Dataset curation#

Converting data to BIDS#

All git-annex datasets should be BIDS-compliant. For more information about the BIDS standard, please visit For some examples of BIDS datasets, visit this page. A quick way to verify compliance with the convention is this online BIDS validator.

When you receive raw data from an external collaborator, save them under a temporary location on one of NeuroPoly’s server, e.g.: duke/temp.

Then, inspect the data and convert them to BIDS. It is recommended to write a script that does the conversion. The script should then be saved under the code folder of the final dataset. Some previous scripts can be found on GitHub or under the code folder of already existing datasets.


Once the data are converted to BIDS and uploaded to git-annex repository, please delete the temporary folder.

Building the raw dataset#

The raw dataset corresponds to the core dataset that contains all the different acquisitions generated for one or several subjects. NO postprocessing steps should be applied to these acquisitions.

Subjects folders in the raw dataset are structured as follows for MRI, with folders corresponding to subjects, [sessions] and MRI modalities:

Raw structure#

Useful BIDS specifications are:

The example below applies for MRI data:

β”œβ”€β”€ dataset_description.json
β”œβ”€β”€ participants.tsv
β”œβ”€β”€ participants.json
β”œβ”€β”€ code/
β”‚   └──
β”œβ”€β”€ sub-<label>/
β”‚   └── [ses-<label>/]
β”‚        β”œβ”€β”€ anat/
β”‚        β”‚   β”œβ”€β”€ sub-<label>[_ses-<label>][_acq-<label>][_ce-<label>][_rec-<label>][_run-<index>][_part-<mag|phase|real|imag>]_<suffix>.json
β”‚        β”‚   └── sub-<label>[_ses-<label>][_acq-<label>][_ce-<label>][_rec-<label>][_run-<index>][_part-<mag|phase|real|imag>]_<suffix>.nii[.gz]
β”‚        β”œβ”€β”€ fmap/
β”‚        β”œβ”€β”€ func/
β”‚        └── dwi/
β”‚            β”œβ”€β”€ sub-<label>[_ses-<label>][_acq-<label>][_rec-<label>][_dir-<label>][_run-<index>][_part-<mag|phase|real|imag>]_dwi.bval
β”‚            β”œβ”€β”€ sub-<label>[_ses-<label>][_acq-<label>][_rec-<label>][_dir-<label>][_run-<index>][_part-<mag|phase|real|imag>]_dwi.bvec
β”‚            β”œβ”€β”€ sub-<label>[_ses-<label>][_acq-<label>][_rec-<label>][_dir-<label>][_run-<index>][_part-<mag|phase|real|imag>]_dwi.json
β”‚            └── sub-<label>[_ses-<label>][_acq-<label>][_rec-<label>][_dir-<label>][_run-<index>][_part-<mag|phase|real|imag>]_dwi.nii[.gz]


[Brackets] are characterizing optional informations

Subject naming convention#

Basic convention: sub-XXX



Multi-institution/Multi-pathology convention: sub-\<site>\<pathology>XXX

Example of Multi-institution dataset:

sub-mon001      # mon stands for Montreal
sub-tor001      # tor stands for Toronto

Example of Multi-institution/Multi-pathology dataset:

In the case of multi-pathology dataset (two or more distinct diseases + healthy controls), it is convenient to include also pathology to the subjectID, for example:

sub-torDCM001      # tor stands for Toronto and DCM stands for Degenerative Cervical Myelopathy
sub-torHC001       # tor stands for Toronto and HC stands for Healthy Controls
sub-zurSCI001      # zur stands for Zurich and SCI stands for Spinal Cord Injury

Raw entities#

Characterized by a key word (sub, ses, acq, etc.) and a value (label = an alphanumeric value, index = a nonnegative integer, etc) separated with a dash -

  • sub-<label>

  • [ses-<label>]

  • [acq-<label>]

  • [ce-<label>]

  • [rec-<label>]

  • [run-<index>]

  • [part-<mag|phase|real|imag>]

  • [dir-<label>]

Multiple entities can be used, but they must be separated using underscores _

Examples of special cases below:

  • If you need to differentiate spinal cord images from the brain within the same dataset, use the acq-cspine tag. For example, sub-001_acq-cspine_T1w.nii.gz. We opted for acq-cspine tag (see BIDS template) because bp-cspine is not currently supported by the BIDS convention (see BEP25 BIDS extension proposal).

  • If you need to differentiate between sequences acquired with different orientations, use the acq-ax, acq-cor, or acq-sag tag. For example, sub-001_acq-ax_T1w.nii.gz.

  • If you need to differentiate between different magnetization transfer (MT) sequences, use the flip-<index>_mt-<on|off> tag. For example, sub-001_flip-1_mt-on_MTS.nii.gz, sub-001_flip-1_mt-off_MTS.nii.gz or sub-001_flip-2_mt-off_MTS.nii.gz.


If you to combine several above mentioned tags, use camelCase. For example, sub-001_acq-cspineSag_T1w.nii.gz.

Raw suffixes#

An alphanumeric string located after all the entities following a final underscore _ (i.e. the <suffix>). This suffix corresponds for MRI to the MRI contrast:

  • T1w


  • dwi

  • etc.

Only ONE suffix can be used within the filename.

Raw extensions#

Files extensions:

  • .nii.gz

  • .json

  • .bval

  • etc.

The is a markdown file describing the dataset in more detail.

Please use the template below:


This is an <MRI/Microscopy> dataset acquired in the context of the <XYZ> project. 
<IF DATASET CONTAINS DERIVATIVES>It also contains <manual segmentation/labels> of <MS lesions/tumors/etc> from <one/two/or more> expert raters located under the derivatives folder.

## Contact Person

Dataset shared by: <NAME AND EMAIL>
<IF THERE IS A PRIMARY PROJECT/MODEL>Repository:<organization>/<repository_name>




The dataset_description.json is a JSON file describing the dataset.

Please use the dataset_description.json template below:

    "BIDSVersion": "1.9.0",
    "Name": "<dataset_name>",
    "DatasetType": "raw"


Refer to the BIDS spec to know what version to fill in here.


The participants.tsv is a Tab-separated value file that lists all subjects in the dataset with useful metadata. Please start off from the example below:

participant_id	source_id	species	age	sex	pathology	institution
sub-001	001	homo sapiens	30	F	HC	montreal
sub-002	005	homo sapiens	40	O	MS	montreal
sub-003	007	homo sapiens	n/a	n/a	MS	toronto

Additional notes:

  • Authorized values for pathology are listed under participants.json.

  • Indicate missing values with n/a (for β€œnot available”), not by empty cells!

  • In the example above, the apparent mismatch between β€˜pathology’ and the values is caused by the tabs

  • Other columns can be added if the metadata are relevant


The participants.json is a JSON file providing a legend for the columns in participants.tsv, with longer descriptions, units, and in the case of categorical variables, allowed levels. Please use the template below:

    "participant_id": {
        "Description": "Unique Participant ID",
        "LongName": "Participant ID"
    "source_id": {
        "Description": "Subject ID in the source unprocessed data",
        "LongName": "Subject ID in the source unprocessed data"
    "species": {
        "Description": "Binomial species name of participant",
        "LongName": "Species"
    "age": {
        "Description": "Participant age",
        "LongName": "Participant age",
        "Units": "years"
    "sex": {
        "Description": "sex of the participant as reported by the participant",
        "Levels": {
            "M": "male",
            "F": "female",
            "O": "other"
    "pathology": {
        "Description": "The diagnosis of pathology of the participant",
        "LongName": "Pathology name",
        "Levels": {
            "HC": "Healthy Control",
            "DCM": "Degenerative Cervical Myelopathy (synonymous with CSM - Cervical Spondylotic Myelopathy)",
            "MildCompression": "Asymptomatic cord compression, without myelopathy",
            "MS": "Multiple Sclerosis",
            "SCI": "Traumatic Spinal Cord Injury"
    "institution": {
        "Description": "Human-friendly institution name",
        "LongName": "BIDS Institution ID"
    "notes": {
        "Description": "Additional notes about the participant. For example, if there is more information about a disease, indicate it here.",
        "LongName": "Additional notes"


The data cleaning and curation script(s) that create the sub-XXX/ folders should be kept with them, under the code/ folder. Within reason, every dataset should have a script that when run like

python code/ path/to/sourcedata ./

unpacks, converts and renames all the images and related files in path/to/sourcedata/ into BIDS format in the current dataset ./.

This program should be committed first, before the curated data it produces. Afterwards, every commit that modifies the code should also re-run it, and the code and re-curated data should be committed in tandem.


Analysis scripts should not be kept here. Keep them in separate repositories, usually in public on GitHub, with instructions about.

Building the derivatives datasets#

First, it is important to understand what are BIDS derivatives folders:

Derivatives are outputs of common processing pipelines, capturing data and meta-data sufficient for a researcher to understand and (critically) reuse those outputs in subsequent processing. Standardizing derivatives is motivated by use cases where formalized machine-readable access to processed data enables higher level processing.

Derivative folders are derived datasets generated from a raw dataset. They must include ONLY processed data obtained from a specific raw dataset (e.g., segmentations, masks, labels).


In this section we decided not to fully follow the BIDS derivatives convention. For more information please see our related issue.

Derivatives structure#

According to BIDS, derived datasets could be stored inside a parent folder derivatives/ β€œto make a clear distinction between raw data and results of data processing”. This folder should also follow the same folder logic as the one used for the raw data.

Derivative data obtained using different processes/workflows should ideally be stored using different derivatives folders. Eg:

  • derivatives/labels/

  • derivatives/sct_5.6/

  • derivatives/fmriprep_2.3/


Despite what is written above, to streamline data identification and reduce the need for extensive folder crawling, we opted for common folder names, such as labels/, that typically contains binary segmentation and point-wise labels.

Derived datasets follow the same structure and hierarchy as the raw dataset, with folders corresponding to subjects, [sessions] and MRI modalities:

β”œβ”€β”€ dataset_description.json
β”œβ”€β”€ participants.tsv
β”œβ”€β”€ participants.json
β”œβ”€β”€ code/
β”œβ”€β”€ sub-<label>/
└── derivatives/
    └── <label>  <-- name of the derivative folder
        └── sub-<label>/]
            └── [ses-<label>/]
                └── data type/  <-- could be 'anat', 'fmap', 'func', etc.
                    └── <source_filename>[_space-<space>][_res-<label>][_den-<label>][_desc-<label>]_<suffix>.<extension>


Entities and suffixes are different from those used with the raw filenames and are specific to data types.


Because derived datasets are datasets, files and folders presented in the raw template section could also be included in this dataset (e.g., code/, etc.)


This element corresponds to the entire source filename, with the omission of the extension. For example, if the source file name is sub-02_acq-MTon_MTS.nii.gz, the <source_filename> to be used for the derivatives is sub-02_acq-MTon_MTS.

Derivative entities#

Characterized by a key word (space, res, den, etc.) and a value (label = an alphanumeric value, index = a nonnegative integer, etc) separated with a dash -

  • [space-<space>]: image space if different from raw space: template space (e.g. MNI305 etc), orig, other etc. (see BIDS)

  • [res-<label>]: for changes in resolution

  • [den-<label>]: for changes related to density

  • [desc-<label>]: should be used to distinguish two files that do not otherwise have a distinguishing entity. (e.g. sub-001_UNIT1_desc-denoised.nii.gz)

  • [label-<label>]: to avoid confusion if multiple masks are available we have to specify the masked structure (i.e. _label-WM for white matter, _label-GM for gray matter, _label-lesion for lesions etc.)

Entities are then separated using underscores _

Derivative suffixes#

An alphanumeric string located after all the entities following a final underscore _ :

Image type (suffix)

Associated entities




Suffix used for binary masks (0 and 1 only). The entity is used to specify the segmented structure in the image.



Suffix used for discrete segmentations representing multiple anatomical structures. The entity is used to specify the atlas used to map the different structures.



Suffix used for soft segmentations representing anatomical structures with values ranging from 0 to 1. The entity is used to specify the segmented structure in the image.



Suffix used for binary labels (0 and 1 only). The entity is used to specify the type of structure labeled in the image.



Suffix used for discrete labels representing multiple anatomical structures. The entity is used to specify the atlas used to label the different structures


Here, the corresponding entity label-<label> is mandatory to specify the labeled region.

Derivatives extensions#

Files extensions:

  • .nii.gz

  • .json

  • etc.

In addition to the subjects folders, derived datasets must include their own dataset_description.json file to track all the processing steps used to create the data. Example:


    "BIDSVersion": "1.9.0",
    "Name": "<dataset_name>",
    "DatasetType": "derivative"


To provide more details about the processing steps (e.g., reorientation, resampling), a descriptions.tsv file may be added at the root of the folder. This file must contain at least two columns:

  • desc_id: contains all the labels used with the desc entity within the filenames accross the entire dataset.

  • description: human readable descriptions

JSON sidecars#

JSON sidecars are companion files linked to data files. They share the same filenames but have a β€œ.json” extension. These files store essential metadata, serving as guidebooks to provide crucial details about the associated data, ensuring organized and comprehensive information.

Therefore, to improve the way we track our data, .json sidecars have to be generated for each data present in derived datasets. Here are few examples of JSON sidecar:

Below is a JSON sidecar describing a fully-manual labels created in the ORIGINAL SPACE:

    "SpatialReference": "orig",
    "GeneratedBy": [
            "Name": "Manual",
            "Author": "Nathan Molinier",
            "Date": "2023-07-14 13:43:10"

If the label was previously produced by an automatic algorithm, append to the GeneratedBy section:

    "SpatialReference": "orig",
    "GeneratedBy": [
            "Name": "sct_deepseg_sc",
            "Version": "SCT v6.1"
            "Name": "Manual",
            "Author": "Nathan Molinier",
            "Date": "2023-07-14 13:43:10"

If the label is created after the data was resampled and cropped, indicate it under SpatialReference:

    "SpatialReference": {
        "ResamplingFactor": "2",
        "Interpolation": "spline",
        "Xmin": 5,
        "Xmax": 95,
        "Ymin": 2,
        "Ymax": 18,
        "Zmin": 4,
        "Zmax": 100
    "GeneratedBy": [
            "Name": "sct_resample",
            "Version": "SCT v6.1"
            "Name": "sct_crop_image",
            "Version": "SCT v6.1"

Another example of a label created in another space than the image (here: the PAM50 template space):

    "SpatialReference": "PAM50",
    "GeneratedBy": [
            "Name": "sct_register_to_template",
            "Version": "SCT v6.1"


For better clarity, if the image space is different between the raw data and the label (as is the case above), the entity space-other MUST also be used in the filename. For templates, the entity space-template or space-<template_name> (e.g. space-PAM50) may be used instead.

Label names#

To be consistent regarding the way anatomical regions will be referred to, please follow this table (based on the BIDS labels):




Spinal Cord


Gray Matter


White Matter


Intervertebral discs, with values following this convention


Vertebrae, with values following this convention


Spinal rootlets


Pontomedullary Junction, indicated as a single voxel with a value β€˜50’


Cerebrospinal Fluid


Spinal canal


Spinal Cord Compression, indicated as a single voxel with a value β€˜1’ at the point of compression. There can be more than one compression.


Lesion (e.g., multiple sclerosis plaques, spinal cord injury lesions). The pathology associated with the lesion is indicated in the file participants.tsv








Axon (used in microscopy datasets)


Myelin (used in microscopy datasets)

When multiple anatomical regions are present in the image, atlases should be used. When specified, these atlases SHOULD be added to a folder atlases/ at the root of the derivative folder or a URL should be included inside the json sidecars.

Examples and use cases#

Here is an example of a dataset structure with a single subject sub-001:

β”œβ”€β”€ dataset_description.json
β”œβ”€β”€ participants.tsv
β”œβ”€β”€ participants.json
β”œβ”€β”€ code/
β”‚   └──
β”œβ”€β”€ sub-001
β”‚   └── anat
β”‚       β”œβ”€β”€sub-001_acq-sag_T2w.nii.gz
β”‚       └──sub-001_acq-sag_T2w.json
└── derivatives
    └── labels
        β”œβ”€β”€ dataset_description.json
        └── sub-001
            └── anat
                β”œβ”€β”€ sub-001_acq-sag_T2w_label-SC_seg.nii.gz  # spinal cord (SC) binary segmentation 
                β”œβ”€β”€ sub-001_acq-sag_T2w_label-SC_softseg.nii.gz  # spinal cord (SC) soft segmentation
                β”œβ”€β”€ sub-001_acq-sag_T2w_label-discs_dlabel.nii.gz  # discrete discs labeling
                β”œβ”€β”€ sub-001_acq-sag_T2w_label-vertebrae_dseg  # vertebrae discrete segmentation (segmented stuctures have different values based on the vertebral levels)
                β”œβ”€β”€ sub-001_acq-sag_T2w_label-rootlets_dseg  # nerve rootlets discrete segmentation
                β”œβ”€β”€ sub-001_acq-sag_T2w_label-compression_label.nii.gz  # binary compression labeling
                β”œβ”€β”€ sub-001_acq-sag_T2w_label-PMJ_dlabel  # Pontomedullary junction, indicated as a single voxel with a value '50'
                └── sub-001_acq-sag_T2w_label-lesion_seg  # lesion binary segmentation

Changelog policy#

We use git log to track our changes. That means care should be taken to write good messages: they are there to help both you and future researchers understand how the dataset evolved.

Good commit message examples:

git commit -m 'Segment spines of subjects 010 through 023
Produced manually, using fsleyes.'


git commit -m 'Add new subjects provided by <email_adress>'

If you choose to also fill in BIDS’s optional CHANGES file make sure it reflects the git log.