Glasgow 7T Database

There are 2 repositories:

The full db is available on EBRAINS (link).
A small sample of data, with manual annotations, is available here.

Full dataset

You can download the Glasgow data, anatomical and segmentation masks, from the EBRAINS website (link). You may need to create an account to access the data.

The database consists of 142 anatomical volumes (76 unique subjects) obtained with a MP2RAGE sequence at 0.63 mm3 isotropic resolution, using a 7-Tesla MRI scanner with 32-channel head coil. All volumes were collected, as reconstructed DICOM images, at the Imaging Centre of Excellence (ICE) at the Queen Elizabeth University Hospital, Glasgow (UK). For every volume, we release the corresponding scans INV1, INV2, and UNI_Images. In addition, we release the automatic segmentations obtained from different methods (see below).

Data structure

In our work, we used plain reconstructed data (no pre-processing). The only pre-process applied to the data is the conversion from DICOM to the NIfTI format, carried out using dcm2niix. All the volumes are in NIfTI format (.nii or .nii.gz). We used Nipy’s NiBabel to handle such MR Images.

The data should be organised as follows, for the subject N session S:

BIDS/
├── sub-00N
│   ├── anat
│   │   ├── sub-00N_ses-00S_INV1.json        -> INV1 descriptors
│   │   ├── sub-00N_ses-00S_INV1.nii.gz      -> INV1 volume
│   │   ├── sub-00N_ses-00S_INV2.json        -> INV2 descriptors
│   │   ├── sub-00N_ses-00S_INV2.nii.gz      -> INV1 volume
│   │   ├── sub-00N_ses-00S_T1w.json         -> MP2RAGE descriptors
│   │   ├── sub-00N_ses-00S_T1w.nii.gz       -> MP2RAGE volume
│   └── seg
│       ├── sub-00N_ses-00S_CEREBRUM7T.nii.gz      -> segmentation by our method (only for testing volumes)
│       ├── sub-00N_ses-00S_Fracasso16.nii.gz      -> segmentation by Fracasso (2016)
│       ├── sub-00N_ses-00S_FreeSurfer_v6.nii.gz   -> segmentation by FreeSurfer v06
│       ├── sub-00N_ses-00S_FreeSurfer_v7.nii.gz   -> segmentation by FreeSurfer v07
│       ├── sub-00N_ses-00S_nighres.nii.gz         -> segmentation by Huntenburg (2018)
│       ├── sub-00N_ses-00S_training_labels.nii.gz -> segmentation mask used for training 

Which is the same structure you can find in the EBRAINS release. Using a different data structure is possible, but the code needs to be slightly modified.

Tissue classes

Along with the publicly available data, 6-class (+ background) segmentation masks are provided. The segmented classes (and the color code used in the notebooks and in the paper) are:

Class ID	Substructure/Tissue	Color
0	Background	Transparent
1	Grey matter	Light green
2	Basal ganglia	Dark green
3	White matter	Red
4	Ventricles	Blue
5	Cerebellum	Yellow
6	Brainstem	Pink

Such ground truth was obtained using ad-hoc procedure (see the manuscript) using the classes used in the MICCAI MRBrainS13 and MRBrainS18 challenges.

Manual segmentation

The dataset with manual annotations (link) is composed by 3 subjects, scanned at the Imaging Centre of Excellence (ICE) at the Queen Elizabeth University Hospital, Glasgow (UK). For every subject, two folders are provided, containing:

anat/
- INV1
- INV2
- UNI_Images (T1w)
derivatives/
- manual annotations of 8 regions of widely interest in neuroscience
  - early visual cortex (EVC)
  - high-level visual areas (HVC)
  - motor cortex (MCX)
  - cerebellum (CER)
  - hippocampus (HIP)
  - early auditory cortex (EAC)
  - brainstem (BST)
  - basal ganglia (BGA)
- automatic segmentation by FreeSurfer (v6 and v7)
- automatic segmentation by Fracasso16
- automatic segmentation by our method (CEREBRUM7T) with probability maps (CEREBRUM7T_probMap)
- automatic segmentation by nighres
- labels used for training our method

Testing on OpenNeuro data

Note: if you are testing the trained model on the dataset published on OpenNeuro, please notice that you need to download the mean and std volumes at this link (psw: rocknroll87q/cerebrum7t).

CEREBRUM 7T