Deep Learning and Fluid Dynamics On-Site CT-FFR Solution Compared to Off-Site FFRct and Invasive FFR.

BACKGROUND

On-site computed tomography (CT)-derived fractional flow reserve (FFR) solutions are increasingly needed to reduce delays, costs, and reliance on external platforms.

OBJECTIVES

This single-center prospective study evaluated the diagnostic performance of an on-site deep learning and fluid dynamic-based CT-FFR algorithm (xFFR, GE HealthCare) against off-site HeartFlow CT-FFR (FFRct) and invasive FFR (iFFR) for coronary artery disease (CAD) assessment.

METHODS

In this single-center prospective study, 250 symptomatic patients at intermediate-to-high CAD risk (mean age: 65 ± 9 years; 76% male) underwent coronary computed tomography angiography (CTA), xFFR, FFRct, and invasive coronary angiography with iFFR. Areas under the curve (AUCs) were calculated for xFFR and FFRct, with Spearman's correlations and Cohen's κ used to assess agreement with iFFR.

RESULTS

Functionally significant CAD was detected in 56.6% (xFFR), 54% (FFRct), and 48% (iFFR) of cases; xFFR showed sensitivity, specificity, and accuracy of 95%, 81%, and 88%, respectively. The overall diagnostic accuracy was comparable to FFRct (AUC: 0.91 vs AUC: 0.89; P = 0.274), superior only for left anterior descending coronary artery assessment (AUC: 0.96 vs AUC: 0.84; P = 0.001). Correlation analysis showed good agreement with iFFR (ρ = 0.67) and FFRct (ρ = 0.53). The mean xFFR analysis time was 8 ± 3.4 minutes.

CONCLUSIONS

This study establishes xFFR as a robust and efficient on-site tool for assessing CAD, demonstrating high diagnostic accuracy, reproducibility, and agreement with invasive methods. Its rapid processing and integration into clinical workflows position xFFR as a promising alternative to off-site FFRct solutions. Further studies are warranted to confirm its generalizability and optimize its implementation.