Deep learning automates detection of wall motion abnormalities via measurement of longitudinal strain from ECG-gated CT images

INTRODUCTION: 4D cardiac CT (cineCT) is increasingly used to evaluate cardiac dynamics. While echocardiography and CMR have demonstrated the utility of longitudinal strain (LS) measures, measuring LS from cineCT currently requires reformatting the 4D dataset into long-axis imaging planes and delineating the endocardial boundary across time. In this work, we demonstrate the ability of a recently published deep learning framework to automatically and accurately measure LS for detection of wall motion abnormalities (WMA). METHODS: One hundred clinical cineCT studies were evaluated by three experienced cardiac CT readers to identify whether each AHA segment had a WMA. Fifty cases were used for method development and an independent group of 50 were used for testing. A previously developed convolutional neural network was used to automatically segment the LV bloodpool and to define the 2, 3, and 4 CH long-axis imaging planes. LS was measured as the perimeter of the bloodpool for each long-axis plane. Two smoothing approaches were developed to avoid artifacts due to papillary muscle insertion and texture of the endocardial surface. The impact of the smoothing was evaluated by comparison of LS estimates to LV ejection fraction and the fractional area change of the corresponding view. RESULTS: The automated, DL approach successfully analyzed 48/50 patients in the training cohort and 47/50 in the testing cohort. The optimal LS cutoff for identification of WMA was −21.8, −15.4, and −16.6% for the 2-, 3-, and 4-CH views in the training cohort. This led to correct labeling of 85, 85, and 83% of 2-, 3-, and 4-CH views, respectively, in the testing cohort. Per-study accuracy was 83% (84% sensitivity and 82% specificity). Smoothing significantly improved agreement between LS and fractional area change (R(2): 2 CH = 0.38 vs. 0.89 vs. 0.92). CONCLUSION: Automated LV blood pool segmentation and long-axis plane delineation via deep learning enables automatic LS assessment. LS values accurately identify regional wall motion abnormalities and may be used to complement standard visual assessments.