Ai Applications In Segmenting Retinal Lesions And Detecting Macula And Optic Disc Areas In Fundus Images

Abstract

A crucial step in the diagnosis and treatment of ocular disorders such as diabetic retinopathy, uveitis, cataracts, macular degeneration, and others is the analysis of lesions that occur in fundus pictures of the retina and the precise determination of the optic disc and macular. However, a lot of current research still heavily depends on the manual resolution work done by subject-matter specialists. Their work is frequently tedious, repetitive, and time-consuming, which leads to a decrease in accuracy and efficiency in evaluating the outcome. It has become more necessary to create an automated computer-aided system to help ophthalmologists provide recommendations that are more accurate and trustworthy. Although there have been significant advances in the deep learning field, particularly with neural networks and other approaches, these current methods still have several issues, such as poor performance, missing minor lesions, etc. To tackle this problem, this study applied deep learning to two tasks, including segmenting retinal lesions and detecting the macula and optic disc. Both of these assist in identifying the macula and the optic disc, two significant eye structures, and the regions of damage in the fundus images aid in identifying retinopathy-related symptoms. The model is constructed using both the Faster R-CNN and Mask R-CNN with backbone combining ResNet 50 or ResNet 101 with Regional Request Network (RPN). In comparison to the IDRiD dataset with the macula and optic disc detection, which only obtained 66.4% mAP on the test dataset, the Faster R-CNN model achieved 93% mAP@50-95 in the DRIVE dataset. Additionally, the Mask R-CNN obtained results in the IDRiD dataset with the segmentation of retinal lesions is 22.56% mAP. In general, normal human eye pictures performed better than in diabetic retinopathy patients in recognizing the macula and optic disc in fundus images. Regarding the segmentation of the lesion, the big plaque regions on the fundus imaging cannot be segmented the most possible.