Diabetic Retinopathy Detection Using Deep Learning

Abstract

Diabetic retinopathy (DR) is a leading cause of vision impairment globally. Early and accurate detection of DR severity is crucial for timely intervention and prevention of blindness. This study investigates the potential of two cutting-edge deep learning models, Swin Transformer and FastViT, for automated DR severity classification from fundus images. To enhance the visibility of pathological features crucial for accurate classification, a novel image preprocessing pipeline is proposed. This pipeline combines Contrast Limited Adaptive Histogram Equalization (CLAHE) with Top-hat and Blackhat morphological operations, aiming to improve contrast and effectively highlight salient features like hemorrhages and exudates. We conduct a rigorous comparative assessment of Swin Transformer and FastViT, applying them to the publicly available APTOS 2019 dataset and DDR dataset, utilizing identical training parameters and the proposed preprocessing technique. This allows for a definitive evaluation of their respective performance in this critical diagnostic task. By systematically evaluating these advanced models under controlled conditions, this research aims to identify the most effective approach for automated DR severity classification, ultimately contributing to the development of reliable diagnostic tools that can aid in reducing the risk of blindness associated with diabetes.