Automatic Quantification of Cardiac MRI for Hypertrophic Cardiomyopathy

Author: ORCID icon
Konda, Anudeep, Computer Science - School of Engineering and Applied Science, University of Virginia
Ordonez-Roman, Vicente, Department of Computer Science, University of Virginia

Hypertrophic cardiomyopathy (HCM) is the most common monogenic heart disease, characterized by unexplained left ventricular hypertrophy, myofibrillar disarray and myocardial fibrosis. Left and right ventricular mass, ejection fraction and myocardium wall thickness at different segments measured from cardiac cine MRI based on LV, RV segmentation and mean myocardial T1 measured from LV segmentation on native T1 maps are critical biomarkers for diagnosis and prognosis of HCM patients. Deep convolutional neural networks (DCNNs) have shown great promise in many medical image segmentation tasks, including cardiac MRI. However, due to the greatly increased variability in shape and size of heart chambers and often reduced image contrast, the segmentation for HCM is more challenging than healthy and other patient populations and the model trained on generic cardiac MRI is very likely to fail on HCM. In this study, we developed a cascaded deep convolutional neural network to automatically segment the epi and endocardium at end-diastole and end-systole phases from cine and native T1 images to calculate all variables of interest based on a database with 100 HCM patients. Ejection fraction, LV and RV mass, mean myocardial T1 and regional wall thickness at 6 automatically localized segments per slice were calculated with promising results. The model greatly reduces the post-processing time and inter/intra-observer variability in biomarker quantifications for HCM patients.

MS (Master of Science)
Hypertrophic Cardiomyopathy, Deep Learning, Cardiac Chamber Segmentation
All rights reserved (no additional license for public reuse)
Issued Date: