Detection of Cardiac Events in Echocardiography Using 3D Convolutional Recurrent Neural Networks

Abstract

A proper definition of cardiac events such as end-diastole (ED) and end-systole (ES) is important for quantitative measurements in echocardiography. While ED can be found using electrocardiography (ECG), ES is difficult to extract from ECG alone. Further, on hand-held devices ECG is not available or cumbersome. Several methods for automatic detection of cardiac events have been proposed in the recent years, such as using a 2D convolutional neural network (CNN) followed by 1D recurrent layers. This structure may be suboptimal, as tissue movement has a spatio-temporal nature which is ignored in the CNN. We propose using a 3D CNN to extract spatio-temporal features directly from the input video, which are fed to long short term memory (LSTM) layers. The joint network is trained to classify whether frames belong to either diastole or systole. ES and ED are then automatically detected as the switch between the two states. The 3D CNN + LSTM model performs favourably at detecting cardiac events on a dataset consisting of standard B-mode images of apical four-and two-chamber views from 500 patients. The mean absolute error between events in the apical four-chamber view is 1.63 and 1.71 frames from ED/ES reference respectively. Model inference is fast, using (30 ± 2) ms per 30 frame input sequence on a modern graphics processing unit.