A Deep Learning Framework for Multi-Level Diabetic Retinopathy Classification Using Fundus Images

Abstract

Diabetic retinopathy is a major cause of visual impairment in the world and thus a proper diagnosis must be carried out early and correctly to be treated. The paper presents a deep learning-based system of multi-level diabetic retinopathy classification using retinal fundus images. The model uses convolutional neural networks to automatically generate hierarchical features and categorize images into five stages of severity, including No DR, Mild, Moderate, Severe, and Proliferative DR. The APTOS 2019 Blindness Detection dataset is used for training and evaluation, ensuring a reliable and standardized benchmark. Preprocessing and data augmentation methods are used to improve model performance and generalization and minimize overfitting. The experiment has shown that the model is able to achieve an accuracy of 95.6% which is higher than the traditional base models like CNN, ResNet50, and EfficientNet. The soundness and permanence of the framework is statistically tested and validated through numerous repetitions of the framework. In addition, explainable AI is added to enhance interpretability, which would allow better clinical interpretation. The suggested system is a reliable and efficient solution to facilitate automated screening of diabetic retinopathy and help ophthalmologists to make clinical decisions.