Multi-attention enhanced encoder-decoder network with hybrid transformer bottleneck for echocardiography image...

Area of Science:

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

Background:

• Accurate segmentation of 2D echocardiograms is crucial for diagnosing heart conditions.
• Existing models like U-Net (CNNs) struggle with long-range dependencies, while transformers may lack fine-grained localization.
• A hybrid approach is needed to leverage the strengths of both CNNs and transformers.

Purpose of the Study:

• To develop a novel deep learning architecture for improved 2D echocardiogram segmentation.
• To combine the local feature extraction of CNNs with the global context modeling of transformers.
• To enhance segmentation accuracy and robustness in cardiac imaging.

Main Methods:

• Proposed a novel encoder-decoder framework with multiple attention mechanisms.
• Integrated a hybrid bottleneck combining a Vision Transformer (ViT) and a Multi Receptive Field Block (MRFB).
• Utilized Atrous Spatial Pyramid Pooling (ASPP) for deep supervision and enhanced feature representation.

Main Results:

• Achieved high Dice Similarity Coefficients on the CAMUS dataset for left ventricle endocardium, epicardium, and left atrium.
• Demonstrated superior performance compared to pure CNN, pure transformer, and other hybrid architectures.
• The model effectively captures both local and global contexts for robust segmentation.

Conclusions:

• The proposed hybrid deep learning model significantly improves 2D echocardiogram segmentation accuracy.
• This advancement holds promise for more precise diagnosis and assessment of heart diseases.
• The model's ability to integrate local and global features offers a robust solution for medical image segmentation.