Identification of Diabetic Retinopathy Using Machine Learning

Abstract

Diagnosing diabetic retinopathy (DR) with colour fundus images is a difficult and time-consuming task due to a complex grading system and the demand for qualified doctors to determine the existence and importance of multiple microscopic characteristics. In this work, we propose a CNN approach for appropriately assessing DR severity from digital fundus images. We build a network with CNN architecture and data augmentations that can identify the intricate components necessary for the classification task, such as micro-aneurysms, exudate, and retinal hemorrhages, and then automatically offer a diagnosis without user input. We use a top-tier graphics processing unit (GPU) to train our network using the publicly available Kaggle dataset, and the results are excellent, especially for a challenging classification test. Our suggested CNN achieves a sensitivity of 95% and an accuracy of 75% on 5,000 validation images on the data set of 80,000 photos used.[1]