Advanced Multi-Level Bidirectional Attention Network for Retinal Vessel Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Retinal vessel segmentation is vital for diagnosing eye conditions using fundus images.
• Existing neural networks struggle with feature loss and context fusion, especially with complex vessel structures and varying image brightness.
• Pathological image segmentation is challenging due to fine vessel curvature and background noise.

Purpose of the Study:

• To propose a novel multi-level bidirectional attention aggregation network for improved retinal vessel segmentation.
• To address feature loss in encoders and insufficient context fusion in skip connections.
• To enhance the segmentation of fine vessels and pathological features in retinal images.

Main Methods:

• Developed a Partial Encoder Block (PEB) to minimize feature loss during encoding.
• Introduced a Dynamic Direction Attention Module (DDAM) in skip connections for better anisotropic geometric representation and detail preservation.
• Implemented a Multi-Feature Fusion Module (MFFM) to integrate multi-level features, enhancing detail retention and noise suppression.

Main Results:

• The proposed network demonstrated significant improvements across multiple datasets (DRIVE, STARE, CHASEDB1).
• Achieved notable gains in AUC, F1-score, Sensitivity, and Specificity compared to existing methods.
• Specifically, DRIVE dataset saw improvements of 0.19% (AUC), 0.43% (F1), and 1.17% (Sensitivity).

Conclusions:

• The multi-level bidirectional attention aggregation network effectively enhances retinal vessel segmentation.
• The proposed modules (PEB, DDAM, MFFM) successfully address limitations of previous methods.
• The network shows superior performance in segmenting fine vessels and pathological features in retinal fundus images.