DL-LVEF: Deep-Learning Measurement of the Left Ventricular Ejection Fraction from Echocardiographic Images

Agnese Sbrollini1, MHD Jafar Mortada1, Selene Tomassini2, Haidar Anbar1, Micaela Morettini1, Laura Burattini1
1Università Politecnica delle Marche, 2Università Politecnica delle Mache


Left ventricular ejection fraction (LVEF) is a commonly used index of cardiac functionality. Thus, accuracy in its measurement is fundamental. LVEF measure is usually manually performed by clinicians from echo-cardiographic images. Use of automatic algorithms could make LVEF measurement more objective. Thus, the aim of the present work is to pre-sent DL-LVEF, a new automatic algorithm for LVEF measurement. DL-LVEF was implemented in Google Colab Pro and includes two computa-tional phases, which are: 1) deep-learning identification and segmenta-tion of the left ventricular endocardium, performed by combining the YOLOv7 algorithm with a U-Net; and 2) LVEF computation, based on the Simpson's rule. DL-LVEF was set up and tested on the CAMUS database, which includes 1800 echocardiographic images acquired from 450 pa-tients with annotated LVEF values and manual segmentation of the left ventricular endocardium. The database was divided into training dataset (70%) and testing dataset (30%); 14% of the training dataset was used as validation dataset for defining the early stopping point. In both training and testing datasets (Table), measured LVEF values (mLVEF, %) and an-notated LVEF values (aLVEF, %) were found to be statistically highly correlated (ρ≥0.89); moreover, median mLVEF value was not statistically different from median aLVEF value. Eventually, mean absolute error (MAE, %) was ≤5%. Thus, DL-LVEF provided objective and accurate LVEF measurement. Future DL-LVEF evolutions will provide segmenta-tion of other cardiac anatomical structures and, thus, will allow meas-urement of other clinically relevant cardiac indexes.