multiPI-TransBTS: A multi-path learning framework for brain tumor image segmentation based on multi-physical...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Brain Tumor Segmentation (BraTS) is crucial for clinical management but challenged by tumor heterogeneity and MRI variability.
• Automated segmentation methods struggle with diverse tumor appearances, sizes, and intensities across different MRI modalities.

Purpose of the Study:

• To propose a novel Transformer-based framework, multiPI-TransBTS, for enhanced brain tumor segmentation accuracy.
• To leverage multi-physical information, including spatial, semantic, and multi-modal data, to address tumor heterogeneity.

Main Methods:

• The multiPI-TransBTS framework utilizes a multi-branch encoder for modality-specific feature extraction.
• An Adaptive Feature Fusion (AFF) module employs channel-wise and element-wise attention for effective feature recalibration.
• A multi-source, multi-scale decoder with Task-Specific Feature Introduction (TSFI) generates segmentation for Whole Tumor (WT), Tumor Core (TC), and Enhancing Tumor (ET) regions.

Main Results:

• multiPI-TransBTS demonstrated superior performance on BraTS2019 and BraTS2020 datasets compared to state-of-the-art methods.
• The model achieved improved Dice coefficients, Hausdorff distances, and Sensitivity scores for brain tumor segmentation.
• Further analysis highlighted the need to balance precision and recall for Enhancing Tumor (ET) segmentation.

Conclusions:

• The proposed multiPI-TransBTS framework offers a significant advancement in automated brain tumor segmentation.
• Integrating multi-physical information effectively addresses the challenges posed by brain tumor heterogeneity.
• This approach holds potential for improving clinical outcomes in brain tumor patient management.