Motion factorization for echocardiogram classification

This paper presents an algorithm for video factorization of echocardiograms for qualitative cardiac function classification. Our method uses global video features and is robust to heart shape variation across patients and transducer motion. Unlike related approaches, our method neither requires locating specific structures nor manual intervention. On real-world data, our algorithm achieves 99% recall and 91% precision on the task of qualitatively classifying unseen A4C echocardio-gram videos as normal or impaired.