Can deep learning identify normal coronary flow reserve from rubidium myocardial PET perfusion?

Abstract

Abstract Funding Acknowledgements Type of funding sources: None. Background Impaired vasodilator function is an early manifestation of coronary artery disease (CAD) and may precede angiographic stenosis. Rest and stress myocardial blood flow (MBF) are calculated from dynamic imaging during rest/stress Rubidium (Rb) myocardial PET perfusion. Coronary flow reserve (CFR) equals stress divided by rest MBF. CFR is an independent predictor of cardiac mortality in patients with known or suspected CAD. We evaluated the prediction of CFR from analysis of stress/rest PET images by deep learning (DL) as compared with standard calculation of CFR using supervised learning applied methodology using within a commercial DL training platform. Methods 1036 patients (625 male, 411 female, mean age: 64.3 years old) were studied. Patients underwent Stress/rest Rb PET perfusion, and CFR calculated using MBF software by an expert user. Abnormal CFR was defined as <2.0. The left ventricle myocardium was segmented using standard software. DL was trained using polar distribution of normalized PET uptake at stress and rest, processed stress and rest images were cropped, the stress images were then subtracted from the rest images. DL was trained using 935 subtracted images and tested using the remaining 101 images. DL was trained with supervision to classify images. The image shows examples of subtracted abnormal cases (1a & 1b). Results Using our supervised training methodology, the commercial MBF software platform reported 465 cases as abnormal, with 48 of these were included in the DL test set. The DL platform produced abnormal output classifiers for all the whole test set. DL accurately detected over 70% of abnormal cases. The commercial MBF software reported 571 cases as normal; with 50 of these contributing to the DL test set. DL was accurate in 48.0% of normal cases. Statistical results are shown in the table. Conclusion We have shown the proof of concept that DL algorithms trained with supervision can detect abnormal CFR. Our work shows that further work is needed to develop supervised learning methodology in order to improve accuracy for clinical use. Statistical Results Statistic Value Sensitivity 63.16% Specificity 56.67% + Predictive Value 48.00% - Predictive Value 70.83% Accuracy 59.18% Abstract Figure.