Multi-Modality Fusion and Tumor Sub-Component Relationship Ensemble Network for Brain Tumor Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning enhances brain tumor segmentation in multi-modality magnetic resonance imaging (MRI).
• Weighted fusion methods identify modality importance but struggle with fusing multi-modal and single-modal features.
• An effective fusion strategy is needed for improved segmentation accuracy.

Purpose of the Study:

• To propose a novel deep learning network for brain tumor segmentation in MRI.
• To address the challenge of fusing multi-modal and single-modal features effectively.
• To improve the accuracy and reliability of brain tumor segmentation.

Main Methods:

• Developed a multi-modality and single-modality feature recalibration network.
• Designed a dual recalibration module to integrate complementary multi-modal and specific single-modal features.
• Utilized the BraTS 2018 dataset for experimental validation.

Main Results:

• The proposed network outperformed existing multi-modal methods on the BraTS 2018 dataset.
• Spatial recalibration significantly improved segmentation results across evaluation metrics.
• Achieved Dice score increases of 1.7% (enhanced tumor core), 0.5% (whole tumor), and 1.6% (tumor core).

Conclusions:

• The proposed feature recalibration network offers an effective solution for MRI brain tumor segmentation.
• The dual recalibration module successfully integrates diverse image features for enhanced accuracy.
• This approach advances the application of deep learning in neuro-oncology diagnostics.