Learning Motion Based Auxiliary Task for Cardiomyopathy Recognition with Cardiac Magnetic Resonance Images

Abstract

Accurate analysis of the patient's heart function, and early diagnosis of myocardial disease can improve the treatment effect and reduce the medical cost significantly. Among the different medical imaging techniques, cardiac magnetic resonance (CMR) has high tissue contrast which is widely used in clinic. However, pro-processing CMR data manually for diagnose is extremely time consuming. To develop an automatic cardiomyopathy recognition algorithm among normal group, hypertrophic cardiomyopathy, and dilated cardiomyopathy group, we employ the CNN and LSTM to extract spatial and motion features. In addition, we propose a motion based auxiliary task to help the main recognition task, without additional annotation. In experiment, compared to C3D [1] and LRCN [2], the proposed method obtains the best performance. Both accuracy and AUC score achieve 0.94.