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

• Published in: The International Journal of Cardiovascular Imaging Vol. 36; no. 6; pp. 1177 - 1185
• Main Authors: Li, Shujiao, Chen, Chihua, Qin, Le, Gu, Shengjia, Zhang, Huan, Yan, Fuhua, Yang, Wenjie
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 01.06.2020; Springer Nature B.V
• Subjects: Algorithms; Angiography; Blood vessels; Cardiac Imaging; Cardiology; Computed tomography; Datasets; Image reconstruction; Imaging; Ischemia; Iterative methods; Learning algorithms; Machine learning; Medical imaging; Medicine; Medicine & Public Health; Original Paper; Plaques; Radiology; Stenosis; Subgroups; Myocardial fractional flow reserve; Coronary computed tomography angiography; Image reconstruction; Machine learning; Coronary stenosis
• ISSN: 1569-5794; 1875-8312; 1573-0743; 1875-8312
• DOI: 10.1007/s10554-020-01807-7

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.