A REVIEW ON EXPLAINABLE CNN FOR EARLY DETECTION OF DIABETIC RETINOPATHY DIAGNOSIS

Abstract: Diabetic retinopathy (DR) is a leading cause of vision impairment globally, emphasizing the urgent need for early and accurate detection methods. Recent advancements in deep learning (DL) have demonstrated significant potential in automating DR diagnosis from retinal fundus images, thereby aiding clinicians in timely intervention. Nevertheless, the opacity of DL models remains a barrier to their widespread clinical adoption, necessitating transparent and explainable solutions. This paper proposes an integrated framework that combines state of the art deep learning architectures with explainable artificial intelligence (XAI) techniques, specifically Grad-CAM, to improve the interpretability of the diagnosis process. The methodology involves training multiple DL models, including a novel customized convolutional neural network (CNN), on high-resolution fundus image datasets, complemented by extensive data augmentation and preprocessing strategies to address class imbalance and image variability. The incorporation of XAI enables visualization of model decisions, fostering trust and facilitating clinical validation. Experimental results demonstrate that the proposed approach achieves high classification accuracy, superior early-stage detection capabilities, and meaningful interpretability insights, potentially enhancing clinical decision support systems for diabetic retinopathy.

Keywords: Diabetic Retinopathy, Retinal Fundus Images, Grad-CAM, Customized Convolutional Neural Network.