Systematic review on diabetic retinopathy segmentation and detection using deep learning techniques: research challenges and future directions

Diabetic Retinopathy (DR) is a critical cause of preventable blindness. The purpose of this article is to systematically review recent deep learning (DL) approaches for DR segmentation and detection using fundus images. A systematic search of the literature was carried according to the PRISMA 2020 guidelines, using the databases, ScienceDirect, Scopus, IEEE Xplore, and PubMed. Inclusion criteria are peer-reviewed English-language studies that used DL for DR detection or segmentation, utilizing fundus images. Exclusion criteria included: non-English articles, articles that described studies that did not experimentally verify the performance of DL to detect DR. 143 studies were examined related to blood vessel segmentation, lesion segmentation, and DR diagnosis, which utilize various deep learning methods like convolutional neural networks (CNNs), U-Net variations, attention-based models, and hybrid methodologies. The performance measures were dice score, accuracy, sensitivity, intersection over union (IoU), specificity, F1 score, and quadratic-weighted kappa coefficient. The reviewed models achieved > 90% sensitivity and specificity; however, issues were found in class imbalance, clinical generalization, and variance in image resolution. DL has provided additional pathways towards DR detection, but clinical applicability is a longer requirement in future research. Future studies should be cross-disciplinary studies and/or leverage new telemedicine technologies.