brain-connectivity-analysis

Overview

brain-connectivity-analysis covers how brain regions connect and communicate: structural connectivity from diffusion MRI tractography, functional connectivity from correlated activity (fMRI, EEG/MEG), effective connectivity (causal modeling), and the graph-theory toolkit used to describe networks. The emphasis is on what each method can and cannot claim — connectivity neuroscience is full of measures that sound causal but aren't.

When to use

• Diffusion MRI: tensor/fODF modeling, tractography, along-tract statistics.
• Functional connectivity: seed-based correlation, ICA networks, parcellation-based connectomes.
• Graph theory: degree, clustering, efficiency, modularity, hubs, rich clubs.
• Effective connectivity: DCM, Granger causality, transfer entropy — when direction matters.
• Comparing networks across groups or conditions with proper statistics (NBS, permutation).
• Multimodal integration: constraining functional claims with structural connectivity.

Core concepts

• Structural connectivity (dMRI). Diffusion measures water displacement; fiber orientation distributions (constrained spherical deconvolution) resolve crossing fibers where the tensor model fails. Tractography reconstructs streamlines — probabilistic tracking quantifies uncertainty. Streamline count is not fiber count; treat it as a connectivity index, not anatomy.
• Tractography pitfalls. False positives are rampant (crossing/kissing fibers, gyral bias). Anatomically constrained tractography (ACT) and SIFT/SIFT2 filtering reduce but don't eliminate them. Validate major findings against known anatomy.
• Functional connectivity. Correlation of time series between regions. It measures statistical dependence, not communication — two regions can correlate via a common third input. Always state this limitation; never write "region A communicated with region B" from correlation alone.
• Parcellation choice. Connectome results depend heavily on the atlas (Schaefer, Glasser, AAL). Finer parcellations increase multiple comparisons; coarser ones blur boundaries. Report the atlas and test robustness to an alternative.
• Confounds. Head motion inflates short-range and attenuates long-range connectivity; global signal regression removes widespread noise but introduces negative correlations and can distort group differences. There is no neutral choice — report both or justify one.
• Graph metrics. Degree/hubs (influential nodes), clustering coefficient (local segregation), path length/efficiency (integration), modularity (community structure), rich club (hub interconnectivity). Normalize against random null networks — raw values are meaningless without comparison.
• Effective connectivity. DCM tests specific mechanistic hypotheses about directed influence (model comparison, not exploratory search); Granger/transfer entropy give directed functional measures but assume stationarity and are confounded by hemodynamic variability in fMRI.
• Network-based statistics (NBS). For group comparisons of connectomes: permutation-based, controls family-wise error over edges, more powerful than edge-wise FDR for connected effects.

Practical workflow

1. Preprocess carefully. Connectivity amplifies preprocessing choices — motion, GSR, and parcellation decisions change results more than in activation studies.
2. Build connectomes. Structural: tractography + SIFT2 → streamline-weighted matrices. Functional: parcellate → extract time series → correlation (Fisher-z) → threshold or keep weighted.
3. QC. Check motion-connectivity correlations (QC-FC); verify tractography against anatomy; confirm networks replicate across sessions (test-retest reliability is often modest — report it).
4. Describe. Graph metrics vs null models; community detection; hub identification.
5. Compare. NBS or permutation-based edge/group tests with proper correction. Preregister the metric of interest — the graph-metric garden of forking paths is large.
6. Interpret conservatively. Correlation ≠ communication; streamline count ≠ axon count; group differences in connectivity need motion and demographic matching to be believable.

Example (Python sketch):

from nilearn.connectome import ConnectivityMeasure
corr = ConnectivityMeasure(kind="correlation").fit_transform([time_series])[0]
# graph metrics
import bct
deg = bct.degrees_und((corr > 0.3).astype(int))

Common pitfalls

• Causal language for correlational connectivity.
• Global signal regression applied (or skipped) without acknowledging the trade-off.
• Motion artifacts presented as network differences between groups.
• Uncorrected edge-wise tests across thousands of connections.
• Treating streamline counts as quantitative anatomy.
• Parcellation shopping until the network "looks right."
• Ignoring test-retest reliability — many connectivity metrics are noisy.