Benchmarking local low rank denoising methods for task-based fMRI data analysis

Local-low-rank denoising in task-based fMRI increases sensitivity to statistical detection of neural activity, without harming specificity. We compared 5 methods (NORDIC, MP-PCA, Hybrid-PCA, Optimal-Threshold, Hybrid-OT) in 4 preprocessing configuration (denoising on magnitude/complex data, before/after motion correction) and their effect on downstream analysis. For best performance, the denoising shoud be done prior to motion correction, and using complex-valued data is only valuable in some settings. In average (n=6), up to 8 times more activations can be detected (p < 0.05, controlling for FDR). We also provide open-source implementations for broader use of Local-Low-Rank denoising methods in fMRI.