Diabetic Retinopathy Severity Classification based on attention mechanism

Abstract

One of the significant factors causing blindness is diabetic retinopathy, a typical microvascular side effect of diabetes. Highly qualified professionals often examine colored fundus photos to identify this catastrophic condition. It takes much time and effort for ophthalmologists to diagnose diabetic retinopathy (DR) manually. The number of diabetes patients has dramatically increased during the last several years, which has made automated DR diagnosis a research hotspot. This paper proposes a hybrid deep learning model using a pre-trained DenseNet architecture integrated with CBAM for feature refinement. The dataset provided by the Kaggle Asia Pacific Tele-Ophthalmology Society (APTOS), having 3662 fundus images, is used in this research. In the multiclass classification experiment, we achieved 86.22% accuracy and 91.44 Kappa score(QWK). The local interpretable model-agnostic explanations (LIME) framework is used to assess predictions further and produce visual explanations, which can assist in decreasing the drawback of black-box models in aiding medical decision-making.