Modified U-Net Architecture Using Fundus Images for Hemorrhages Semantic Segmentation

Abstract

Diabetic retinopathy is a severe eye disease that affects diabetic patients worldwide due to fluctuations in blood glucose levels. Early detection of DR is crucial in preventing blindness, and CADS have been instrumental in aiding ophthalmologists in the diagnosis phase of the disease using fundus images. The researches focuses on detection the type and location of abnormalities in fundus images efficiently. This paper introduces a modified UN et architecture model trained to segment hemorrhages using fundus images. The first step involves pre-processing images and masks. Followed by training the model using IdRID dataset. The performance of the modified UNet model is compared to other methods using DRIVE, STARE, CHASE-DBl, and IOSTAR datasets, and it outperforms them in accuracy. The application of data augmentation techniques has significantly enhanced the model's performance, making it robust and effective even on images from different datasets. On DRIVE and STARE datasets, the modified UNet achieved Se and Acc values of 0.8293, 0.9675, 0.8973, and 0.9833, respec-tively, showing its superiority in small vessel segmentation and hemorrhage segmentation. Additionally, cross-dataset and cross-modality experiments were conducted on the DRIVE, STARE, and IOSTAR datasets to evaluate the architecture's generalization and extendibility capabilities, and the results demonstrate that it has remarkable abilities in both areas.