Diabetic retinopathy detection using supervised and unsupervised deep learning: a review study

Abstract

The severe progression of Diabetes Mellitus (DM) stands out as one of the most significant concerns for healthcare officials worldwide. Diabetic Retinopathy (DR) is a common complication associated with diabetes, particularly affecting individuals between the ages of 18 and 65. As per the findings of the International Diabetes Federation (IDF) report, 35–60% of individuals suffering from DR possess a diabetes history. DR emerges as a leading cause of worldwide visual impairment. Due to the absence of ophthalmologists worldwide, insufficient health resources, and healthcare services, patients cannot get timely eye screening services. Automated computer-aided detection of DR provides a wide range of potential benefits. In contrast to traditional observer-driven techniques, automatic detection allows for a more objective analysis of numerous images in a shorter time. Moreover, Unsupervised Learning (UL) holds a high potential for image classification in healthcare, particularly regarding explainability and interpretability. Many studies on the detection of DR with both supervised and unsupervised Deep Learning (DL) methodologies are available. Surprisingly, none of the reviews presented thus far have highlighted the potential benefits of both supervised and unsupervised DL methods in Medical Imaging for the detection of DR. After a rigorous selection process, 103 articles were retrieved from four diverse and well-known databases (Web of Science, Scopus, ScienceDirect, and IEEE). This review provides a comprehensive summary of both supervised and unsupervised DL methods applied in DR detection, explaining the significant benefits of both techniques and covering aspects such as datasets, pre-processing, segmentation techniques, and supervised and unsupervised DL methods for detection. The insights from this review will aid academics and researchers in medical imaging to make informed decisions and choose the best practices for DR detection.