Deep-Attention Feature Fusion Network for Automated Diagnosis of Diabetic Retinopathy Using Fundus Photographs

Diabetic Retinopathy (DR) is a leading cause of global blindness, affecting millions of individuals with diabetes. Early detection and precise classification of DR is essential for timely intervention and effective treatment, significantly reducing the risk of severe vision loss. The diagnosis of DR often relies on retinal fundus images, yet the manual assessment by trained ophthalmologists remains laborious, time-consuming, and subject to subjective interpretation. In this work, we frame diabetic retinopathy grading as a classification problem and develop a deep attention feature fusion network to diagnose DR using fundus photographs. The proposed framework initially extracts convolutional feature maps from multiple CNN models, each trained to capture different aspects of retinal images. These feature maps are then fused using simple yet effective convolutional attention to extract the most salient and localized features from the fused feature map. We conducted experiments using the large publicly available APTOS dataset. To evaluate our results, we performed an ablation study to assess the effectiveness of deep feature fusion and attention-based feature fusion. Our findings demonstrate that the proposed framework significantly enhances overall DR stage classification performance, achieving an average accuracy of 87% and markedly improving the performance of minority classes.