FAT-Net: Frequency-Domain Attention-Guided Topology-Refinement Network for Coronary Artery Segmentation in Invasive...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Coronary artery disease (CAD) is a major global health concern.
• Accurate identification of arterial stenosis in invasive coronary angiography (ICA) is challenging due to image quality issues and complex vessel structures.

Purpose of the Study:

• To introduce the Frequency-Domain Attention-Guided Topology-Refinement Network (FAT-Net) for enhanced coronary artery segmentation and stenosis detection in ICA.
• To improve the accuracy and robustness of automated analysis of coronary angiograms.

Main Methods:

• Developed FAT-Net integrating a frequency-domain Multi-Level Self-Attention (MLSA) mechanism and a cascaded fusion strategy.
• Incorporated a Low-Frequency Decomposition Module (LFDM) for denoising and preserving vascular topology using wavelet decomposition.
• Restored high-frequency details for refining arterial edges and small branches.

Main Results:

• FAT-Net achieved a mean Dice coefficient of 0.87 for coronary artery segmentation, indicating high accuracy.
• The network demonstrated a True Positive Rate (TPR) of 0.61 for stenosis detection, surpassing previous automated methods.
• The results suggest clinically meaningful performance in identifying arterial stenosis.

Conclusions:

• FAT-Net shows significant potential for improving the accuracy of CAD diagnosis.
• The developed network can aid in more precise treatment planning for patients with coronary artery disease.
• FAT-Net offers a robust solution for analyzing challenging coronary angiogram images.