Sparse Dynamic Volume TransUNet with multi-level edge fusion for brain tumor segmentation

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Neuro-oncology imaging

Background:

• Accurate 3D MRI brain tumor segmentation is vital for clinical diagnosis and treatment.
• Existing methods often overlook voxel information, inter-layer connections, and detailed features.
• This limitation impacts the precise localization and spatial distribution analysis of brain tumors.

Purpose of the Study:

• To propose a novel 3D brain tumor segmentation network, SDV-TUNet (Sparse Dynamic Volume TransUNet).
• To enhance segmentation accuracy by effectively integrating voxel information, inter-layer feature connections, and intra-axis details.
• To improve the clinical utility of 3D MRI in brain tumor diagnosis and treatment planning.

Main Methods:

• Developed a 3D encoder-decoder network (SDV-TUNet) incorporating Sparse Dynamic (SD) encoder-decoder and Multi-level Edge Feature Fusion (MEFF) modules.
• Employed a 3D extended shifted window strategy with multi-head self-attention and sparse dynamic adaptive fusion in the SD module for global feature extraction.
• Integrated multi-level local edge information via the MEFF module and skip connections to improve spatial edge information propagation.

Main Results:

• The SDV-TUNet method demonstrated superior performance on the BraTS2020 and BraTS2021 benchmarks.
• Achieved more accurate segmentation results compared to existing state-of-the-art brain tumor segmentation techniques.
• Effectively combined global semantic features with detailed voxel and inter-layer information.

Conclusions:

• The proposed SDV-TUNet network significantly advances 3D MRI brain tumor segmentation accuracy.
• The integration of diverse feature types (voxel, inter-layer, edge) is crucial for robust segmentation.
• This method holds promise for improving clinical decision-making in neuro-oncology.