generates disease reliant features and enhances the overall accuracy of DR and DME grading. The proposed method’s efficiency was evaluated using the Messidor-1 benchmark dataset. The evaluation revealed that the model’s accuracy exceeded that of the state-of-the-art model by >10%.KeywordsDiabetic Macular EdemaDiabetic RetinopathyJoint gradingTransformerAttentionDeep learning