Denoising for fMRI pattern analysis


I’m looking for some input on denoising task fMRI data, particularly for improving the reliability of multi-voxel patterns. I’ve seen several resources and suggestions for denoising for resting-state connectivity analyses but I’m struggling to identify current best practices for task fMRi data.

I’m aware of Charest et al. (2018) that evaluates the effect of denoising with glmdenoise on pattern analysis. Is there any other work that improves on this approach, or are there other approaches that people have tried?

Any leads or advice would be very helpful!


1 Like

There is this paper that came out recently

Mascali, D, Moraschi, M, DiNuzzo, M, et al. Evaluation of denoising strategies for task‐based functional connectivity: Equalizing residual motion artifacts between rest and cognitively demanding tasks. Hum Brain Mapp. 2020; 1– 24. doi: 10.1002/hbm.25332

But I don’t think they look at MVPA workflows.
Also I don’t know to what degree the conclusions of this type of benchmark study generalize across implementation packages, so the relevance may depend on what software you are using.


Thanks for pointing out the Mascali et al. (2020) paper.


Hi, I am interested by this topic as well.

I found this bioRXiv paper that implemented high-end acquisition and processing strategies that uses in particular an improved version of glmdenoise to improve the beta estimation.

A massive 7T fMRI dataset to bridge cognitive and computational neuroscience
Allen et al. (2021)

There is also a DNN denoising strategy targeting specifically task fmri but they do not look in this paper at MVPA workflows:
A robust deep neural network for denoising task-based fMRI data: An application to working memory and episodic memory
Yang et al, Medical Image Analysis, (2020)


Maybe an old, classic, cheap solution: compcor.

This is implemented in Nilearn, and I warmly recommend it: