MAEF-Net: Multi-attention efficient feature fusion network for left ventricular segmentation and quantitative...

Area of Science:

• Medical Imaging and Diagnostics
• Artificial Intelligence in Healthcare
• Cardiovascular Disease Analysis

Background:

• Clinical diagnosis of heart disease relies on cardiac chamber segmentation and functional metric quantification in dynamic echocardiography.
• Manual segmentation of the left ventricle and identification of end-diastolic and end-systolic frames for LVEF calculation are time-consuming and tedious.

Purpose of the Study:

• To develop a fully automated deep learning-based method for echocardiographic analysis.
• To automatically segment the left ventricle and detect cardiac phases for accurate LVEF computation.

Main Methods:

• Proposed a multi-attention efficient feature fusion network (MAEF-Net) for automated left ventricular segmentation.
• Implemented a multi-attention mechanism, deep supervision, and spatial pyramid feature fusion to enhance feature extraction.
• Automatically detected end-diastolic frames (EDFs) and end-systolic frames (ESFs) for LVEF calculation.

Main Results:

• On the EchoNet-Dynamic dataset: Left ventricular segmentation DSC of 93.10% ± 2.22%, cardiac phase detection MAE of 2.36 frames ± 2.23, and LVEF prediction MAE of 6.29%.
• On a private clinical dataset: Left ventricular segmentation DSC of 92.81% ± 2.85%, cardiac phase detection MAE of 2.25 frames ± 2.27, LVEF prediction MAE of 5.91%, and Pearson correlation coefficient r of 0.96.

Conclusions:

• The proposed MAEF-Net method offers a robust and automated solution for left ventricular segmentation and quantitative analysis in 2D echocardiography.
• This approach has the potential to significantly improve the efficiency and accuracy of cardiac diagnosis, aiding clinical decision-making.
• The study provides publicly available code and trained models to facilitate further research and clinical adoption.