Related Experiment Video
Updated: Jan 7, 2026
Estimating Bilateral Atrial Function by Cardiovascular Magnetic Resonance Feature Tracking in Patients with Paroxysmal Atrial Fibrillation
Published on: July 20, 2022
MultiEchoNet: a multi-task network for left ventricular ejection fraction and mitral annulus diameter calculation
Mengli Zhou1, Mingen Zhong2, Kang Fan3
1Xiamen University of Technology, Xiamen, 361024, Fujian, China.
Insights
MultiEchoNet automates left ventricular function analysis using weakly supervised learning, improving cardiovascular disease diagnosis. This deep learning model accurately quantifies ejection fraction and annulus diameter from ultrasound images, enhancing efficiency and accuracy.
Area of Science:
- Cardiovascular Imaging
- Medical Artificial Intelligence
- Echocardiography Analysis
Background:
- Accurate quantification of left ventricular function is crucial for diagnosing cardiovascular diseases.
- Current clinical methods rely on time-consuming manual segmentation of ultrasound images.
- Existing automated methods may lack efficiency and accuracy in complex cardiac analyses.
Purpose of the Study:
- To develop an automated system, MultiEchoNet, for quantifying left ventricular ejection fraction (LVEF) and mitral annulus diameter (MAD).
- To address limitations of manual segmentation by employing a weakly supervised learning strategy.
- To enhance the efficiency and accuracy of cardiovascular disease diagnosis through advanced AI.
Main Methods:
- Introduced MultiEchoNet, a multi-task deep learning network utilizing weakly supervised learning.
- Integrated a novel task propagation module for efficient global semantic information capture and reduced computational cost.
- Employed a multi-task Transformer module for cross-task information extraction and mutual guidance, enabling concurrent segmentation and keypoint localization.
- Utilized peak detection for identifying end-systolic and end-diastolic frames for precise parameter calculation.
Main Results:
- Achieved high performance on public datasets (EchoNet-Dynamic, CAMUS) with Dice similarity coefficients of 93.51% and 93.18% for segmentation.
- Obtained excellent keypoint similarity scores (0.958 and 0.940) and high correlation coefficients for LVEF (0.845, 0.82) and MAD (0.971, 0.963).
- Demonstrated robust support for auxiliary diagnosis of cardiovascular diseases.
Conclusions:
- MultiEchoNet effectively automates the quantification of left ventricular function parameters.
- The proposed weakly supervised multi-task learning approach significantly improves accuracy and efficiency in echocardiographic analysis.
- This AI-driven tool shows strong potential for clinical application in cardiovascular disease diagnosis.
Abstract:
Quantification of left ventricular function is essential for diagnosing cardiovascular diseases. Current clinical practice requires interactive segmentation of ultrasound images to delineate the left ventricular region and identify keypoints such as the apex and mitral annulus, a process that is both time-consuming and inefficient. To address these limitations, we introduce MultiEchoNet, a multi-task network employing a weakly supervised learning strategy to automatically calculate the left ventricular ejection fraction (LVEF) and mitral annulus diameter (MAD). Our approach integrates a novel task propagation module designed to improve the network's ability to capture global semantic information for each task at reduced computational cost, thereby minimizing task interference and enhancing task-specific feature extraction. Furthermore, we developed a multi-task Transformer module to facilitate the extraction of complementary modality information across tasks, promoting mutual guidance and optimization. This enables concurrent left ventricular segmentation and keypoint localization. In addition, peak detection is utilized to identify the end-systolic frame and end-diastolic frame in the echocardiographic sequence generated by the network, allowing for the precise calculation of related parameters. Experimental evaluations on public datasets EchoNet-Dynamic and CAMUS demonstrate that our algorithm achieves Dice similarity coefficients of 93.51% and 93.18%, respectively, and the highest keypoint similarity scores were 0.958 and 0.940, respectively. Additionally, the correlation coefficients between the predicted and true LVEF values were 0.845 and 0.82, respectively, while those for MAD were 0.971 and 0.963, respectively. These results suggest that MultiEchoNet offers robust support for the auxiliary diagnosis of cardiovascular diseases. Code is available at https://github.com/zzzmmmlll965/MultiEchoNet .
