Left Ventricular Ejection Fraction Estimation for Pediatric Patients using Convolutional Neural Network Model

Ejection Fraction (EF) estimation is a major and time consuming task in diagnosis of cardiac pathologies. This mandates the requirement of trained cardiologists with greater years of experience and is prone to subjective errors. Ejection fraction below
$30{\mathrm{\% }}$
is considered as severe Left Ventricular Dysfunction, due to weakness of heart muscles and valves. The data set consists of videos of pediatrics who undertook an echocardiographic evaluation. The proposed work is of two modules: segmentation and ejection fraction estimation. Segmentation is done on the frames to estimate the left ventricular region of the heart using U-Net architecture. Individual frames of the videos are abstracted and annotated using binary masking. Diastolic and systolic areas are found from the PSAX view and length of the left ventricular is found from the
$\mathrm{A}4\mathrm{C}$
view. Volume of the left ventricular region is determined using systolic and diastolic volumes with the help of the Bullet Method. Various regression models have been developed and tested out to estimate the ejection fraction of the left ventricular region of the heart. The project contributes to quicker ejection fraction estimation with
$R2$
score of 0.8689 which is better compared to other existing solutions.