DEEP LEARNING-BASED EARLY DETECTION OF DIABETIC RETINOPATHY USING FUNDUS IMAGES

Abstract

This paper proposes a deep learning model that together with high-resolution fundus images can be used to detect diabetic retinopathy (DR) at an early stage. The efficiency of the proposed architecture is to utilize EfficientNetv2 to extract representation in a usefully robust fashion and a Vision Transformer (ViT) to incorporate long-range spatial contextual information to improve lesion localization and classification performance. Data augmentation and gradually resizing images during the training process in the model helps to fix the class imbalance to enhance generalization. Besides, it is combined with Grad-CAM to enable interpretability by visual recognition of areas affecting the predictions. It has been trained and tested on more than 100 000 fundus images gathered on EyePACS and Messidor-2. It attains a high classification performance (AUC-ROC, 96.3%, sensitivity, 94.1%, specificity, 92.8%) impressive than the standard CNN epitomes like ResNet 50 and DenseNet 121. Qualitative Grad-CAM heatmaps validate the value of the model to identify clinically relevant areas that validates its capability of applications in transparent and scalable DR screening in real-world low-resource scenarios.