Filtering in tractography using autoencoders (FINTA).

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Diffusion MRI tractography generates streamlines to map white matter pathways.
• Current methods face challenges like anatomical inaccuracy and signal under-representation.

Purpose of the Study:

• To introduce a novel autoencoder-based method, FINTA, for filtering streamlines in diffusion MRI tractography.
• To enhance the reliability and accuracy of tractograms.

Main Methods:

• FINTA utilizes raw, unlabeled tractograms to train an autoencoder for learning streamline representations.
• A nearest neighbor algorithm filters streamlines based on the learned embedding.
• The method was tested on synthetic and in vivo human brain data.

Main Results:

• FINTA achieved over 90% accuracy on the test set.
• Demonstrated superior filtering performance compared to conventional and state-of-the-art methods like RecoBundles.
• Showed applicability to partial tractograms and generalization across different tracking methods and datasets.

Conclusions:

• FINTA offers a flexible and powerful deep learning framework for white matter streamline filtering.
• The method significantly reduces computation time for large tractograms.
• Enhances tractometry and connectivity analyses by providing more reliable tractograms.