Abstract
2259
Introduction: Low-dose ungated CT attenuation correction (CTAC) scans are commonly obtained with SPECT/CT myocardial perfusion imaging. We aimed to develop a deep learning model to compute CAC scores automatically and rapidly from CTAC scans and evaluate the prognostic significance of such scores.
Methods: We adapted a novel deep learning (DL) model from video analysis applications, that is based on convolutional long-short term memory network and has ability to processes multiple adjacent CT slices for additional spatial information. The model was trained to automatically quantify CAC on CTAC scans using 6,608 studies (that included 4781 low-dose ungated SPECT CTAC and 1827 ECG-gated CAC scans) and evaluated in an external site cohort of 2236 patients undergoing SPECT MPI with CTAC and with no history of prior coronary artery disease. Major adverse cardiac event (MACE) was defined as death, late revascularization (more than 90 days after scan), acute myocardial infarction or unstable angina. We compared the MACE risk stratification in 4 CAC score categories (0; 1-100; 101-400; >400) between quantitative CAC scores manually derived from SPECT CTAC maps by experienced readers (expert scores) and scores obtained fully automatically by deep learning (DL-CTAC scores) using univariable and multivariable Cox regression models, net reclassification index (NRI) and areas under receiver operating curve (ROC). Agreement between expert reader annotations and DL scores was analyzed in categories using Linearly Weighted Cohen’s Kappa.
Results: The testing cohort was 50% male, median age was 67 (95% confidence interval [CI]: 58, 74), and 277 patients experienced MACE in median follow up period of 2.9 years (95% CI: 1.8, 4.2). DL-CTAC were obtained fully automatically in less than 2 seconds per scan and exhibited very good category-wise agreement with expert scores (Cohen’s Kappa of 0.8, 95% CI 0.7, 0.82). Both expert and DL-CTAC scores were effective for MACE stratification for each CAC category, with stepwise increase in hazard ratios (HR) – up to 5.2 (95% CI 3.6, 7.7) in the highest CAC score category for DL-CTAC scores and up to 5.1(95% CI: CI: 3.5,7,4) for expert scores. The HRs remained significantly increased for all CAC score categories, both by expert reader and deep learning, after adjusting for age, sex and total perfusion deficit – up to 3 (95% CI: 2.0, 4.6) for DL CTAC scores and up to 2.9 (95% CI: 1.9, 4.5) for expert scores. NRI was non-significant between expert and DL-CTAC scores at -0.03 (95% CI: -0.16, 0.08). Area under ROC for the prediction of future MACE was similar for expert and DL-CTAC scores (0.69 for both types of scores).
Conclusions: CAC scores obtained by DL from low-dose ungated SPECT CTAC scans show a very good agreement with expert reader annotations and predict cardiovascular risk similarly to CAC scores derived by manual scoring but without consuming expert time and effort. The use of deep learning CAC scoring in all SPECT CTAC scans could provide additional prognostic information that is independent of perfusion deficit at no extra cost