A Systematic Review on Diabetic Retinopathy Detection Using Deep Learning Techniques

Segmentation is an essential requirement to accurately access diabetic retinopathy (DR) and it becomes extremely time-consuming and challenging to detect manually. As a result, an automatic retinal fundus image segmentation (RFIS) system is required to precisely define the region of interest and help ophthalmologists in the rapid diagnosis of DR. This systematic review provides a comprehensive overview of the development of deep learning (DL) based approach for RFIS to diagnose DR at an early stage. This review is fivefold: (1) retinal datasets, (2) pre-processing approaches, (3) DR segmentation and detection methods, (4) performance evaluation measures, and (5) proposed methodology. Articles on RFIS for DR detection were identified using the query “Deep Learning Techniques”, “Diabetic Retinopathy”, and “RFIS”, alone and in combination using PubMed, Google Scholar, IEEE Xplore, and Research Gate databases until 2021 using PRISMA principle. Approximately 340 publications were searched and 115 relevant studies focused on the DL approaches for RFIS for DR diagnosis were chosen for study. According to the survey, 66% of researchers employed DL approaches for Blood vessel (BV) segmentation, 36% of researchers used DL approaches for lesion detection, and 15% of researchers used DL approaches for optic disc and optic cup (OD & OC) segmentation for DR Diagnosis. This systematic review provides detailed literature of the state-of-the-art relevant articles for RFIS of BV, Lesions, OD & OC for non-proliferative DR diagnosis and discusses future directions to improve the performance of DR and overcome research challenges. Finally, this article highlights the outline of the proposed work to improve the accuracy of existing models.