Greedy low-rank algorithm for spatial connectome regression.

Area of Science:

• Computational Neuroscience
• Neuroimaging Analysis
• Matrix Computations

Background:

• Reconstructing brain connectivity from tract tracing data is a key computational challenge.
• Existing methods for mesoscopic connectome reconstruction do not scale to whole-brain analyses.
• Large-scale, ill-conditioned matrix equations hinder whole-brain connectome reconstruction.

Purpose of the Study:

• To develop a scalable computational method for whole-brain connectome reconstruction.
• To address the limitations of existing techniques in handling large-scale, complex matrix equations.
• To enable more accurate and efficient mapping of brain connectivity.

Main Methods:

• A greedy low-rank algorithm approximating the solution via rank-one updates.
• Exploitation of sparse and positive definite problem structure.
• Development of judicious stopping criteria and efficient solvers for sub-problems, including a GPU implementation.

Main Results:

• The proposed algorithm significantly outperforms previous methods in speed for connectome reconstruction.
• Moderate ranks provide a good approximation of the whole-brain connectome.
• The GPU implementation effectively addresses the computational bottleneck for large datasets.

Conclusions:

• The greedy low-rank algorithm offers a scalable solution for whole-brain connectome reconstruction.
• This computational advance facilitates the estimation of increasingly large-scale connectomes.
• The method enables more efficient analysis of brain connectivity data from tracing experiments.