Validation and diagnostic performance of a fast on-site deep learning-based CT-FFR algorithm

Abstract

Background/Introduction CT-based fractional flow reserve (CT-FFR) has been extensively studied and established as a valuable tool for clinical decision making over the past decade. Nevertheless, clinical implementation has not been systematically adopted due to economic and technical reasons. Among the latter, the turn-around time for the computation and analysis' results potentially plays an important role. Purpose To evaluate the feasibility and diagnostic accuracy of CT-FFR computed on-site with a novel, deep learning-based algorithm using invasive hemodynamic indices as the reference standard. Methods Sixty-one patients who underwent clinically indicated coronary computed tomography angiography and invasive FFR (iFFR) and/or instantaneous wave-free ratio (iFR) measurements were retrospectively included. CT-FFR analysis was performed in 77 arteries using an on-site prototype software based on deep learning algorithms for coronary anatomy segmentation and prediction of the pressure drop under rest and hyperemia. The diagnostic performance of CT-FFR to detect hemodynamic significant lesions was assessed using iFFR (≤0.8) and/or iFR (≤0.89) as the standard of reference (60 patients with iFFR, 11 patients with iFR, and 3 patients with both) and the receiver operating characteristic area under the curve (AUC) was calculated. Furthermore, correlation analysis and Bland-Altman (BA) analysis was performed. Time for analysis including processing and manual edits to the lumen segmentation was recorded. Results CT-FFR analysis was successful in 59 (97%) patients and 74 (96%) arteries. In 74 arteries, 31 of 74 coronary lesions were invasively found to be hemodynamically significant. Total mean time for per patient CT-FFR analysis was 7 minutes and 55 seconds. Compared with invasive indices, per-lesion sensitivity and specificity of CT-FFR were 90%, and 98%, respectively. The AUC of CT-FFR vs. invasive indices for hemodynamic significance was 0.94, (95% confidence interval: 0.86–0.98). Compared to iFFR, CT-FFR correlated well (r=0.77) with only a very small bias (0.02) and narrow BA limits of agreement (−0.14 to 0.17). The per-lesion accuracy, sensitivity and specificity vs. iFFR were 96%, 93%, and 100%, respectively. Conclusion A novel deep learning-based CT-FFR algorithm yields excellent diagnostic accuracy compared to invasive hemodynamic indices to detect lesion-specific ischemia and offers the potential to be readily implemented into clinical practice given that it can be performed fast and on-site. Funding Acknowledgement Type of funding sources: None.