Robust Bayesian brain extraction by integrating structural subspace-based spatial prior into deep neural networks.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Accurate brain extraction (skull stripping) is critical for neuroimaging research.
• Brain image data heterogeneity poses a significant challenge for robust segmentation.
• Existing methods struggle to capture complex spatial-intensity distributions.

Purpose of the Study:

• To develop a novel Bayesian brain extraction method.
• To enhance accuracy and robustness in segmenting brain images.
• To address data heterogeneity in brain imaging across diverse populations and conditions.

Main Methods:

• Integration of a structural subspace-based prior (mixture-of-eigenmodes) with deep learning classification.
• Utilizing a structural subspace model for global spatial-structural distributions.
• Employing a multi-resolution, position-dependent neural network for local spatial-intensity distributions.
• A patch-based fusion network to combine global and local features.

Main Results:

• Demonstrated superior segmentation accuracy and robustness compared to state-of-the-art methods.
• Successfully evaluated on multi-institutional datasets, including healthy scans, images with lesions, noise, and artifacts.
• The proposed method effectively captures high-dimensional spatial-intensity distributions.

Conclusions:

• The novel Bayesian approach offers accurate and robust brain extraction.
• The method shows promise for improving practical clinical applications in neuroimaging.
• Addresses limitations of current techniques in handling data heterogeneity.