Artificial intelligence machine learning-based coronary CT fractional flow reserve (CT-FFR_ML): Impact of iterative...

The influence of computed tomography (CT) reconstruction algorithms on the performance of machine-learning-based CT-derived fractional flow reserve (CT-FFR_ML) has not been investigated. CT-FFR_ML values and processing time of two reconstruction algorithms were compared using an on-site workstation.

CT-FFR_ML was computed on 40 coronary CT angiography (CCTA) datasets that were reconstructed with both iterative reconstruction in image space (IRIS) and filtered back-projection (FBP) algorithms. CT-FFR_ML was computed on a per-vessel and per-segment basis as well as distal to lesions with ≥50% stenosis on CCTA. Processing times were recorded. Significant flow-limiting stenosis was defined as invasive FFR and CT-FFR_ML values ≤ 0.80. Pearson's correlation, Wilcoxon, and McNemar statistical testing were used for data analysis.

Per-vessel analysis of IRIS and FBP reconstructions demonstrated significantly different CT-FFR_ML values (p ≤ 0.05). Correlation of CT-FFR_ML values between algorithms was high for the left main (r = 0.74), left anterior descending (r = 0.76), and right coronary (r = 0.70) arteries. Proximal and middle segments showed a high correlation of CT-FFR_ML values (r = 0.73 and r = 0.67, p ≤ 0.001, respectively), despite having significantly different averages (p ≤ 0.05). No difference in diagnostic accuracy was observed (both 81.8%, p = 1.000). Of the 40 patients, 10 had invasive FFR results. Per-lesion correlation with invasive FFR values was moderate for IRIS (r = 0.53, p = 0.117) and FBP (r = 0.49, p = 0.142). Processing time was significantly shorter using IRIS (15.9 vs. 19.8 min, p ≤ 0.05).

CT reconstruction algorithms influence CT-FFR_ML analysis, potentially affecting patient management. Additionally, iterative reconstruction improves CT-FFR_ML post-processing speed.