An Improved Residual U-Net for Segmentation of Multiple Structures in Fundus Images

Abstract

Diabetic Retinopathy screening helps with early detection and prompt treatment of this vision-threatening condition. To facilitate the screening procedure, deep learning-based segmentation method is designed to identify and segment the fundus image’s regular markers like the optic disc and blood vessels along with the DR lesion namely exudates. Based on a standard U-Net framework with minor changes to the encoder and decoder parts of the model, this study presents an improved residual U-Net for the segmentation process. Three U-paths are obtained by the IRU-Net, each of which is made up of three upsizing paths and one downsizing path. IRU-Net can improve the related feature fusion and acquire more information from fundus images with the use of a structure with three U-paths. Additionally, IRU-Net builds a residual block to retrieve highly realistic features, and it integrates a channel attention module alongside the decoder component to properly combine the feature data. Furthermore to address the imbalance in fundus image class, a revised weighted focus loss function is additionally included. To segment the image, identify the regions of the retinal image that are associated to blood vessels, and assess the suggested strategy for diagnosing retinal disease, the DRIVE and IDRiD image libraries are used here. Comparing IRU-Net to various classic approaches and other contemporary U-Nets, the numerical findings show that IRU-Net is a potential network for use in clinical imaging segmentation in terms of sensitivity, DSE, and IoU.