3D TractFormer: 3D Direct Volumetric White Matter Tract Segmentation with Hybrid Channel-Wise Transformer

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Accurate white matter tract segmentation in diffusion-weighted magnetic resonance imaging (dMRI) is crucial for brain health analysis.
• Current 2D slice-based methods overlook 3D contextual information and face challenges with volumetric data complexity and boundary handling.

Purpose of the Study:

• To propose an efficient and accurate 3D direct volumetric segmentation method for white matter tracts in dMRI.
• To overcome limitations of existing 2D methods by leveraging 3D contextual information and addressing data complexity.

Main Methods:

• A novel U-shaped network architecture deeply interleaving convolutional and transformer blocks for integrated spatial and global feature extraction.
• A channel-wise transformer incorporating depth-wise separable convolution and attention mechanisms to manage 4D data challenges.
• Training a symmetric network with volumetric patches to address limited 3D training data and reduce computational costs.

Main Results:

• The proposed 3D method demonstrates superior performance in white matter tract segmentation compared to state-of-the-art techniques.
• Experimental validation on a large tractogram dataset confirms the method's effectiveness and efficiency.

Conclusions:

• The developed 3D direct volumetric segmentation method offers a significant advancement for dMRI analysis.
• This approach provides a more robust and computationally efficient solution for segmenting white matter tracts, aiding brain health research.