CABNet

CABNet integrates category and global attention blocks to improve fine-grained diabetic retinopathy grading from fundus images by addressing class imbalance, intra-class variation, and small lesion detection such as microaneurysms, hemorrhages, and soft exudates.

Key Features:

• Category Attention Block (CAB): The CAB extracts region-wise discriminative features to mitigate imbalanced data distributions and emphasize regions relevant to each diabetic retinopathy grade.
• Global Attention Block (GAB): The GAB generates class-agnostic global attention feature maps to capture fine-grained details and small lesions such as microaneurysms, hemorrhages, and soft exudates.
• Integration with Backbone Networks: CAB and GAB modules aggregate with various backbone networks for end-to-end training while adding minimal additional parameters.

Scientific Applications:

• Diabetic retinopathy grading: Fine-grained classification of DR severity from fundus images under intra-class variation and class imbalance.
• Small lesion detection in retinal images: Improved identification of microaneurysms, hemorrhages, and soft exudates through attention-driven feature localization.

Methodology:

CABNet aggregates Category Attention Block (CAB) and Global Attention Block (GAB) modules with a backbone network, using attention mechanisms to extract region-wise discriminative features and generate class-agnostic global attention feature maps for end-to-end training.