A novel vessel extraction technique for a three-way classification of diabetic retinopathy using cascaded classifier

Diabetic Retinopathy (DR) is one of the severe complications of diabetes which if not timely diagnosed and treated, can lead to severe retinal damage including irreversible vision loss. The overburdened healthcare system, lack of skilled ophthalmologists, and long delays in reporting the results of fundus images are the main reasons for untimely intervention. AI-based diagnostic systems can improve the situation. In fact, World Economic Forum has listed AI-facilitated healthcare as one of the leading emerging technologies. Many deep learning algorithms have been proposed for the diagnosis of DR in recent years but these studies only considered one or two symptoms for detecting diabetic retinopathy such as exudates, vessels, or hemorrhages, and not all of them together. In this paper, we propose a diagnostic system that takes into account all symptoms for detecting diabetic retinopathy and proposed a novel retinal vessels extraction algorithms that overcame the other studies vessel extraction algorithm. In this study, we also classify abnormal images as having exudates or not. For this purpose, fundus image datasets were annotated by experts for the presence or absence of exudates, these retinal images are then passed through pre-processing stage for extraction of haemorrhages and retinal vessels separately, these pre-processed images are then fused to one image and passed to two classifier, one classifier trained on normal and abnormal retinal images and second classifier trained to classify DR due to other symptoms (haemorrhages/neovascularization) and DR due to exudates. These classifiers are arranged in a cascaded way for identifying image with no DR, DR with exudates and DR with other symptoms. Results show that the proposed extraction algorithms improve the performance of the state-of-the-art DR classification model. Our model achieved 0.88 accuracy, with 0.91 precision, 0.86 recall, and 0.88 F1 score.