Oral Abstract Session
SCMR 22nd Annual Scientific Sessions
Background: The purpose was to assess the potential of multi-parametric perfusion mapping using an automated machine learning (ML) approach to predict not only obstructive epicardial coronary artery disease (CAD) vs normal, but also microvascular dysfunction (MVD). We hypothesized that using ML with multi-parametric maps would outperform detection of CAD by experts.
Methods: 50 patients (40 male, aged 63±8 years) with stable angina and 15 healthy controls (12 male, aged 45±8) underwent adenosine stress CMR at the Royal Free Hosp, London, UK. This study was approved by the local Ethics Committee with written informed consent for all subjects. Truth standards (CAD vs MVD vs normal) were derived from invasive coronary angiography with fractional flow reserve (FFR) and index of microcirculatory resistance (IMR) measured in all major epicardial arteries (angiographically critical stenoses unsafe to pass a pressure wire were assumed to have FFR<0.80), categorizing each subject as CAD (FFR<0.80), MVD (FFR>0.80 & IMR>25), and normal (FFR>0.8 & IMR< 25, or control). In-line perfusion mapping  generated quantitative myocardial blood flow (MBF) and other co-registered physiological parameter maps : MBF, myocardial blood volume, permeability surface area product, interstitial volume, arterial delay (TA), and capillary delay (TC). Manually contoured data (16 segments) were further divided into endo and epicardial sectors for each param. These data were used to generate a large number of potential features from which a subset was selected for training of various ML models with 5-fold cross validation. Hyperparameter optimization was used for feature and parameter selection. Scores included overall accuracy, sensitivity, and specificity. Both raw images and maps were classified as CAD/nonCAD by 2 experts.
Results: 27 patients had CAD (8 three-vessel) and 23 had non-obstructed arteries (7 IMR-, 16 IMR+). Example pixel-wise perfusion maps are shown in Fig 1. Overall CAD/nonCAD classification accuracy (Fig 2) was 65, 72, and 83%, with sensitivity of 80, 87, and 87% and specificity of 55, 62, and 81%, for expert using raw images vs raw image plus maps vs ML, respectively. Using ML, overall 3-category classification accuracy (Fig 3) was 77%, with sensitivity/specificity of 87/79%, 52/92%, and 82/92% (CAD, MVD, and normal). Hyperparameter optimization selected 8 features: global endo MBF, min endo MBF, global endo/epi MBF ratio, min MBF endo/epi ratio, min transmural MBF, max TC, max TA, min MBF endo-remote ratio.
Conclusion: Overall accuracy was high with a spectrum of ischemic heart disease consisting of patients with obstructive CAD, MVD, and a “normal” group consisting of 7 symptomatic patients with FFR>0.8 and IMR<25, and controls (mean age 45). ML classification of myocardial perfusion with multi-parametric perfusion mapping has potential for objective, automatic detection of disease and classification of epicardial vs microvascular disease.