A Lightweight Multi-scale and Multi-contextual Feature Fusion Approach for Diabetic Retinopathy Detection Using Deep Learning

Diabetic retinopathy (DR) is a leading cause of blindness worldwide, necessitating early detection for effective treatment and prevention of vision loss. Traditional methods for DR detection rely on manual assessment, which is time-consuming and prone to errors. Deep learning-based solutions have demonstrated great potential in automating DR detection, but challenges remain in capturing both fine-grained and large-scale lesion patterns effectively. This paper introduces a lightweight, multi-scale, and multi-contextual feature fusion approach for DR detection, integrating an enhanced MobileNetV3 backbone with advanced modules. Specifically, we propose the Cross-Scale Transformer Atrous Spatial Pyramid Pooling (CST-ASPP) module, which refines multi-scale lesion representation by incorporating self-attention mechanisms and cross-scale feature fusion. The Frequency Aware Dilated Convolutional Block Attention Module (FAD-CBAM) is introduced to improve feature extraction by emphasizing significant spatial and frequency-domain information. The final model leverages a multi-branch feature aggregation strategy to enhance lesion discrimination while maintaining computational efficiency. Experiments on standard DR datasets demonstrate that our proposed model outperforms existing state-of-the-art approaches in terms of accuracy, sensitivity, and F1-score, achieving robust lesion detection with minimal computational overhead. These results confirm the effectiveness of our approach in improving DR screening, making it a promising solution for real-world clinical applications.