Dynamic Severity-Aware Fusion Network for Accurate and Real-Time Diabetic Retinopathy Classification Using Multi-Scale Attention Mechanisms

Abstract

Diabetic retinopathy (DR) is a major cause of vision loss worldwide, requiring accurate and timely diagnosis for effective treatment. This study aims to address the limitations of existing machine learning and deep learning approaches, such as poor generalization, high computational complexity, and limited ability to capture severity-specific features. To achieve this, a novel Dynamic Severity-Aware Fusion Network (DSAF-Net) is proposed for efficient and reliable DR classification. The model integrates multi-scale convolutional neural network (CNN) feature extraction, a severity estimation module, and dual attention mechanisms (channel and spatial) to dynamically emphasize critical retinal features while suppressing irrelevant information. Additionally, adaptive feature fusion is employed to combine hierarchical representations, improving overall learning capability. The proposed method is evaluated using a publicly available Kaggle diabetic dataset with standardized preprocessing and optimized training settings. Experimental results demonstrate that DSAF-Net achieves superior performance with an accuracy of 97.8%, precision of 96.5%, recall of 96.1%, and F1-score of 96.0%, along with a low latency of 19 ms, outperforming traditional and existing deep learning models. These results indicate that the proposed model provides robust, accurate, and real-time classification. In conclusion, DSAF-Net offers an effective solution for automated diabetic retinopathy detection and has strong potential for deployment in clinical decision support systems.