CSA ‐Net: A Channel‐Spatial Attention Guided Network for Classification of Glaucoma Stages

Glaucoma is a chronic blinding eye disease, where early detection and intervention can reduce the occurrence of irreversible vision damage. Existing deep learning methods based on convolutional neural networks (CNNs) have achieved excellent performance in glaucoma staging tasks, but most approaches focus on binary classification (normal or glaucoma) without considering precise multi‐stage classification of disease progression. Due to the high clinical similarity between different developmental stages of glaucoma and the progressive morphological changes in lesion areas, precise staging of glaucoma faces significant challenges. To address these issues, this paper constructs a novel channel‐spatial attention guided network called CSA‐Net. Specifically, a channel‐spatial attention module (CSAM), composed of local and global channel‐spatial attention blocks, is embedded after the feature extraction layers of the backbone network. This module effectively captures fine‐grained discriminative features in fundus images associated with disease progression by integrating local feature analysis, global context awareness, and establishing multi‐dimensional feature correlation mechanisms. Experimental results demonstrate that CSA‐Net outperforms FJA‐Net in key evaluation metrics: accuracy, F1‐score, and AUC improve by 0.25%, 0.71%, and 0.61% on the HDV1 dataset, and by 1.93%, 2.86%, and 1.65% on the LMG dataset, respectively.