GCHA-Net: Global context and hybrid attention network for automatic liver segmentation

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer-Aided Diagnosis

Background:

• Liver segmentation is crucial for liver cancer diagnosis and surgical planning.
• U-Net is effective for medical image segmentation but suffers from spatial information loss due to downsampling.
• Existing methods require improvement in capturing both global context and local details.

Purpose of the Study:

• To propose a novel Global Context and Hybrid Attention Network (GCHA-Net) for improved liver segmentation.
• To address the loss of spatial information in traditional U-Net architectures.
• To enhance the adaptive capture of structural and detailed features in medical images.

Main Methods:

• Development of GCHA-Net incorporating a Global Attention Module (GAM) for channel and positional interdependencies.
• Integration of a Feature Aggregation Module (FAM) with a Local Attention Module (LAM) to focus on relevant liver regions.
• Utilized LiTS2017 and 3Dircadb datasets for experimental validation.

Main Results:

• GCHA-Net achieved high performance on the LiTS2017 dataset with Dice Per Case (DPC) of 96.5% and Dice Global (DG) of 96.9%.
• The model demonstrated superior performance compared to state-of-the-art models in liver segmentation.
• Experiments on the 3Dircadb dataset confirmed GCHA-Net's highest accuracy among closely related models.

Conclusions:

• GCHA-Net effectively captures global context and local details, improving liver segmentation accuracy.
• The proposed model overcomes limitations of U-Net by preserving spatial information.
• GCHA-Net shows significant potential for clinical applications in liver cancer diagnosis and surgical planning.