fmri-analysis

Overview

fmri-analysis covers the standard fMRI pipeline: preprocessing (motion correction, distortion correction, normalization, smoothing), first-level modeling with the general linear model (GLM), and group-level inference. It also covers the quality-control culture that fMRI requires — head motion alone can manufacture "activations," and the field's replication history demands conservative statistics.

The focus is task fMRI with the GLM; resting-state connectivity gets its own coverage in resting-state-fmri, and preprocessing details live in mri-preprocessing.

When to use

• Designing a task fMRI experiment: block vs event-related designs, jittering, trial timing, run length.
• First-level GLM: building design matrices, HRF convolution, contrasts.
• Group analysis: one-sample/two-sample t-tests, mixed-effects (FLAME), covariates.
• Multiple-comparisons correction: cluster correction, FWE, FDR, TFCE, permutation (randomise).
• ROI analysis: anatomical vs functional ROIs, percent signal change extraction.
• MVPA/decoding: searchlight and ROI-based multivariate pattern analysis.
• Reporting: what parameters, thresholds, and QC metrics a methods section needs.

Core concepts

• BOLD is indirect. The signal reflects blood oxygenation, not spikes — delayed ~4-6 s, sluggish, and confounded by vasculature, motion, and physiology. Effect sizes are small (often <1% signal change); design accordingly.
• The hemodynamic response function. Model neural events convolved with a canonical HRF (double-gamma). Add temporal/dispersion derivatives to absorb HRF variability. Misspecified timing (wrong onsets, unmodeled durations) is the commonest GLM error.
• Design efficiency. Block designs maximize detection power; event-related designs (with jittered ISIs) allow estimating the HRF shape and separating close events. Efficiency depends on the contrast — optimize the design for the contrast you care about.
• Nuisance regressors. Six motion parameters plus derivatives, CSF/white-matter signals (CompCor), physiological regressors (RETROICOR) when available, and spike regressors for high-motion volumes. Motion correlates with task in many paradigms — unmodeled, it becomes "activation."
• Preprocessing order matters. Slice-timing → motion correction → distortion correction (fieldmaps/TOPUP) → coregistration → normalization → smoothing. Each step's errors propagate; check registrations visually for every subject.
• Smoothing trade-off. 6-8 mm FWHM boosts SNR for univariate GLM but blurs fine patterns — skip or minimize smoothing for MVPA. Match smoothing to the analysis, not to habit.
• Multiple comparisons. ~100k+ voxels tested. Valid options: voxel-wise FWE, cluster-level correction with proper cluster-forming thresholds (p<0.001, not p<0.01 — see Eklund et al.), TFCE, or permutation testing. Uncorrected p<0.001 with k>10 is not a correction.
• Circular analysis. Defining ROIs from the same contrast you then test inflates everything. Use independent localizers, anatomical atlases, or split-half data.

Practical workflow

1. Design. Choose block vs event-related for the question; jitter ISIs; keep runs <10 min to limit motion; pilot the timing outside the scanner.
2. QC raw data. Check motion (framewise displacement; flag runs with mean FD >0.5 mm or many spikes), coverage, ghosting, and signal dropout (orbitofrontal/temporal regions).
3. Preprocess. fMRIPrep is the recommended default — it handles the boilerplate reproducibly. Visually verify coregistration and normalization for every subject.
4. First-level GLM. Build the design matrix with correct onsets/durations, HRF convolution, nuisance regressors, and high-pass filtering (cutoff ~1/128 Hz or matched to design). Check design orthogonality and efficiency; inspect residual maps.
5. Contrasts. Task > baseline, condition A > B, parametric modulators. Keep contrasts few and preregistered; exploratory whole-brain contrasts get full correction.
6. Group analysis. Mixed-effects model (subjects as random effects), covariates for age/sex/ motion, proper multiple-comparisons correction. Report peak coordinates, cluster sizes, and effect sizes — not just "significant blobs."
7. Report. Preprocessing pipeline + versions, motion exclusion criteria, first-level model details, group model, correction method with thresholds, and unthresholded maps (NeuroVault).

Example (nilearn sketch):

from nilearn.glm.first_level import FirstLevelModel
fm = FirstLevelModel(t_r=2.0, hrf_model="spm", drift_model="cosine",
                     smoothing_fwhm=6, mask_img=mask)
fm.fit(run_imgs, events=events_df, confounds=confounds_df)
z_map = fm.compute_contrast("task-baseline", output_type="z_score")

Common pitfalls

• Cluster correction with lenient cluster-forming thresholds (the Eklund problem).
• Motion-task correlation passed off as activation (check motion regressors' correlation with design).
• Double-dipping ROIs: selecting voxels by the effect you then quantify.
• Interpreting BOLD as direct neural activity or as "the brain lighting up."
• Reverse inference: "amygdala active, therefore fear" — regions do many things.
• Smoothing before MVPA, destroying the patterns you want to decode.
• Small samples (n<20) with flexible analyses — the replication-crisis recipe.