Predicting Ejection Fractions from Echocardiogram Videos Using Deep Learning

Echocardiography is a widely used imaging modality for assessing cardiac function, with ejection fraction (EF) being a critical metric for diagnosing heart conditions. Analyzing video echocardiograms poses special problems because of the dynamic Data type and the requirement to capture temporal relationships. This study proposes 3D Convolutional Neural Networks (3DCNN) and (2DCNN with LSTM) working parallel to analyze Echonet’s dynamic dataset. The primary objective is to accurately estimate the heart’s ejection fraction from input video echocardiograms. The proposed model employs 3DCNNs to capture spatial patterns across different frames and LSTM layers to model temporal dependencies. By combining the strengths of both architectures, the model aims to extract informative representations and capture the dynamic changes in the heart’s structure and function. The experimental results on Echonet’s dynamic dataset demonstrate the effectiveness of the parallel 3DCNNand 2DCNN + LSTM model for ejection fraction estimation. The model achieves competitive performance with RMSE of 1.1, MAE of 3.2, and R2 of 0.80, indicating its potential for accurate and reliable evaluation of cardiac function from video echocardiograms. In conclusion, the proposed model offers a promising approach for estimating ejection fraction from video echocardiograms. Its ability to capture spatial and temporal information improves accuracy and provides a valuable tool for diagnosing and monitoring heart conditions. Also, the proposed model outperforms all other recent models that use the same dataset.