Automated detection and grading of Diabetic Retinopathy using deep convolutional neural networks

点击翻译按钮获取中文摘要

Developing automated algorithms for screening and assessing diabetes and its complications using digital**retinography can reduce the risks of visual impairment and the severity of disease complications. With the**prevalence of diabetes, which affects more than 400 million people worldwide, and the long-term micro/macro**vascular complications of the disease, which may affect seriously the kidneys, the heart, the nervous system**and the eyes, there is an urgent need of rapid, accurate and specific tests. This project focuses on automated and**early detection/grading of Diabetic Retinopathy (DR), one of the most frequent and serious complications of**the disease and the leading the cause of blindness amongst working-aged patients. In current practices,**health-care professionals inspect retinal images visually to detect and grade DR. However, the overwhelming**prevalence of the disease and the limited number of qualified professionals impose an enormous burden on**retina specialists. Furthermore, human inspection of these images is subjective. To enhance the chances of**success of different possible treatments, it is critical to identify and grade DR as early as possible, and in an**objective manner. This is where state-of-the-art artificial intelligence algorithms, such as deep convolutional**neural networks (CNNs), can make a significant impact, leveraging large-scale data sets of retinal images. The**overall objective of this project is to design a fully automated, efficient and accurate algorithm for detecting**and grading DR in retinal images. We intend to tackle a 4-class problem, which uses 4 levels of severity**following a clinical standard: mild, moderate, severe and proliferative. The focus will be on state-of-the-art**deep CNN approaches. Specific sub-objectives are: (1) investigating a novel semi-supervised learning function,**which embeds priors that are relevant to the problem, leveraging large sets of non-labeled images and**mitigating the limited sizes of training sets annotated by human experts; (2) designing, implementing and**testing a CNN architecture that accounts for the specific context of retinal images; and (3) evaluating the**algorithm by comparing the results to ground-truth annotations.