A Deep Learning Based Approach for Automatic Cardiac Events Identification

Visually identifying End-Diastolic (ED) and End-Systolic (ES) frames from 2D echocardiographic videos without electrocardiogram is time-consuming but a fundamental step in routine clinical practice for assessment of cardiac structure and functionality. In recent years, several algorithms such as segmentation-based methods and regression-based methods, have been proposed to automatically identify ED and ES frames for automated cardiac function assessment. However, these methods require high quality images or large datasets which are difficult to obtain. In this work, we propose the first classification-based method, which is combined with a three-step postprocess for ED and ES identification from 2-D echocardiographic videos. We explored a deep-learning based model for this task which has lightweight structure with 175 kb parameters (in hdf5 format) and has no large demand for the number of echocardiographic videos compared with regression-based methods. In particular, we propose a weights-shared convolutional neural network module as the backbone, which is for two adjacent echocardiographic video frames feature extraction; and the backbone is combined with a classification module, which predicts the two frames’ relationship for volume-time prediction. Based on this, ED and ES can be effectively identified. We trained and evaluated the proposed method on apical 4 chamber views and apical 2 chamber views. For both cardiac views, the accuracies are above 0.95 and the AUCs (area under the ROC) are above 0.99. For ultimate ED and ES prediction, all the precisions are above 97% and all average Frame Distances(aFDs) are less than 2 frame errors. Moreover, we demonstrate that our method has considerable generalizability.