Author: Abdusalam Abdukerim 👨🔬
Affiliation: Institute for Medical Imaging Technology, Ruijin Hospital 🌍
Purpose:
Coronary computed tomography angiography (CCTA) is the gold-standard non-invasive test for coronary artery disease (CAD), but iodine contrast agents (ICA) pose limitations in specific populations. We trained a generative diffusion model (DM) to generate synthetic CCTA (syn-CCTA) from non-contrast CT (NCCT) scans, enabling contrast-free CAD and coronary artery stenosis (CAS) evaluation.
Methods:
DM was trained and validated on 1,459 paired NCCT-CCTA images (internal cohort: n=1,005; three external cohorts: n=75/102/277). Data preprocessing comprised image quality screening, region of interest segmentation, and non-rigid image registration to ensure optimal alignment. Performance was assessed via structural metrics (SSIM, PSNR, Dice) and clinical validation using automated CAS assessment software.
Results:
The model achieved Dice scores of 0.875 (internal cohort) and 0.862/0.867/0.870 (external cohorts). SSIM values were 0.783(internal cohort) and 0.769/0.775/0.778 (external cohorts); PSNR was 31.56 (internal cohort) and 30.21/30.75/31.12 (external cohorts). Clinical validation demonstrated good agreement between syn- and real CCTA images, with kappa coefficients of 0.723 and 0.739 for plaque type classification, and 0.772 and 0.812 for CAS in external cohorts 2 and 3, respectively. While most diagnostic metrics—including sensitivity, specificity, accuracy, negative predictive value (NPV), and AUC—demonstrated higher values (~0.80), positive predictive value (PPV) was notably lower, particularly for non-calcified plaques (~0.35) and mild stenosis cases (~0.45).
Conclusion:
We developed a diffusion-based generative model capable of synthesizing CCTA images from NCCT scans, achieving strong diagnostic performance in assessing CAS. The model demonstrated consistent sensitivity and specificity across multiple validation cohorts; however, its lower PPV—attributable to inherent limitations in the 2D architecture—resulted in false-positive cases. This contrast-free approach represents a clinically viable alternative for CAD screening in patients contraindicated for contrast agents.