A novel hybrid layer-based encoder-decoder framework for 3D segmentation in congenital heart disease.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Cardiovascular diagnostics

Background:

• Accurate segmentation of 3D cardiac anatomy is vital for diagnosing and planning treatments for congenital heart disease (CHD).
• Existing deep learning models struggle with 3D CHD images due to limited data and complex tissue variations.
• This necessitates advanced methods for precise segmentation of cardiac structures.

Purpose of the Study:

• To develop and evaluate a novel hybrid layer-based encoder-decoder framework for 3D CHD image segmentation.
• To enhance the accuracy of segmenting the whole heart (WH) and great vessels in 3D medical images.
• To address the limitations of current deep learning approaches in handling complex cardiac anatomies.

Main Methods:

• Proposed a novel hybrid layer-based encoder-decoder framework incorporating global volume mixing and local multihead attention modules.
• Utilized a self-attention mechanism to capture local and global dependencies within 3D image data.
• Trained and evaluated the model on public datasets (ImageCHD and HVSMR-2.0) for 3D CHD image segmentation.

Main Results:

• The proposed framework demonstrated superior performance in segmenting the whole heart (WH) and great vessels compared to existing popular networks.
• Achieved high Dice coefficients and Intersection over Union (IoU) indices, indicating excellent segmentation accuracy.
• Effectively learned shape boundary features crucial for accurate anatomical segmentation.

Conclusions:

• The novel hybrid framework offers a significant advancement in 3D medical image segmentation for congenital heart disease.
• The model's ability to capture both local and global dependencies improves the segmentation of complex cardiac and great vessel structures.
• This approach holds promise for improving diagnostic accuracy and treatment planning in CHD patients.