GPU-accelerated connectome discovery at scale.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biophysics

Background:

• Diffusion MRI and tractography estimate in vivo human brain connectivity.
• Tractography algorithms vary, yielding inconsistent results without validation.
• Streamline pruning improves accuracy but is computationally intensive, limiting big-data applications.

Purpose of the Study:

• Introduce ReAl-LiFE, a GPU-accelerated streamline pruning algorithm for enhanced brain connectome estimation.
• Overcome computational limitations of existing methods for large-scale neuroimaging.
• Improve accuracy, reliability, and scalability of in vivo brain connectivity mapping.

Main Methods:

• Developed a GPU-based implementation of the LiFE (Linear Fascicle Evaluation) streamline pruning algorithm, named ReAl-LiFE.
• Achieved >100x speedup compared to previous CPU-based implementations.
• Applied ReAl-LiFE to generate sparser, more accurate, and highly reliable brain connectomes.

Main Results:

• ReAl-LiFE demonstrated superior test-retest reliability in estimating brain connections.
• The method outperformed competing tractography approaches in accuracy and efficiency.
• Connectome features derived from ReAl-LiFE predicted inter-individual cognitive variations.

Conclusions:

• ReAl-LiFE offers a significant advancement for accurate, large-scale, individualized brain connectome discovery.
• The GPU-accelerated approach overcomes previous computational bottlenecks in streamline pruning.
• The underlying non-negative least-squares optimization is broadly applicable to other computational problems.