A Multi-Scale Soft-Thresholding Attention Network for Diabetic Retinopathy Recognition

Diabetic retinopathy (DR) is a major cause of preventable vision loss, and its early detection is essential for timely clinical intervention. However, existing deep learning-based DR recognition methods still face two fundamental challenges: substantial lesion-scale variability and significant background noise in retinal fundus images. To address these issues, we propose a lightweight framework named Multi-Scale Soft-Thresholding Attention Network (MSA-Net). The model integrates three components: (1) parallel multi-scale convolutional branches to capture lesions of different spatial sizes; (2) a soft-thresholding attention module to suppress noise-dominated responses; and (3) hierarchical feature fusion to enhance cross-layer representation consistency. A squeeze-and-excitation module is further incorporated for channel recalibration. On the APTOS 2019 dataset, MSA-Net achieves 97.54% accuracy and 0.991 AUC-ROC for binary DR recognition. We further evaluate five-class DR grading on APTOS2019 with 5-fold stratified cross-validation, achieving 82.71 ± 1.25% accuracy and 0.8937 ± 0.0142 QWK, indicating stable performance for ordinal severity classification. With only 4.54 M parameters, MSA-Net remains lightweight and suitable for deployment in resource-constrained DR screening environments.