MSFP-Net: a multi-scale and multi-frequency enhanced 3D network with learnable priors for coronary artery segmentation in CCTA images

• MSFP-Net integrates multi-scale and multi-frequency cues for CCTA artery segmentation. • MFMS enhances low-contrast and fine-branch segmentation via spatial–frequency fusion. • PIE adds explicit learnable priors to strengthen coronary structural coherence. Coronary artery segmentation is crucial for imaging examinations, providing clinicians with important information, such as vascular structure, to guide treatment plans for coronary artery disease. However, coronary artery segmentation is extremely challenging due to its small and complex vessel branches, low contrast with surrounding tissues, and significant inter-patient anatomical variability. To deal with these issues, in this study, we propose MSFP-Net, a multi-scale and multi-frequency enhanced 3D network with learnable priors. Specifically, MSFP-Net integrates spatial and frequency-domain information through a multi-frequency enhanced multi-scale (MFMS) module and incorporates a Prior Information Enhancement (PIE) module that adaptively guides the segmentation with learnable structural priors. This hybrid design enables the network to capture both global context and fine vessel details while improving robustness to individual variability. We evaluated the MSFP-Net model using our in-house dataset CCTA80 and two publicly available datasets, ASOCA and ImageCAS. Experimental results demonstrate that our method consistently outperforms state-of-the-art methods, with over 2% improvement in Dice coefficient and significant reductions in Hausdorff Distance compared to 3D U-Net. These results highlight the potential of MSFP-Net as an effective and generalizable solution for coronary artery segmentation in clinical practice. The code is available at: https://github.com/wangak2/MSFP-Net .