Metric definitions

Precise definitions of every quantity RABIES computes and reports.

Nuisance regressors

Selected with --nuisance_regressors at the confound correction stage.

mot_6

3 rotations (Euler angles in radians) and 3 translations (in mm) measured from the rigid-body head realignment algorithm.

mot_24

The 6 motion parameters together with their temporal derivatives, plus 12 additional parameters obtained by taking the squared terms — the Friston 24 parameters [FWH+96]:

\[ mot24_t = [mot6_t,(mot6_t-mot6_{t-1}),(mot6_t)^2,(mot6_t-mot6_{t-1})^2] \]

with \(mot24_t\) representing the list of 24 regressors for timepoint \(t\).

WM/CSF/vascular/global signal

The mean signal computed within the corresponding brain mask (WM, CSF, vascular or whole-brain).

aCompCor_percent

Principal component timecourses derived from timeseries within the combined WM and CSF masks — the aCompCor technique [MNC+14]. From the timeseries within the WM/CSF masks \(Y_{WM/CSF}\), a principal component analysis (PCA) decomposition is conducted to derive

\[ Y_{WM/CSF} = W_{aCompCor}C^T \]

with \(C\) a set of spatial principal components and \(W\) their associated loadings across time. The first components explaining 50% of the variance are kept, and their loadings \(W_{aCompCor}\) provide the aCompCor nuisance regressors. The PCA is conducted on partially cleaned timeseries after step 4 of the confound correction pipeline (see The confound correction workflow).

aCompCor_5

As aCompCor_percent, but the first 5 components are kept instead of a set explaining 50% of the variance.

Temporal scan diagnosis

Head motion translation and rotation parameters: 3 rotations (Euler angles in radians) and 3 translations (in mm) measured from the rigid-body head realignment algorithm.

Framewise displacement

For each timepoint, this is the displacement between the current and the previous frame, averaged over all brain voxels. For each brain voxel within the referential space for head realignment (i.e. the 3D EPI provided as reference for realignment) and for each timepoint, the inverse transform of the head motion parameters from the corresponding timepoint is applied to obtain the voxel position pre-motion correction. Framewise displacement is then computed for each voxel as the Euclidean distance between the pre-motion-correction positions for the current and previous timepoints. The mean framewise displacement \(FD_t\) at timepoint \(t\) is therefore

\[ FD_t = \frac{1}{n}\sum_{i=1}^{n}\sqrt{(x_{i,t}-x_{i,t-1})^2+(y_{i,t}-y_{i,t-1})^2+(z_{i,t}-z_{i,t-1})^2} \]

using the 3D \(x\), \(y\) and \(z\) spatial coordinates in mm for timepoints \(t\) and \(t-1\) and voxel indices \(i\). Framewise displacement for the first frame, which has no past timepoint, is set to 0.

DVARS

The estimation of temporal shifts in global signal at each timepoint, measured as the root-mean-square of the timeseries’ temporal derivative

\[ DVARS_t = \sqrt{\frac{1}{n}\sum_{i=1}^{n}(Y_{i,t}-Y_{i,t-1})^2} \]

where \(Y_{i,t}\) is the BOLD signal in brain voxel \(i\) at timepoint \(t\). The first timepoint is set to 0, having no previous timepoint.

Framewise distance from mean

The mean square error (MSE) between each frame and the average EPI, with the average computed as the tri-mean across time voxelwise. The input is the resampled timeseries output from the preprocess stage.

Whole-brain/Edge/WM/CSF mask

The mean signal across a given brain mask: whole-brain (the global signal), WM, CSF or brain edge.

CR_var

The variance estimated and removed by nuisance regression at each timepoint, computed as the root-mean square (RMS) of \(Y_{CR}\) across voxels, where \(Y_{CR}\) is the predicted confound timeseries.

CR R²

The proportion of variance explained, and removed, by nuisance regression. Obtained with \(CR_{R^2}= 1-\frac{Var(\hat{Y_t})}{Var(Y_t)}\) for a given frame at timepoint \(t\), where \(Y_t\) and \(\hat{Y_t}\) are the BOLD frame at timepoint \(t\) pre- and post-regression, and \(Var(x) = \frac{1}{n}\sum_{i=1}^{n}(x_i - \mu_x)^2\) is the variance with \(\mu\) the mean.

Mean amplitude of network VS confound sources

A set of timecourses averaged as \(\frac{1}{n}\sum_{i=1}^{n}|X_i|\), where \(X_i\) is timecourse \(i\). The timecourses correspond to one of:

DR confounds: timecourses from the first stage of dual regression, using the confound components provided to --prior_confound_idx
DR networks: network timecourses from the first stage of dual regression as specified with --prior_bold_idx
SBC networks: network timecourses derived from the seeds provided in --seed_list

Spatial scan diagnosis

BOLD_SD: The temporal standard deviation computed for each voxel from the cleaned BOLD timeseries.
CR_SD: The temporal standard deviation computed for each voxel from the predicted confound timeseries during confound regression, i.e. \(Y_{CR}\).
CR R²: The proportion of variance explained by confound regression at each voxel. Obtained with \(CR_{R^2}= 1-\frac{Var(\hat{Y_v})}{Var(Y_v)}\) for a given voxel \(v\), where \(Y_v\) and \(\hat{Y_v}\) are the timeseries pre- and post-regression for that voxel, and \(Var(x) = \frac{1}{n}\sum_{i=1}^{n}(x_i - \mu_x)^2\) is the variance of \(x\) with \(\mu\) the mean.
Global signal covariance (GS_cov): The covariance between the mean whole-brain signal (i.e. global signal) timecourse and the timeseries at each voxel.
DR network X: The linear coefficients resulting from the second regression with dual regression, corresponding to a network amplitude map, for the Xth network specified with --prior_bold_idx.
SBC network X: The voxelwise correlation coefficients (Pearson’s r) estimated with seed-based connectivity, for the Xth seed provided in --seed_list.

Distribution plot

Network amplitude: The overall network amplitude, summarised by computing the L2-norm across a network connectivity map from a subject-level analysis. Such a map can be derived from seed-based correlation, or correspond to the linear coefficients from the second regression \({\beta}_{SM}\) for dual regression.
Network specificity: The network map for a given analysis (seed-based or dual regression) and a given fMRI scan is compared relative to a reference network map to assess specificity. The reference map is either the original ICA component for dual regression, a manually provided input map with --seed_prior_list for seed-based analysis, or the dataset average itself if selecting --group_avg_prior. To compute specificity, both the reference and individual scan network maps are thresholded to retain the top X% of voxels with highest connectivity, X% being defined by --brainmap_percent_threshold, and the overlap of the thresholded area is computed using Dice overlap.

Important

If the reference network map was generated from the dataset average, it is important to validate that the average can indeed provide an adequate representation of the expected network connectivity. If the reference does not represent a network, then the network specificity metric is meaningless. The average network used for those computations can be visualised in the group statistical report.

Dual regression confound correlation

The timecourse for a single network, from a seed or from dual regression, is correlated with the timecourse from each confound component (provided using --prior_confound_idx) modelled through dual regression. The absolute mean correlation is then computed to obtain the average amplitude of confound correlations for that network analysis.

FD-DVARS corr.

For each scan, the correlation between the framewise displacement timecourse and the DVARS timecourse computed post-confound correction — this is not the DVARS plotted in the temporal diagnosis figure. Censored timeframes are excluded from both timecourses, and DVARS is recomputed after applying confound correction, so this metric represents residual associations between spontaneous motion and the cleaned global signal fluctuations.

Total \(CR_{SD}\)

The total standard deviation across the predicted confound timeseries \(Y_{CR}\).

Mean framewise displacement

The mean framewise displacement computed across time, including only frames remaining after the censoring applied for confound correction.

Temporal degrees of freedom

The degrees of freedom remaining after confound correction:

tDOF = Original number of timepoints
     - Number of censored timepoints
     - Number of AROMA components removed
     - Number of nuisance regressors

Group statistical QC report

These quantities are stored inside the analysis_QC/*_stats/*_QC_stats.csv output files.

Specificity of network variability: As with network specificity in the distribution plot, the network variability map and the corresponding reference network map are thresholded to include the top X% of voxels (X% defined by --brainmap_percent_threshold), and the overlap is estimated using Dice overlap.
Mean confound correlation: For each confound correlation map (\(CR_{SD}\), mean FD or tDOF), the mean is computed across voxels within the thresholded area of the reference network map, giving a mean correlation within the network’s core region.