RapidParc: A global-context transformer for parallel, accurate, and lesion-robust tractogram parcellation

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Diffusion MRI tractography generates tractograms representing white matter architecture.
• Tractogram parcellation is crucial for structural connectivity studies and clinical applications.
• Registration-free methods offer advantages over traditional registration-based approaches, especially for pathological cases.

Purpose of the Study:

• Introduce RapidParc, a novel transformer-based method for registration-free tractogram parcellation.
• Evaluate RapidParc's speed, accuracy, and generalization capabilities compared to state-of-the-art methods.
• Investigate factors contributing to improved generalization in registration-free parcellation.

Main Methods:

• Developed RapidParc, a transformer-based approach treating streamlines as parallel tokens for global context.
• Compared RapidParc against TractCloud on diffusion MRI data from healthy controls and patients with epilepsy (post-hemispherotomy, sAH).
• Employed a novel augmentation strategy to enhance model robustness and generalization.

Main Results:

• RapidParc demonstrated a two-orders-of-magnitude speed improvement over TractCloud with comparable or slightly higher accuracy.
• RapidParc exhibited superior generalization to structurally altered brains, even when trained solely on healthy control data.
• Robust tractogram centering was identified as a key factor for improved generalization.

Conclusions:

• RapidParc offers a highly efficient and accurate solution for registration-free tractogram parcellation.
• The method's robustness to anatomical alterations makes it suitable for clinical applications involving diverse patient populations.
• Future registration-free parcellation methods can benefit from focusing on robust streamline centering rather than complex local context construction.