Brain tumor segmentation by combining MultiEncoder UNet with wavelet fusion.

Area of Science:

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

Background:

• Accurate brain tumor segmentation from multimodal MRI is crucial for clinical diagnosis and surgical planning.
• Current deep learning methods often use early fusion, which can ignore complementary information between MRI modalities and limit performance.
• The localized nature of convolutional operations hinders the capture of long-range voxel relationships.

Purpose of the Study:

• To develop a novel multimodal segmentation network for improved brain tumor detection.
• To address limitations of early fusion strategies by employing a late fusion approach.
• To enhance the capture of both complementary inter-modal features and long-range spatial dependencies within brain tumors.

Main Methods:

• Proposed a multimodal segmentation network utilizing a late fusion strategy with specialized encoders for distinct MRI modalities.
• Incorporated a feature fusion module employing 3D discrete wavelet transform to extract complementary inter-modal features.
• Introduced a 3D global context-aware module to capture long-range dependencies among tumor voxels at a high feature level.

Main Results:

• The proposed model demonstrated competitive performance against state-of-the-art methods on the BraTS2018 and BraTS2021 datasets.
• Experimental results validate the effectiveness of the late fusion strategy and the global context-aware module.

Conclusions:

• The developed approach offers a novel concept for multimodal fusion in deep neural networks for brain tumor segmentation.
• The network achieves more accurate and promising segmentation results, showing potential to aid physicians in diagnosis.
• This method enhances the utilization of complementary information and long-range dependencies for superior segmentation outcomes.