A stepwise strategy integrating dynamic stress CT myocardial perfusion and deep learning–based FFRCT in the work-up of stable coronary artery disease

Objectives To validate a novel stepwise strategy in which computed tomography–derived fractional flow reserve (FFRCT) is restricted to intermediate stenosis on coronary computed tomography angiography (CCTA) and computed tomography myocardial perfusion imaging (CT-MPI) was reserved for vessels with gray zone FFRCT values. Materials and methods This retrospective study included 87 consecutive patients (age, 58 ± 10 years; 70% male) who underwent CCTA, dynamic CT-MPI, interventional coronary angiography (ICA), and fractional flow reserve (FFR) for suspected or known coronary artery disease. FFRCT was computed using a deep learning–based platform. Three stepwise strategies (CCTA + FFRCT + CT-MPI, CCTA + FFRCT, CCTA + CT-MPI) were constructed and their diagnostic performance was evaluated using ICA/FFR as the reference standard. The proportions of vessels requiring further ICA/FFR measurement based on different strategies were noted. Furthermore, the net reclassification index (NRI) was calculated to ascertain the superior model. Results The CCTA + FFRCT + CT-MPI strategy yielded the lowest proportion of vessels requiring additional ICA/FFR measurement when compared to the CCTA + FFRCT and CCTA + CT-MPI strategies (12%, 22%, and 24%). The CCTA + FFRCT + CT-MPI strategy exhibited the highest accuracy for ruling-out (91%, 84%, and 85%) and ruling-in (90%, 85%, and 85%) functionally significant lesions. All strategies exhibited comparable sensitivity for ruling-out functionally significant lesions and specificity for ruling-in functionally significant lesions (p > 0.05). The NRI indicated that the CCTA + FFRCT + CT-MPI strategy outperformed the CCTA + FFRCT strategy (NRI = 0.238, p < 0.001) and the CCTA + CT-MPI strategy (NRI = 0.233%, p < 0.001). Conclusions The CCTA + FFRCT + CT-MPI stepwise strategy was superior to the CCTA + FFRCT strategy and CCTA+ CT-MPI strategy by minimizing unnecessary invasive diagnostic catheterization without compromising the agreement rate with ICA/FFR. Clinical relevance statement Our novel stepwise strategy facilitates greater confidence and accuracy when clinicians need to decide on interventional coronary angiography referral or deferral, reducing the burden of invasive investigations on patients. Key Points • A stepwise CCTA + FFR CT + CT-MPI strategy holds promise as a viable method to reduce the need for invasive diagnostic catheterization, while maintaining a high level of agreement with ICA/FFR. • The CCTA + FFR CT + CT-MPI strategy performed better than the CCTA + FFR CT and CCTA + CT-MPI strategies. • A stepwise CCTA + FFR CT + CT-MPI strategy allows to minimize unnecessary invasive diagnostic catheterization and helps clinicians to referral or deferral for ICA/FFR with more confidence. Supplementary Information The online version contains supplementary material available at 10.1007/s00330-023-10562-x.

Motion correction was performed for all images.The influx of contrast medium was measured using the arterial input function (AIF).The AIF was sampled in the descending aorta by including both cranial and caudal sections.MBF was derived from the time attenuation curves for each voxel.A dedicated semiautomated parametric deconvolution algorithm based on a 2compartment model of intra-and extravascular space was used.MBF was calculated as the ratio between the maximum slope of the fit curve and the peak AIF.To sample the MBF, the region of interest (ROI) was manually placed on a short-axis view on a segment basis according to the American Heart Association (AHA) 17-segment model.The ROI was placed to cover the whole segment without perfusion defects or cover the area (at least 50 mm 2 ) of suspected perfusion defects within the segment.According to the AHA recommendation, individual myocardial segments were assigned to the 3 major coronary artery territories and were adjusted for differences in dominance.Two experienced radiologists who were blinded to the participants' clinical history independently analyzed the CT-MPI data.For quantitative analysis, in stenosis-subtended territories, the absolute MBF was calculated as the mean value of MBF for segments with perfusion defects.Within reference territories, the absolute MBF was calculated as the mean value of MBF for all segments assigned to this territory.
CCTA and CT-MPI image quality was assessed with a 4-point scale (1 = poor; 2 = moderate; 3 = good; 4 = excellent).CCTA and CT-MPI images with poor image quality were excluded from the analysis.

Machine learning based FFRCT assessments
Deep learning-based FFR was computed by using a novel tree-structured recurrent neural network solution (TreeVes-Net).The framework was trained using a database of 13,000 synthetic coronary trees based on the geometric parameters of the coronary trees and was validated using a database of 180 real coronary trees with invasive FFR measurements.In the training database, the values of the geometric parameters were randomly prespecified in the appropriate ranges.The input to TreeVes-Net is the hydrodynamics-related geometric feature vector including local vascular features, local stenotic features, and global features for one position at the centerline of the coronary artery extracted from the input image of these synthetic coronary arteries.The distribution of FFR values along the coronary centerline is calculated by solving the Navier-Stokes equations with the finite element method.In TreeVes-Net, a fully connected multilayer neural network was used to embed the input image features, and bidirectional recursive neural network with long short-term memory was used to tackle the problem of spatial long-term dependency among fluid states at different points on coronary centerlines.For the output of the results, 3D color-coded coronary maps were generated to visualize the outcome.The lesion-specific FFRCT values were measured 20 mm distal to the stenosis.

Figure
Figure S1 Relationship between CCTA, FFRCT and CT-MPI and presence of

Figure
Figure S2 Correlation analysis and Bland-Altman Plots.

Figure
Figure S3Flowchart algorithm for the stepwise strategy: Sensitive analysis I

Figure
Figure S4 Flowchart algorithm for the stepwise strategy: Sensitive analysis II

Table S1 Sensitivity analysis of the diagnostic ability of stepwise strategies (intermediate diameter stenosis was defined as 40%-90%, and the grey zone of FFRCT value was defined as 0.76-0.86).
Data are n (%) or percentage, unless otherwise specified.CCTA, coronary computed tomography angiography; CT-MPI, computed tomography myocardial perfusion imaging; FFRCT, computed tomography derived flow fractional reserve; ICA, invasive coronary angiography; FFR, fractional flow reserve; FN, false negative; FP, false positive; TN, true negative; TP, true positive; NPV, negative predictive value; PPV, positive predictive value.Eur Radiol (2023) Lyu L, Pan J, Li D et al.