Vol.:(0123456789) 1 3 The International Journal of Cardiovascular Imaging https://doi.org/10.1007/s10554-018-1419-9 ORIGINAL PAPER Diagnostic performance of machine-learning-based computed fractional fow reserve (FFR) derived from coronary computed tomography angiography for the assessment of myocardial ischemia verifed by invasive FFR Xiuhua Hu 1  · Minglei Yang 2  · Lu Han 1  · Yujiao Du 1 Received: 8 May 2018 / Accepted: 23 July 2018 © Springer Nature B.V. 2018 Abstract To explore the diagnostic performance of a machine-learning-based (ML-based) computed fractional fow reserve (cFFR) derived from coronary computed tomography angiography (CCTA) in identifying ischemia-causing lesions verifed by invasive FFR in catheter coronary angiography (ICA). We retrospectively studied 117 intermediate coronary artery lesions [40–80% diameter stenosis (DS)] from 105 patients (mean age 62 years, 32 female) who had undergone invasive FFR. CCTA images were used to compute cFFR values on the workstation. DS and the myocardium jeopardy index (MJI) of coronary stenosis were also assessed with CCTA. The diagnostic performance of cFFR was evaluated, including its correlation with invasive FFR and its diagnostic accuracy. Then, its performance was compared to that of combined DS and MJI. Of the 117 lesions, 36 (30.8%) had invasive FFR ≤ 0.80; 22 cFFR were measured as true positives and 74 cFFR as true negatives. The average time of cFFR assessment was 18 ± 7 min. The cFFR correlated strongly to invasive FFR (Spearman’s coefcient 0.665, p < 0.01). When diagnosing invasive FFR ≤ 0.80, the accuracy of cFFR was 82% with an AUC of 0.864, which was signifcantly higher than that of DS (accuracy 75%, AUC 0.777, p = 0.013). The AUC of cFFR was not signifcantly diferent from that of combined DS and MJI (0.846, p = 0.743). cFFR ≤ 0.80 based on CCTA showed good diagnostic performance for detecting ischemia-producing lesions verifed by invasive FFR. The short calculation time required renders cFFR promising for clinical use. Keywords Coronary computed tomographic angiography · Fractional fow reserve · Invasive coronary angiography · Diagnostic performance · Machine-learning-based cFFR Introduction Invasive fractional fow reserve (FFR) is a well-established clinical standard for identifying ischemia-causing lesions to guide revascularization for better prognosis [1–3]. Recently, the application of computational fuid dynamics (CFD) to coronary computed tomography angiography (CCTA) allowed for CT-FFR (FFR CT of HeartFlow and cFFR of Siemens) to be calculated noninvasively, achieving satis- factory diagnostic performance, using invasive FFR as the reference standard [4–9]. However, the CFD-based model, especially the three-dimensional FFR CT , has high compu- tational demands, which limits on-site clinical use [10]. Recently, a new version of cFFR based on machine learning was introduced after being trained on a large database gen- erated from CCTA anatomies and CFD-base computation * Xiuhua Hu huxiuhua_srrsh@zju.edu.cn Minglei Yang minglei.yang@siemens-healthineers.com Lu Han 279845920@qq.com Yujiao Du 296176893@qq.com 1 Department of Radiology, School of Medicine, Sir Run Run Shaw Hospital, Zhejiang University, 3 East Qingchun Road, Hangzhou 310006, Zhejiang, People’s Republic of China 2 Biomedical Engineering, CT Collaboration of Siemens Healthineers, No. 278, Zhouzhu Road, Pudong New District, Shanghai 201318, People’s Republic of China