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

Coronary artery disease (CAD) remains one of the leading causes of mortality worldwide. Although invasive coronary angiography (ICA) is widely used in clinical practice, accurately identifying arterial stenosis is still challenging due to low contrast, heavy noise, and complex vessel morphology. This study introduces the Frequency-Domain Attention-Guided Topology-Refinement Network (FAT-Net) to enhance coronary artery segmentation and stenosis detection in ICA. FAT-Net integrates a frequency-domain Multi-Level Self-Attention (MLSA) mechanism with a cascaded fusion strategy, enabling effective modeling of vascular structures and contextual dependencies across high- and low-frequency components, while improving robustness against background noise. Additionally, the proposed Low-Frequency Decomposition Module (LFDM) performs multi-level wavelet decomposition to progressively denoise ICAs and preserve global vascular topology. High-frequency details are then restored via inverse fusion, continuously refining arterial edges and small branches. Extensive experiments demonstrate that FAT-Net achieves a mean Dice coefficient of 0.87 for coronary artery segmentation and a True Positive Rate (TPR) of 0.61 for stenosis detection. The high Dice coefficient indicates accurate vascular segmentation, while the TPR slightly exceeds the levels reported in prior automated stenosis assessment studies, suggesting clinically meaningful detection performance. These results suggests that FAT-Net has strong potential to support accurate CAD diagnosis and treatment planning.