Clinical Report: Evaluation of Diagnostic Efficacy of Deep Learning and CFD for CT-FFR
Overview
This study compares the diagnostic performance of deep learning (DL) and computational fluid dynamics (CFD) approaches for CT-derived fractional flow reserve (CT-FFR) against invasive FFR.
Background
Coronary artery disease (CAD) is a leading cause of morbidity and mortality. CT-derived FFR (CT-FFR) offers a non-invasive alternative to invasive FFR, which is the gold standard for identifying ischemia-producing lesions. Understanding the performance of DL and CFD methods for CT-FFR is important for non-invasive functional assessments.
Data Highlights
Method
AUC
Correlation with Invasive FFR
Deep Learning
0.90 (95% CI: 0.88-0.93)
rho = 0.71 (p < 0.001)
Computational Fluid Dynamics
0.89 (95% CI: 0.86-0.92)
rho = 0.68 (p < 0.001)
Key Findings
DL and CFD methods showed high diagnostic performance with AUCs of 0.90 and 0.89, respectively.
Both methods were strongly correlated with invasive FFR (DL: rho = 0.71; CFD: rho = 0.68).
Stable performance was observed across different coronary vessels and cardiac phases.
In gray-zone lesions, DL and CFD had comparable correct classification rates (86.4% vs. 84.6%).
No significant differences in diagnostic performance were found between the two methods (p > 0.05).
Clinical Implications
The findings indicate that both DL-based and CFD-based CT-FFR can be reliably used for non-invasive assessment of hemodynamically significant stenosis. This supports their potential integration into clinical workflows for evaluating coronary artery disease.
Conclusion
DL and CFD approaches for CT-FFR exhibit similar diagnostic performance.
Shear wave velocity measurements in the basal anteroseptal and right ventricular walls differed between transthyretin and light chain cardiac amyloidosis when conventional echocardiographic parameters did not.
