A Noninvasive Framework for Heart Function Assessment by Multitask Learning

Abstract

Accurate assessment of cardiac function is vital for preventing and managing cardiovascular diseases (CVDs). Recent advancements in machine learning, especially convolutional neural networks (CNNs) and multitask learning (MTL), have improved the precision of echocardiogram evaluations. However, existing methods often overlook the intrinsic relationships among ejection fraction (EF), end-diastolic volume (EDV), and end-systolic volume (ESV), which are essential for accurate assessments. We propose a noninvasive framework for heart function assessment (FHFA) using MTL that utilizes a 3-D CNN to extract key spatiotemporal features from echocardiogram videos. By employing an MTL strategy and weight distribution mechanism, this framework enhances the accuracy of EF predictions and provides a comprehensive assessment of cardiac structure and function. This approach ensures that the model effectively integrates auxiliary task information while focusing on the primary task, resulting in a more precise analysis of cardiac function. The experimental results on the Echonet-Dynamic dataset demonstrate that our method achieves an average absolute error of 3.89, a root-mean-square error (RMSE) of 5.13, and an $R^{2}$ value of 0.82, outperforming existing methods. Future work will focus on automatic weight optimization, model compression, and improving computational efficiency for broader clinical applications.