Development of fully automated echocardiographic data interpretation technique

Abstract

The segmentation of cardiac structures is a routine task performed by clinicians, which is usually done by manually delineating each structure to extract indices to help in the diagnosis. Over the last years, deep learning techniques had been studied to automatize the whole echocardiogram interpretation process, including the view recognition and segmentation of the images, as well as the index extraction. In this work, we developed and validated a convolutional neural network architecture, based on U-Net, for 2D echocardiogram multi-structure segmentation. The dataset used for training was from CAMUS, an Open-Source dataset from 500 patients that includes annotations of the left ventricle cavity, the left ventricle myocardium, and the left atrium cavity. The results obtained showed a mean Dice coefficient of 0.93, 0.86, and 0.88 on the three structures, respectively. The model’s performance was further evaluated using other datasets such as the EchoNet-Dynamic dataset and data from Hospital Sant Pau, Barcelona. For the last one, ground-truth images were generated and validated to obtain quantitative metrics to evaluate the model performance. In addition, Doppler echocardiograms from hospital data were used to evaluate a view-recognition model for Doppler classification, trained with a different vendor echocardiographer. Results were analysed to identify errors due to multiple machine variability and difference echocardiogram view planes.