Objective To construct a deep learning-based artificial intelligence model to automatically quantify left ventricular ejection fraction (LVEF) using static views of echocardiography.
Methods The study included data of 1, 902 adults with left ventricular multi-slice echocardiographic views at end-systole and end-diastole. The collected dataset was divided into development set (1, 610 cases, with 1, 252 cases for model training and 358 cases for parameter adjustment), internal test set (177 cases for internal validation), and external test set (115 cases for external validation and generalization testing). The model achieved left ventricular segmentation and automatic quantification of LVEF through precise identification of the left ventricular endocardial boundary and inspection of key points. The Dice coefficient was employed to evaluate the performance of the left ventricular segmentation model, while the Pearson correlation coefficient and the intraclass correlation coefficient were used to assess the correlation and consistency between the automatically measured LVEF and the reference standard.
Results The left ventricular segmentation model performed well, with Dice coefficients ≥ 0.90 for both the internal and external independent test sets; the agreement between the automatically measured LVEF and the cardiologists' manual measurements was moderate, with Pearson correlation coefficients ranging from 0.46 to 0.71 and intragroup correlation analysis agreements from 0.39 to 0.57 for the internal test set; and Pearson correlation coefficients for the independent external test set were 0.26 to 0.54 and intra-group correlation analysis agreement of 0.23 to 0.50.
Conclusion In this study, a left ventricular segmentation model with better performance is constructed, and initial application of the model for automatic quantification of LVEF for two-dimensional echocardiography has general performance, which requires further optimisation of the algorithm to improve the model generalisation.
DownLoad: