Abstract

Info

Title:: Joint grading of diabetic retinopathy and macular edema based on improved ResNeSt50

Author(s):: LIU Yuxuan; GU De; School of Internet of Things Engineering, Jiangnan University, Wuxi 214122, China

Keywords:: diabetes; retina; ResNeSt50; partial convolution; attention mechanism

Abstract:: To address the challenge of joint grading caused by the diverse lesion morphologies associated with differentstages of diabetic retinopathy, such as hemorrhages, microaneurysms, and neovascularization, which often obscure thelesions of diabetic macular edema, an improved ResNeSt50-based joint grading network is proposed. The modifiedResNeSt50 with a novel convolutional operation (partial convolution) replacing the standard 3×3 convolution in the residualblocks is used to extract image features for exploring the specificities of diabetic retinopathy and macular edema.Subsequently, a disease-related attention module is introduced to capture the intrinsic correlation between diabeticretinopathy and macular edema. Experiments conducted on the Messidor and IDRID datasets show that the proposedapproach achieves joint accuracies of 83.7% and 64.1%, respectively. The results demonstrate that the proposed algorithmcan significantly improve the performance of joint grading for diabetic retinopathy and diabetic macular edema.