A Synergistic Multi-Scale Attention and Composite Feature Extraction Network for Coronary Artery Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Surgery

Background:

• Accurate coronary artery segmentation is vital for robot-assisted percutaneous coronary intervention (PCI).
• Existing methods struggle with artifacts, fine branches, and discontinuities in Digital Subtraction Angiography (DSA) images.
• These limitations hinder the precision required for surgical navigation.

Purpose of the Study:

• To develop a novel deep learning framework for precise coronary artery segmentation from DSA images.
• To improve the accuracy and topological connectivity of segmented vessels for enhanced surgical guidance.

Main Methods:

• A U-shaped deep learning architecture incorporating a Composite Feature Extraction Module (CFEM) and a Multi-scale Composite Attention Module (MCAM).
• CFEM captures tubular vascular features across scales; MCAM enhances fine branch perception and long-range dependencies via attention mechanisms.
• A combined Dice-Focal loss function optimized boundary accuracy and class imbalance.

Main Results:

• The proposed method achieved a Dice coefficient of 76.74%, clDice of 50.30%, and HD95 of 57.84 pixels on the ARCADE dataset.
• Demonstrated significant improvements over state-of-the-art approaches in segmentation accuracy and vascular connectivity.
• The framework effectively addresses challenges posed by artifacts and fine vascular structures.

Conclusions:

• The novel deep learning framework offers robust and accurate coronary artery segmentation for robot-assisted PCI.
• Enhanced segmentation precision and connectivity have substantial clinical potential for interventional surgery.
• This approach provides reliable vascular structure information, improving surgical navigation and outcomes.