Towards Interpretable Diabetic Retinopathy Detection: Combining Multi-CNN Models with Grad-CAM

Abstract

Diabetic retinopathy (DR) is a leading cause of vision impairment and blindness, necessitating accurate and early detection to prevent severe outcomes. This paper discusses the utility of ensemble learning methodologies in enhancing the prediction accuracy of Diabetic Retinopathy detection from retinal images and the prospective utilization of Gradientweighted Class Activation Mapping (Grad-CAM) to maximize model interpretability. Using a dataset of 1,437 color fundus images, we explored the potential of different pre-trained convolutional neural networks (CNNs), including Xception, VGG16, InceptionV3, and DenseNet121. Their respective accuracies on the test set were 89.27%, 91.44%, 89.06%, and 93.35%. Our objective was to improve the accuracy of diabetic retinopathy detection. We explored methods to combine predictions from these four models we began with weighted voting, which achieved an accuracy of 93.95%, and subsequently employed meta-learners, achieving an improved accuracy of 94.63%. These approaches surpassed individual models in distinguishing between non-proliferative and proliferative phases of DR. These findings underscore the potential of these approaches in developing robust diagnostic tools for diabetic retinopathy. Furthermore, techniques like Grad- CAM enhance interpretability, opening the door for further advancements in early-stage detection and clinical integration automatically while maximising accuracy and interpretability.