An Explainable AI-based Deep Learning Model for Classification of Diabetic Retinopathy Stages Using Retinal Fundus Images

Diabetic Retinopathy (DR) is a vision-related illness in humans that causes retinal degeneration and can eventually lead to blindness. Diabetic retinopathy is one of the main causes of vision impairment globally and a major consequence of diabetes mellitus. Diabetic retinopathy detection in early stage is a must, in order to prevent total blindness. Numerous physical examinations, including visual clarity assessment, pupil dilation, and optical coherence tomography, are used to diagnose diabetic retinopathy; nevertheless, they are time-consuming and affect patients. This research outlines an innovative approach using Convolutional Neural Networks (CNN) for the detection of Diabetic Retinopathy (DR). This study builds a diversified dataset of retinal fundus images in order to detect five phases of Diabetic Retinopathy (DR) using a customized deep CNN architecture. After considerable research and parameter fine-tuning, the CNN outperformed existing approaches in terms of accuracy, sensitivity, and specificity in diagnosing DR stages. This study uses 3 publicly available datasets, previously imbalanced images across 5 classes, was oversampled to new images, resolving the imbalance issue. The suggested CNN architecture attained a testing accuracy of 95.25%, with testing precision, recall, and F1 score values of 96%, 95%, and 95%, respectively, across the three datasets. However, as a black-box model, its operation is difficult to explain, raising worries about the reliability of its results. To address this issue, we employed Gradient-Weighted Class Activation Map (Grad-CAM) for elucidation, which emphasizes significant areas within the photographs. This approach delivers classification results as well as insights, which help medical practitioners make decisions. An interpretability study revealed important visual regions for decision-making. This method not only advances diagnostic systems, but it also shows promise in early DR detection, potentially assisting in timely identification in clinical settings. It underscores the capabilities of convolutional neural networks (CNNs) in medical image processing, offering the prospect of enhanced accuracy in diabetic retinopathy (DR) screening, hence benefiting diabetic patients and diminishing healthcare expenses.