The impact of iterative reconstruction algorithms on machine learning-based coronary CT angiography-derived fractional flow reserve (CT-FFRML) values

Abstract

To evaluate the impact of an iterative reconstruction (IR) algorithm (advanced modeled iterative reconstruction, ADMIRE) on machine learning-based coronary computed tomography angiography-derived fractional flow reserve (CT-FFR_ML) measurements compared with filtered back projection (FBP). 170 plaque-containing vessels in 107 patients were included. CT-FFR_ML values were measured and compared among 5 imaging reconstruction algorithms (FBP and ADMIRE at strength levels of 1, 2, 3 and 5). The plaques were classified as, 'calcified" or "noncalcified" and "≥ 50% stenosis" or "< 50% stenosis', a total of four subgroups by consensus. There were no significant differences of CT-FFR_ML values among the FBP and ADMIRE 1, 2, 3 and 5 groups wherever comparisons were done at the level of subgroups (P = 0.676, 0.414, 0.849, 0.873, respectively) or overall (P = 0.072). There were 20, 21, 19, 19 and 29 vessels with lesion-specific ischemia (CT-FFR_ML ≤ 0.80) in FBP and ADMIRE 1, 2, 3 and 5 datasets, respectively, but no statistical differences were found (P = 0.437). Compared with CT-FFR_ML value of FBP dataset, the CT-FFR_ML values of 9 (5.3%) vessels from 8 patients (7.5%) in ADMIRE5 dataset switched from above 0.8 to below or equal to 0.8. There were no significant differences of the CT-FFR_ML values among the FBP and IR image algorithms at different strength levels. However, high iterative strength level (ADMIRE 5) was not recommended, which might have an impact on diagnosis of lesion-specific ischemia, although changes only occurred in a modest number of subjects.

Keywords: Coronary computed tomography angiography; Coronary stenosis; Image reconstruction; Machine learning; Myocardial fractional flow reserve.