An Efficient and Explainable Hybrid CNN–Transformer Model for Retinal Image Classification

Diabetic Retinopathy (DR) is a progressive microvascular complication of diabetes and one of the leading causes of blindness worldwide, underscoring the need for early and automated screening systems. Existing deep learning methods for binary DR classification (DR vs. non-DR) often rely on complex architectures that require substantial computational resources, limiting their use in resource-constrained clinical environments. To address this limitation, we propose an efficient parallel feature fusion framework designed for retinal image analysis, combining a lightweight MobileNetV2 branch for robust local feature extraction with a MobileViTv2 branch for capturing essential global contextual information. Evaluated on a publicly available fundus image dataset, the proposed hybrid model achieved an accuracy of 97.84% for binary DR classification, outperforming established baseline architectures, including CNN-based models (ResNet18, EfficientNet-B4, VGG16, and MobileNetV3) and transformer-based models (ViT-Base, DeiT-Base, Swin-Base, and Swin-Tiny). The model also maintains computational efficiency with only 4.93 million parameters, and its interpretability is supported by Grad-CAM visualization, which confirms attention to clinically relevant regions. The combination of high accuracy, interpretability, and low computational cost positions the proposed MobileNetV2–MobileViT architecture as a practical solution for scalable and real-time DR screening in healthcare applications.