Abstract:
Diabetic retinopathy (DR) is a severe condition of the eyes associated with metabolic abnormal ities caused by diabetes, characterized by damage to the retinal vasculature. This may result in
irreversible vision loss if not detected at early stages. Regardless of the essential importance
of early intervention, the diagnosis of DR is often delayed due to the time-intensive nature of
retinal assessments and the shortage of ophthalmologists. This highlights the urgent need for
automated and accurate diagnostic tools to assist in the timely detection of DR. In this research,
we introduce an advanced deep-learning framework utilizing Grad-CAM++ to enhance fea ture extraction and improve the precision of DR classification. The framework employs three
DenseNet models, DenseNet-121(a), DenseNet-169, and DenseNet-121(b), to evaluate retinal
images from the APTOS 2019 dataset, which includes 44,570 images classified into five stages
of DR. Grad-CAM++ is used to highlight critical regions in the images, aiding in more accu rate classification of early-stage DR. Model 3 outperformed the other models with a training
accuracy of 95.63%, precision of 0.9921, recall of 0.9930, and an F1-score of 0.9925. These
findings reflect the potential of our method to increase diagnostic accuracy and minimize the
cost of healthcare systems. The proposed framework exhibits considerable advancements in the
analysis of retinal images, offering an adaptive solution for early DR detection.
