BiTr-Unet: a CNN-Transformer Combined Network for MRI Brain Tumor Segmentation.

Area of Science:

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

Background:

• Convolutional Neural Networks (CNNs) excel at medical image segmentation.
• Vision Transformers show promise in 2D image classification.
• Transformers offer advantages in capturing long-range dependencies via self-attention.

Purpose of the Study:

• To develop an advanced model for brain tumor segmentation on multi-modal MRI scans.
• To leverage the strengths of both CNNs and Transformers for improved segmentation accuracy.
• To introduce the novel BiTr-Unet architecture for this specific application.

Main Methods:

• Proposed a hybrid CNN-Transformer model named BiTr-Unet.
• Modified the architecture to effectively process multi-modal MRI data.
• Evaluated the model on the BraTS2021 dataset for brain tumor segmentation.

Main Results:

• Achieved high median Dice scores on the BraTS2021 validation set: 0.9335 (whole tumor), 0.9304 (tumor core), 0.8899 (enhancing tumor).
• Obtained competitive median Hausdorff distances on the validation set: 2.8284, 2.2361, 1.4142 respectively.
• Demonstrated strong performance on the BraTS2021 testing dataset with Dice scores of 0.9257, 0.9350, and 0.8874.

Conclusions:

• The BiTr-Unet model effectively segments brain tumors from multi-modal MRI scans.
• The hybrid CNN-Transformer approach enhances the extraction of both local and global image features.
• The model's performance on the BraTS2021 dataset validates its potential for clinical applications in neuro-oncology.