AI-based approach to improve coronary artery disease diagnosis from cardiac PET

Wenhao Zhang, Hidesato Fujito, Aakash Dhananjay Shanbhag, Jirong Yi, Robert Miller, Buchwald Mikolaj, Mark Lemley, Joanna Liang, Jianhang Zhou, Bryan Bednarski, Valerie Builoff, Marcelo DiCarli, Damini Dey, Daniel Berman, Piotr Slomka

Abstract:

Introduction: Positron emission tomography (PET)/CT for myocardial perfusion imaging (MPI) provides multiple imaging biomarkers, which are currently evaluated separately. We developed a machine learning (ML) model that utilizes PET MPI data to improve the diagnostic accuracy for coronary artery disease (CAD) by integration of several complementary quantitative parameters related to flow, perfusion, function, and atherosclerosis.

Methods: Patients without known CAD (N=391) who underwent 82Rb PET-MPI and invasive coronary angiography within a 6-month period between 2010 and 2018 were retrospectively enrolled. We developed XGBoost ML models for obstructive CAD detection (defined as left main stenosis ≥50% or ≥70% in other arteries). Age, gender, deep learning model coronary artery calcium score (CAC) from CT attenuation correction maps, rate pressure product (RPP), left ventricle ejection fraction (LVEF), minimum segmental blood flow, flow reserve and perfusion measurements were included in the prediction. 10-fold cross-validation was employed on the entire dataset. Within each fold, the ML model was individually trained on 90% of the entire dataset and tested on the left-out data. We compared the model developed from image data alone (Image AI) with the model that additionally includes age, gender, RPP, and LVEF (Image AI+). These two ML models were also compared to ischemic total perfusion deficit (iTPD), stress total perfusion deficit (Stress TPD), minimum segmental myocardial flow reserve (MFR), and CAC. We assessed the area under the receiver operating characteristic curve (AUC).

Results: The cohort included 34.5% female, with a median age of 71 (inter-quartile range: 65-80). The prevalence of CAD was 236/391. The Image AI+ model (AUC 0.83, 95% Confidence Interval [CI] 0.78-0.88) had similar performance to the Image AI model (AUC 0.82 [0.76-0.88], p=0.065). Both Image AI and Image AI+ models had higher AUC compared with iTPD (0.78, [0.71-0.85], p<0.05), Stress TPD (0.77 [0.70-0.84], p<0.05), MFR (0.75 [0.68-0.82], p<0.05), and CAC (0.74 [0.66-0.82], p<0.05). iTPD was the most influential in contributing to the final predictions for the two ML models (Figure).

Conclusions: ML integrating perfusion flow and CAC scoring have the potential to significantly increase the diagnostic accuracy of PET MPI and provide these results in a fully automated and objective manner.