Accelerating permutation testing in voxel-wise analysis through subspace tracking: A new plugin for SnPM.

Area of Science:

• Neuroimaging
• Computational Statistics
• Non-parametric methods

Background:

• Permutation testing is crucial for calculating corrected p-values in neuroimaging, ensuring strong control of false positives.
• The computational demands of traditional permutation testing can be prohibitive for large neuroimaging datasets.

Purpose of the Study:

• To develop a computationally efficient algorithm for permutation testing in neuroimaging.
• To reduce the runtime of permutation testing while preserving the accuracy of the max null distribution.

Main Methods:

• Viewing permutation testing as constructing a large permutation testing matrix (T).
• Exploiting the low-rank plus low-variance residual structure of T.
• Applying low-rank matrix completion techniques to estimate the max null distribution.
• Introducing the RapidPT algorithm for efficient voxel-wise analysis.

Main Results:

• RapidPT achieves significant speedups compared to Statistical NonParametric Mapping (SnPM13) and NaivePT.
• Speedups range from 1.5x-1000x depending on dataset size and baseline.
• Optimal performance is observed on medium-sized datasets (50≤n≤200).
• RapidPT demonstrates substantial gains on larger datasets (n≥200) and when numerous permutations are required.

Conclusions:

• RapidPT offers a computationally efficient alternative for permutation testing in neuroimaging.
• The algorithm effectively estimates the max null distribution, enabling faster voxel-wise analyses.
• RapidPT can be used as a standalone toolbox or integrated with existing software like SnPM13.