SGEResU-Net for brain tumor segmentation.

Area of Science:

• Medical image analysis
• Artificial intelligence in oncology
• Neurosurgery and neurology

Background:

• Accurate brain tumor segmentation is crucial for diagnosis and treatment planning.
• The variability in tumor characteristics presents significant challenges for automated segmentation.
• U-Net based architectures have shown promise in improving segmentation accuracy.

Purpose of the Study:

• To develop a novel 3D U-Net model, SGEResU-Net, for precise brain tumor segmentation.
• To enhance feature learning and reduce noise in segmentation using spatial group-wise enhance (SGE) attention blocks.
• To improve overall segmentation performance through the integration of residual blocks and a self-ensemble module.

Main Methods:

• Construction of a novel 3D U-Net architecture named SGEResU-Net.
• Embedding residual blocks and spatial group-wise enhance (SGE) attention blocks within the 3D U-Net.
• Utilization of a self-ensemble module to further boost segmentation accuracy.

Main Results:

• SGEResU-Net demonstrated effectiveness on the BraTS 2020 and 2021 benchmarks.
• Achieved Dice Similarity Coefficient (DSC) scores of 83.31% (enhancing tumor), 91.64% (whole tumor), and 86.85% (tumor core) on BraTS 2021.
• Obtained Hausdorff distances (95%) of 19.278, 5.945, and 7.567 for the respective tumor subregions on BraTS 2021.

Conclusions:

• The proposed SGEResU-Net model significantly improves brain tumor segmentation accuracy.
• The integration of SGE attention and residual blocks enhances feature learning while minimizing noise.
• The model shows strong potential for clinical application in brain tumor diagnosis and treatment planning.