Ejection Fraction estimation using deep semantic segmentation neural network

The Ejection Fraction value denotes how much blood is pumped out of the heart to different parts of the body. It is a routine clinical procedure in heart function assessment, where the left ventricle of the heart has to be manually outlined by doctors in clinical settings to measure the Ejection Fraction value which is time-consuming and highly varies by the observer. Most of the state-of-the-art automated Ejection Fraction estimation methods applied statistical or neural network models to generic and expensive clinical procedures like 3D ultrasound, MRI, and CT imaging. However, 2D echocardiography is a specialized diagnosis method that is inexpensive and routinely used in clinical settings to diagnose heart diseases. This paper proposed an automated Ejection Fraction estimation system from 2D echocardiography images using deep semantic segmentation neural networks. Two parallel pipelines of deep semantic segmentation neural network models have been proposed for efficient left ventricle (LV) segmentation in its systolic (contracted) and diastolic (expanded) states. The three different semantic segmentation neural networks, namely UNet, ResUNet, and Deep ResUNet, have been implemented in those parallel pipelines, and the performance of the proposed model has been studied on a standard 2D echocardiography data set. The most accurate model among the three achieved a Dice score of 82.1% and 86.5% in LV segmentation on end systole and end diastole states, respectively. The Ejection Fraction value is then determined by applying the volume measurement formula to the output of the left ventricle segmentation network. Therefore, the proposed automated Ejection Fraction system can be used in clinical settings to remove the eyeball estimation practice and reduce the inter-observer variability problem.

Automated summary

This paper proposes an automated Ejection Fraction estimation system from 2D echocardiography images using deep semantic segmentation neural networks. Two parallel pipelines of deep semantic segmentation neural network models have been proposed for efficient left ventricle segmentation, and three different neural networks, UNet, ResUNet, and Deep ResUNet, have been implemented. The most accurate model achieved high Dice scores for left ventricle segmentation in both systolic and diastolic states. The proposed system can remove the eyeball estimation practice and reduce inter-observer variability.