Variational Autoencoders for Generating Synthetic Tractography-Based Bundle Templates in a Low-Data Setting.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Automatic segmentation of white matter tracts relies on standard atlases, which are time-consuming to create and may not represent diverse populations.
• Existing atlas-based methods face challenges in accurately segmenting complex white matter structures across different subject groups.

Approach:

• Utilized a deep generative model, specifically a Convolutional Variational Autoencoder, to map complex white matter streamlines into a low-dimensional latent space.
• Employed Kernel Density Estimation (KDE) on streamline embeddings to generate synthetic population-specific bundle templates from a limited sample size (50 subjects from ADNI3).

Key Points:

• Quantitative shape analysis revealed that the generated synthetic bundle templates better capture the shape distribution compared to traditional atlas data.
• The framework demonstrated successful direct bundle segmentation from whole-brain tractograms, highlighting its practical applicability.
• The approach addresses the limitations of standard atlases by enabling the creation of customized, population-specific templates.

Conclusions:

• The deep generative model approach provides an efficient and effective method for creating synthetic population-specific white matter bundle templates.
• This framework enhances the accuracy and applicability of white matter tract segmentation in neuroimaging studies.
• The study offers a promising solution for overcoming the challenges associated with traditional atlas-based segmentation methods.