Diagnostic Accuracy Of Deep Learning For Myocardial Perfusion Imaging In Men And Women With A High-Efficiency Parallel-Hole-Collimated Cadmium-Zinc-Telluride Camera: Multicenter Study

92 Purpose: We developed a deep learning (DL) model for combined analysis of raw upright and supine stress myocardial perfusion imaging (MPI) polar maps from a parallel-hole-collimated camera with cadmium-zinc-telluride detectors and we sought to evaluate DL performance compared to total perfusion deficit (TPD) and visual assessments in men vs. women. Methods: MPI and invasive angiography were performed within a 6-month interval in 1,160 patients (64% men) in 4 separate centers without known coronary artery disease (CAD) obtained from REFINE SPECT registry. Images were acquired in supine and upright positions after stress. Standard 17-segment summed stress score (SSS) was performed visually. Diameter stenosis ≥50% of the left main artery, or ≥70% in the left anterior descending, left circumflex, or right coronary artery, was considered obstructive CAD. DL was developed using raw polar maps, with specification of sex, and body mass index (BMI) as auxiliary variables. Training and testing data-sets included both men and women for prediction of obstructive CAD using repeated leave-one-center-out external validation (4 models built from 3 centers and tested in 4th center). Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visualize regions contributing to disease prediction on polar maps. Stress TPD was obtained for upright (U-TPD) and supine (S-TPD) acquisitions. Receiver operating characteristic analysis and pairwise comparisons of the area under the curve (AUC) were performed to test diagnostic performance separately in men and women on the stacked data from 4 models. For TPD and DL, the diagnostic cutoff values were set to match the specificity of visual read with SSS≥4 threshold. Results: The AUC for detection of obstructive CAD by DL was higher than SSS, S-TPD and U-TPD in men (p for all