REFA-DINO: A Relative-Efficient Fusion Attention Framework for Binary and Multiclass Diabetic Retinopathy Classification.
Authors
Affiliations (4)
Affiliations (4)
- International Center for Materials Sciences and Technology, Western Caspian University, Baku, Azerbaijan, wcu.edu.az.
- Jadara Research Center, Jadara University, Irbid, Jordan, jadara.edu.jo.
- Department of Software Engineering, University of Engineering and Technology, Taxila, Taxila, Pakistan, vnu.edu.vn.
- Department of Computer Science, Aalto University, Espoo, Finland, aalto.fi.
Abstract
Diabetic retinopathy (DR) has become a major cause of preventable blindness, and its early detection is crucial for maintaining vision in diabetic patients. Existing computer-aided diagnosis (CAD) systems still face challenges in detecting early-stage lesions, integrating local and global features, and maintaining consistent performance in binary and multiclass settings. To address these limitations, this study proposes a relative-efficient fusion attention framework (REFA-DINO) that integrates a convolutional local-feature learning module with a transformer-based global contextual method. The architecture employs a relative-efficient fusion adapter (REFA) block comprising two branches, EfficientNet and attention-enhanced patch embedding. EfficientNet extracts lesion-sensitive retinal features, and an attention-enhanced patch-embedding module is used to generate token representations. An adaptive relative position-aware cross-attention mechanism is introduced to enable effective interaction among heterogeneous features, which are further processed through the DINO-based transformer module. The extracted global representations are passed to the linear classification head for DR classification. REFA-DINO is evaluated on two diverse and large-scale fundus imaging datasets, EyePACS and APTOS, for binary and multiclass classification. The proposed framework achieved accuracies of 98.60% and 93.77% for binary classification and 85.52% and 87.51% for multiclass classification on the APTOS and EyePACS datasets, respectively. A comparative analysis against state-of-the-art methods, cross-corpora evaluation, and an ablation study are also conducted to demonstrate the effectiveness of the proposed REFA-DINO. Compared with state-of-the-art methods, the REFA-DINO framework enhances binary classification on both the EyePACS and APTOS datasets, with significant improvements over the strongest baselines. The proposed approach attains up to 6.19% and 9.39% improvement in the binary and multiclass classification accuracy on EyePACS. Likewise, it achieves an improvement of 2.64% on the binary classification score on APTOS with competitive multiclass classification performance. Experimental findings highlight the robustness and generalizability of the REFA-DINO framework in classifying DR.