Automated assessment of image quality in 2D echocardiography using deep learning

Echocardiography is the most used modality for assessing cardiac functions. The reliability of the echocardiographic measurements, however, depends on the quality of the images. Currently, the method of image quality assessment is a subjective process, where an echocardiography specialist visually inspects the images. An automated image quality assessment system is thus required. Here, we have reported on the feasibility of using deep learning for developing such automated quality scoring systems. A scoring system was proposed to include specific quality attributes for on-axis, contrast/gain and left ventricular (LV) foreshortening of the apical view. We prepared and used 1,039 echocardiographic patient datasets for model development and testing. Average accuracy of at least 86% was obtained with computation speed at 0.013ms per frame which indicated the feasibility for real-time deployment.