Modular preprocessing pipelines can reintroduce artifacts into fMRI data.

Area of Science:

• Neuroimaging
• Data Analysis
• Signal Processing

Background:

• Functional magnetic resonance imaging (fMRI) preprocessing involves modular, sequential steps like motion regression and temporal filtering.
• Current pipelines offer flexibility in the order of these preprocessing steps.

Purpose of the Study:

• To demonstrate how the sequential application of preprocessing steps in fMRI can inadvertently reintroduce artifacts.
• To highlight the non-commutative nature of linear filtering operations in fMRI data analysis.

Main Methods:

• Theoretical explanation of preprocessing steps as geometric projections onto subspaces.
• Empirical validation using a test-retest resting-state fMRI (rs-fMRI) dataset.
• Analysis of artifact reintroduction due to non-orthogonal projections.

Main Results:

• Sequential regression steps can lead to subspaces that are no longer orthogonal to previously removed nuisance covariates.
• This non-orthogonality results in the reintroduction of artifacts into the fMRI data.
• The order of operations significantly impacts the final processed data quality, particularly in rs-fMRI.

Conclusions:

• The order of preprocessing steps is critical in fMRI analysis, contrary to the flexible approach often taken.
• Remedies include combining all steps into a single linear filter or performing sequential orthogonalization.
• Addressing this issue is crucial for accurate interpretation of rs-fMRI findings.