Uncertainty-guided transformer for brain tumor segmentation.

Area of Science:

• Medical image analysis
• Artificial intelligence in medicine
• Neuroimaging

Background:

• Multi-modal data enhances brain tumor segmentation but can introduce redundant information.
• Tumor boundaries and shapes present challenges for accurate segmentation quality estimation.

Purpose of the Study:

• To develop a robust brain tumor segmentation method using multi-modal data.
• To address challenges posed by redundant information and ambiguous tumor characteristics.

Main Methods:

• Exploited an uncertainty-guided U-shaped transformer with multiple heads.
• Constructed drop-out format masks (boundary, prior probability, conditional probability) for focused training on uncertainty regions.

Main Results:

• Achieved comparable or higher results than state-of-the-art methods on the BraTS2021 dataset.
• Obtained an average Dice coefficient of [Formula: see text] and a Hausdorff distance of 4.91.

Conclusions:

• The proposed uncertainty-guided transformer effectively handles multi-modal data for improved brain tumor segmentation.
• The method demonstrates robustness in segmenting tumors with ambiguous boundaries and irregular shapes.