Abstract
252017
Introduction: Several techniques have been proposed to generate 3D parametric maps of myocardial blood flow (MBF) for 82Rb positron emission tomography (PET). Artificial intelligence (AI), in particular, has shown to be a strong candidate method due to the high speed of image generation and accuracy in reproducing the results from conventional nonlinear least-squares regression of the one tissue compartment model. Previous studies have generally validated different methodologies of parametric mapping from an analytic point-of-view by correlating the MBF values obtained with reference 2D polar map processing and 3D parametric maps for the whole left ventricle (LV) and the three vessel territories (LAD, RCA, and LCx). However, no study has yet to verify that parametric map-derived MBF values for 82Rb PET correlate with the presence of coronary artery disease (CAD). In this study, we validated a previously trained AI model to perform MBF parametric mapping by evaluating its diagnostic ability to predict CAD and compared its performance with traditional polar map kinetic modeling.
Methods: Our dataset comprised N=3,253 patients who underwent rest and stress 82Rb cardiac PET studies and invasive coronary angiography (ICA) within three months at two centers (center A: N=1,518 and center B: N=1,735). The only exclusion criteria were patients with previous bypass surgery. Conventional 2D polar processing was performed using FlowQuant to generate MBF and myocardial flow reserve (MFR) polar maps as well as arterial input functions. These same arterial input functions were used to generate MBF parametric maps for the same patients using an AI model previously trained previously on N=550 independent patients from center A. In that respect, the center B data served as a completely independent external validation site. Patient-specific LV segmentations were used to bring the MBF parametric maps into polar map space for calculation of whole LV and vessel-wise metrics. Using a normative database of N=40 low-likelihood patients for CAD, total perfusion deficit (TPD) and focally impaired myocardium extent from an integrated myocardial flow reserve (iMFR) map were calculated for reference polar map processing and parametric mapping of patients from both centers. For each metric and image processing method (i.e., conventional polar vs 3D parametric mapping), we trained and evaluated a logistic regression model with 5-fold cross validation to predict significant stenosis ≥ 70%. The receiver operator characteristic (ROC) area under the curve (AUC) across the test fold predictions against ground truth labels was computed for both models and compared using permutation tests.
Results: CAD prevalence was 77.3% in center A and 73.9% in center B. For all models, there were no significant differences in the AUC from the parametric- or polar-derived TPD or iMFR either at the whole LV or per-vessel level for either center, with highly overlapping confidence intervals (average ΔAUC=0.005±0.005, p>0.05 for all). The AUCs for the models between both sites were comparable, ranging from 0.67-0.76 depending on the vessel territory and metric, corroborating the robustness of TPD and iMFR for diagnosing significant obstructive disease in these unselected general patient populations.
Conclusions: We have provided a large multicenter clinical validation of our AI model for deriving 3D MBF parametric maps, demonstrating that they recover the same whole-LV and vessel-wise blood flow information as conventional polar map processing. These maps can thus be used reliably for accurate quantification of MBF to effectively diagnose patients with significant stenosis due to CAD.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.