Skip to content

Functional MRI derivatives

This section describes files deriving from functional MRI (fMRI) recordings, namely BOLD and CBV series.

Functional derivatives maps

A derivative map is a measure derived from a functional series, mapped onto spatial locations as defined by voxels in a volume or vertices on a surface.



for example:


Suffixes take the form <source_suffix>map:

Original suffix Derivative map suffix
bold boldmap
cbv cbvmap

The following table lists allowed labels for the stat entity and their corresponding measures:

stat-<label> Measure
mean Mean across the temporal/4th dimension of the data
std Standard deviation across the temporal/4th dimension of the data
tsnr Temporal SNR (mean / std)
sfs Signal fluctuation sensitivity
alff Amplitude low frequency fluctuations
falff Fractional amplitude of low frequency fluctuations
reho Regional homogeneity (voxelwise only)
dcb, dcw Voxelwise degree centrality binary and weighted
ecb, ecw Voxelwise eigenvector centrality binary and weighted
lfcdb, lfcdw Local functional connectivity density
vmhc Voxel mirrored homotopic connectivity

Sidecar JSON (*_boldmap.json/*_cbvmap.json)

The following metadata JSON fields are valid for derivative maps:

Key name Requirement Level Data type Description
SoftwareFilters OPTIONAL, but REQUIRED for stat-alff and stat-falff object of objects or "n/a" Object of temporal software filters applied, or "n/a" if the data is not available. Each key-value pair in the JSON object is a name of the filter and an object in which its parameters are defined as key-value pairs (for example, {"Anti-aliasing filter": {"half-amplitude cutoff (Hz)": 500, "Roll-off": "6dB/Octave"}}).
Neighborhood OPTIONAL, but REQUIRED for stat-reho string A description of the neighborhood used for calculating regional measures.
Threshold OPTIONAL, but REQUIRED for stat-{dcb,dcw,ecb,ecw} number or string String or number describing a threshold used to calculate the measure contained in a file.
Method OPTIONAL, but REQUIRED for stat-{dcb,dcw,ecb,ecw} string Description of method used to calculate the measure contained in a file.

Time series

A time series is a chronologically ordered series of numeric values. Time series will generally be stored as tables, with a row of column headers indicating the name of the series. In the case where the time series varies spatially, the data SHOULD be stored in an appropriate file, such as a 4D NIfTI file, a .func.gii GIFTI file, or a .dtseries.nii CIFTI-2 file.

All time series files MUST be accompanied by a data dictionary in JSON format, consistent with the format described in Common principles, which describes metadata for each column name. In the case of spatial time series files, the notion of column name does not apply, so column-level metadata may be applied to the entire file. In addition, the following fields apply to the entire file in all cases:

Field name Definition
SamplingFrequency REQUIRED. Sampling frequency (in Hz) of all columns in the file. Special value "TR" indicates one sample per volume of a corresponding BOLD series.
StartTime OPTIONAL. Start time (in seconds) in relation to the start of acquisition of the first volume in the corresponding imaging file (negative values are allowed).

Note that there are several differences with these fields in Physiological and other continuous recordings. The "TR" sampling frequency serves to indicate that no resampling is needed to use the series as a regressor for BOLD data, including BOLD series with non-uniform sampling, such as clustered sparse acquisition. StartTime is assumed to be 0 and therefore OPTIONAL. Additionally, because time series TSV files have column headers, the Columns field is omitted.

General time series



for example:


Any time series with common timing information may be stored in a timeseries file. If a time series is specified in another sub-section, then any additional required metadata MUST be stored in the JSON description of the column.

Column names

Column names are unique alphanumeric values that are defined in a relevant JSON sidecar file. For a series of related columns, it is RECOMMENDED to use zero-based indexing as a suffix. For example, the first six aCompCor components may be named a_comp_cor_00 ... a_comp_cor_05. Custom column names must not conflict with reserved names specified in this document.

Reserved names are specified in the remaining subsections.

Time series transformations

Any column may have the following suffixes appended to indicate transformations applied to the original time series data:

Transformation Description
_shift_back The time series has been lagged by one TR. x_shift_back[i] = x[i-1]
_dt The discrete first derivative of the time series. x_dt[i] = x[i+1] - x[i]
_sq The square of the time series
_var_norm The time series has been variance normalized
_centered The time series has had its mean subtracted

For example, rot_z_shift_back_sq means the square of the lagged version of the Z rotation (see Motion-related time series).

ROI-based time series extraction



for example:


ROI-based time series will generally be stored as tables, with a row of column headers indicating the name of the time series. In the case where every voxel has a time series (that is, voxel-wise regressors, as in ANATICOR), then the time series should be saved as a NIfTI file.

Column metadata special fields

Atlas (optional) - A label indicating an atlas that defines a region or set of regions in the volume. An atlas may be three- or four-dimensional, which affects the interpretation of the roi index.

ROI (optional) - For 3D atlases, the ROI label should be numeric, corresponding to the value of the voxels in the ROI. For 4D atlases, the ROI label should be numeric, corresponding to the volume index containing the ROI mask. The following special labels correspond to common ROIs that may be defined by many atlases or segmentation algorithms:

ROI Value Description
WhiteMatter Signal derived from white matter ROI.
CSF Signal derived from cerebro-spinal fluid ROI.
Background Signal derived from background (out of brain) ROI.
GrayMatter Signal derived from gray matter ROI.
Ventricles Signal derived from ventricles ROI.
CircleOfWillis Signal derived from circle of Willis ROI.
GlobalSignal Vector of mean values within the brain mask. Can be used for Global Signal Regression.

Column names

Column names are unique alphanumeric values that are defined in a relevant JSON sidecar file. A naming convention might be to concatenate the atlas, ROI index, summarization method (defined below) and transformations (see Time series transformations) using snake case. For example, harvard_oxford_cortical_4_pc could indicate ROI 4 of the Harvard-Oxford cortical atlas, summarized by taking the first principal component.

Summarization methods

To indicate the summarization method applied to construct a single time series for an ROI, the following column suffixes are defined:

Column suffix Description of the transformation
_mean The mean of voxel time series
_median The median of voxel time series
_pc[_<i>] The ith eigenvariate from principal component analysis, where i is 0 indexed. If i is not specified, the first component is implied (that is, pc_0).
_spat_reg Spatial regression


white_matter_mean global_signal_mean ventricles_mean
12                98                 11
11                34                 53
54                34                 34
    "SamplingFrequency": "TR",
    "white_matter_mean": {
        "ROI": "WhiteMatter"
    "global_signal_mean": {
        "ROI": "GlobalSignal"
    "ventricles_mean": {
        "ROI": "Ventricles"



Column names

The six basic motion parameters derived from motion correction have the following names and units:

Column name Units Description
trans_x, trans_y, trans_z mm Translation parameters
rot_x, rot_y, rot_z radians Rotation parameters

Transformations (see Time series transformations) of motion parameters may be included in the same file. For example, rot_z_shift_back_sq means square of the lagged version of Z rotation.

The following columns indicate summarized motion, as defined in the corresponding references:

Column name Units Description
framewise_displacement mm Framewise displacement (Power, et al., 2012)
rmsd mm Root mean square deviation (Jenkinson, 1999)
rms mm Root mean square of translation parameters (Van Dijk, et al., 2012)

Temporal outlier masks



Outlier masks are columns of zeros (0), with ones (1) indicating volumes that have been identified as outliers by some method.

Column names

Column names for outliers correspond to the type of outlier or method for identifying the outlier. Column names should take the form method[XX], where XX is an optional index.

For example, if a BOLD series included two initial dummy scans, non_steady_state may be a column with a 1 for the first two volumes, or non_steady_state_00 and non_steady_state_01 may be columns with a 1 in the first and second position, respectively.

The following methods are defined as reserved words:

Column name Description
non_steady_state_<x> Initial non-steady-state volumes. One column per volume.

Other time series

Time series that are not otherwise specified MUST be placed in a _timeseries.tsv file (see General time series).

Column names

The following time series are defined as reserved words:

Column name Description
dvars Change in variance (Brett, 2006, Power, et al., 2012)
std_dvars Standardized DVARS (Nichols, 2013)
cosine_<X> Discrete cosine basis vectors
legendre_<X> Legendre polynomial basis vectors

Spatiotemporal decompositions



for example:


Spatiotemporal decompositions produce either spatial or temporal components, along with conjugate mixing matrices that can be represented as time series and spatial maps, respectively. The combination of suffix and extension indicates which class of algorithm produced the outputs.

Mixing matrices may be omitted for algorithms that are temporal decompositions with no spatial component, such as CompCor variants.

Spatiotemporal decomposition files MUST be accompanied by a data dictionary in JSON format, consistent with Raw-BIDS section 4.2, which describes metadata for each column name. In addition to column names, the following MANDATORY field is added:

Key name Description
Method REQUIRED. Algorithm name and (if applicable) version.

Column names

Column names in spatiotemporal decompositions take the form of <decomposition>_<index>, where <decomposition> is the name of the decomposition algorithm, and <index> is a numeric identifier. Indices SHOULD start at 0. If there is a natural ordering, indices should reflect that ordering. For example, the aCompCor component which explains the most variance should be named a_comp_cor_00.

The following reserved words indicate common algorithms:

Column name Description
[a\|t\|w\|c]_comp_cor_<x> CompCor (Behzadi, et al., 2007) calculated with voxels chosen based on: a: anatomically derived ROIs (white matter and CSF), t: temporal variance, w: white matter voxels only, c: CSF voxels only
melodic_<x> Columns from the mixing matrix in FSL MELODIC