Diabetic Retinopathy - An Ensemble Approach

Abstract

Diabetes is a lifestyle disease that affects many people all over the world, with India leading the count of diabetic patients. The most important organ in the human body is the eye. Any anomaly will impact the functioning of life in its operation. The main component of the eye's internal surface, the fundus, is checked to spot any anomalies. In this study, neural networks were used to classify retinal fundus images. Methods of transfer learning are used to put the image into a category based on how bad the diabetic retinopathy is. Diabetes mellitus often evolves into diabetic retinopathy (DR), leading to lesions in the retina that impair vision. Through this paper, we propose an ensemble approach to respectively diagnose diabetes and diabetic retinopathy from blood reports and digital fundus images and accurately classify its severity. In order to do so, we first determine whether the patient has diabetes or not. This has been made possible by using machine learning classification algorithm – K Nearest Neighbors. A high-end graphics processing unit (GPU) was used to train the ensembled network on the publicly accessible APTOS-19[16] dataset, and the results are outstanding, particularly for a high-level classification task. Our proposed method worked more than 95% of the time. It has also been tested against the custom Messidor and EyePACS datasets.