Automated Estimation of Motion Patterns of the Left Ventricle Supports Cardiomyopathy Identification.

We propose a method to extract 3D shape and motion features of the left ventricle from 2D cine scans and investigate their utility in a cardiomyopathy detection task. To this aim, we develop an automatic processing pipeline that builds a 3D model of the left ventricular endocardium and epicardium from 2D cine cardiac MRI exams. We analyze a database of 1,045 clinical MRI studies including a combination of healthy subjects and cardiomyopathy patients with clinically reported disease labels. We use manifold learning techniques to extract shape and motion features from sequences of 25 3D shapes generated for each patient data set and representing a cardiac cycle. Finally, we train a disease classifier using a combination of simple metrics, shape features and motion features. On a testing set of 197 subjects, this classifier achieves the best performance in identifying patients from healthy subjects (AUC 0.922), as well as distinguishing 4 different classes of cardiac diseases (AUC 0.909).