A novel micro-lesion segmentation network for diabetic retinopathy based on global axial-attention
Dingmeng Zhao, Hejiang YeObjective
Precise segmentation of diabetic retinopathy lesions remains challenging because of substantial morphological variability and severe foreground-background imbalance, particularly for small and sparsely distributed lesions such as soft exudates and microaneurysms. This study aims to improve the representation and discrimination of these lesions in fundus images.
Methods
We propose a Global Axial-Attention Network (GAA-Net) for retinal lesion segmentation. A global axial-attention decoder block is introduced to regulate skip-connection fusion, suppress irrelevant physiological structures, and preserve lesion-related features. In addition, an imbalance-aware training strategy combining lesion-centered targeted patch sampling and a stability loss is employed to reduce background-dominated learning. GAA-Net was evaluated on the public DDR and IDRiD datasets.
Results
On DDR, GAA-Net achieved the highest pixel-level AUPR for EX, SE, HE, and MA, reaching 62.79%, 49.17%, 53.58%, and 13.21%, respectively. On IDRiD, it achieved the highest pixel-level AUPR for OD, EX, SE, HE, and MA, with corresponding values of 92.50%, 80.90%, 35.99%, 32.81%, and 41.77%. Image-level analysis showed statistically significant AUPR improvements for all evaluated targets on both datasets (p<0.05). Although competing methods obtained higher F1-score or IoU for some categories, GAA-Net consistently performed best on the primary AUPR metric. Feature-map visualizations further indicated that the proposed attention mechanism more selectively emphasized pathological regions than the baseline U-Net.
Conclusions
GAA-Net improves probability-based retinal lesion discrimination under severe class imbalance, particularly for small and sparsely distributed lesions. The results support its effectiveness within the evaluated datasets, although further improvements in threshold-dependent mask overlap and external validation remain necessary.