Optimized Deep Learning Framework for Diabetic Retinopathy Detection and Classification Using Fundus Imaging

Abstract

Diabetic Retinopathy (DR) is one of the primary causes of blindness all around the world, early and accurate detection to avoid loss of human vision is important. This research presents an optimized hybrid deep learning approach for DR detection from fundus images. Deep learning models such as Convolutional Neural Networks (CNNs) are proficient in capturing local features of an image while on the other hand, Vision Transformers (ViTs) excel at modeling global context, both have certain limitations in a standalone context. This research offers a novel hybrid feature fusion framework to overcome these individual limitations. We extracted deep feature vectors simultaneously from a pre-trained AlexNet (CNN) and a Swin Transformer model, then we combined them into a complete picture. The proposed strategy was rigorously tested on the public APTOS 2019 dataset, where it surpassed all baseline models significantly. The Random Forest classifier, when trained on the combined features, achieved a state-of-the-art accuracy of 98.2% and a huge refinement of 70% compared to the best individual deep learning model. Our study demonstrates that by combining CNN and ViT we can create a strong and reliable tool for detecting DR. Our approach presents an effective way to build automated systems/tools that could help the healthcare providers in finding out about the disease early on and saving the eyesight of their patients.