A New Horizon in Diabetic Retinopathy Detection and Grading: A Narrative Review of Deep Learning Models

• Main Author: Daniel Filipe Teixeira de Melo Santiago
• Publication Date: 2023
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: https://hdl.handle.net/10216/150362
• topic: Medicina clínica; Clinical medicine

Diabetic retinopathy (DR) is a microvascular complication of diabetes mellitus and is one of the most frequent causes of blindness in the adult population worldwide, particularly in low-income countries. Early diagnosis and treatment are paramount in the management of this condition and, subsequentially, to ensurea better prognosis. Also, it is recommended by most clinical practice guidelines that there should be a universal screening for DR amongst diabetic patients. For this matter, in recent years, many studies have been conducted to assess the capability of artificial intelligence models for aiding on the early detection of DR. In particular, the use of deep learning models based on convoluted neural networks, which are trained to automatically recognize pathological patterns and extract lesion features from retinal fundus images, allows DR grading according to the five classes of the International Clinical Diabetic Retinopathy Disease Severity Scale and between referable or non-referable DR. This paper aimed to evaluate the accuracy of deep learning networks used in recent studies in detecting and classifying DR and how well they compare to the manual grading of the disease. For this purpose, we concluded that there are already several models, considered to be state-of-the-art, that can detect and grade DR with very high levels of accuracy, sensitivity and specificity, surpassing the capabilities of health care professionals, with the advantage of having a very high reproducibility and being able to operate continuously without experiencing the effects of fatigue. However, there are still some issues with the lack of interpretability of the models' reasoning and the lack of clinical validation in different contexts. Also, the deep learning models need an enormous amount of data to be trained and there is a limited availability of retinal fundus images with good quality. In this regard, we have also addressed the importance of preprocessing methods for the retinal images and data augmentation techniques in order to increase the quantity of material from where the models can be trained.