Cardiac multi-structure segmentation network based on the fused dual attention mechanism

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Diseases

Background:

• Accurate cardiac segmentation is vital for diagnosing cardiovascular diseases.
• Computed tomography (CT) images present challenges like blurred boundaries and similar tissue intensities, hindering segmentation.
• Existing methods struggle with precise delineation of cardiac structures.

Purpose of the Study:

• To develop an advanced 3D deep learning network for multi-structure cardiac segmentation in CT images.
• To enhance the accuracy of cardiac function quantification through improved segmentation.
• To overcome limitations of blurred boundaries and adjacent tissue interference in CT scans.

Main Methods:

• Proposed a 3D cardiac multi-structure segmentation network incorporating Multi-scale Channel Enhancement Attention (MCEA) and Spatial Decomposition with Channel Fusion Attention (SD-CA) modules.
• MCEA enhances multi-scale feature representation by integrating channel information across different scales.
• SD-CA generates parallel spatial and channel attention weights, focusing on regions of interest and reducing interference.

Main Results:

• Achieved high segmentation accuracy for left ventricle, myocardium, left atrium, right ventricle, and right atrium, with average Dice coefficients of 94.21% and 93.9%.
• Demonstrated strong performance with average 95% Hausdorff distances of 6.5483 and 4.36.
• Quantitative predictions of left ventricular ejection fraction (LVEF) and substructure volumes showed correlation coefficients > 0.9587 and high agreement in Bland-Altman tests (>94.8%).

Conclusions:

• The proposed 3D network effectively addresses challenges in cardiac CT image segmentation, particularly blurred boundaries and similar intensities.
• The MCEA and SD-CA modules significantly improve segmentation performance and the ability to quantify cardiac function indicators.
• Results indicate high accuracy and reliability for clinical diagnosis and treatment planning of cardiovascular diseases.