Mindboggle-101 / README.md
Angelou0516's picture
Upload dataset
d7f19a9 verified
---
license: cc-by-nc-sa-3.0
pretty_name: Mindboggle-101
tags:
- medical
- mri
- brain
- t1-weighted
- cortical-parcellation
- brain-segmentation
- neuroimaging
task_categories:
- image-segmentation
size_categories:
- n<1K
dataset_info:
features:
- name: subject_id
dtype: string
- name: cohort
dtype: string
- name: source_archive
dtype: string
- name: sex
dtype: string
- name: age
dtype: string
- name: image_shape
dtype: string
- name: num_slices
dtype: int32
- name: voxel_spacing
dtype: string
- name: num_cortical_regions
dtype: int32
- name: image
dtype: image
- name: mask
dtype: image
- name: overlay
dtype: image
splits:
- name: train
num_bytes: 9419820
num_examples: 101
download_size: 9426186
dataset_size: 9419820
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---
# Mindboggle-101
Manually-labeled cortical parcellations for **101 healthy-adult T1-weighted brain MRI** volumes,
following the Desikan-Killiany-Tourville (DKT) protocol. This is the largest set of consistently,
manually labeled human brains and is the reference behind the DKT cortical atlas shipped inside
FreeSurfer.
> **Scope of this mirror — read before use.** This repository is the **volumetric** subset of the
> full Mindboggle-101 distribution: image volumes + manual **cortical** label volumes. The
> FreeSurfer **surface** meshes/labels and the **shape-feature tables** that the official deposit
> also contains are **not** included (they are not voxel segmentation masks) — obtain them from the
> Harvard Dataverse DOIs below. The **whole-brain (cortex+subcortex) volume labels** for the
> OASIS-TRT-20 cohort, and the per-subject `+aseg` (cortical+subcortical) variants, are **excluded**
> — see *Licensing*.
## Contents
**101 subjects**, one T1 brain each, organized by **source cohort** (no train/val/test split):
| Cohort | N | Origin archive |
|---|---|---|
| NKI-RS-22 | 22 | Nathan Kline Institute – Rockland Sample |
| MMRR-21 | 21 | Multi-Modal MRI Reproducibility Resource |
| NKI-TRT-20 | 20 | NKI Test–Retest |
| OASIS-TRT-20 | 20 | OASIS cross-sectional, test–retest |
| Extra-18 | 18 | HLN-12 + MMRR-3T7T-2 + Twins-2 + Colin27-1 + Afterthought-1 |
| **Total** | **101** | |
Per subject — `<Cohort>_volumes/<Subject-ID>/`:
| File | Description |
|---|---|
| `t1weighted.nii.gz` | Native-space T1 (≈256³, 1 mm isotropic, `int16`) |
| `t1weighted_brain.nii.gz` | Skull-stripped native T1 |
| `t1weighted.MNI152.nii.gz` | T1 affinely registered to MNI152 |
| `t1weighted_brain.MNI152.nii.gz` | Skull-stripped T1 in MNI152 |
| `t1weighted_brain.MNI152.affine.txt` | Native→MNI152 affine |
| **`labels.DKT31.manual.nii.gz`** | **Gold-standard manual DKT31 cortical labels (native space)** |
| `labels.DKT31.manual.MNI152.nii.gz` | Same labels in MNI152 space |
Also at the repo root: `subject_sources_Mindboggle101.txt`, `subject_list_Mindboggle101.txt`,
`label_definitions.txt`, `label-issues_201903.txt`, `subject_table_Mindboggle101.pdf`,
`scans/` (per-cohort acquisition info), and the Apache-2.0 `code_*` helper files.
## Ground truth & label scheme
**Recommended evaluation reference: `labels.DKT31.manual.nii.gz`** — manually-edited cortical
parcellation, **62 cortical regions** (31 per hemisphere) + background. FreeSurfer/`aparc`-style
integer codes are stored as `float32` (cast to int on load):
- Left hemisphere `1002–1035`; right hemisphere `2002–2035`. Codes are **non-contiguous** (e.g.
1004/1032/1033 absent — those regions are excluded under DKT31); **remap to a contiguous 0…62
range** for training/scoring.
- Region names per code: see **`label_definitions.txt`**.
- Cortex is a thin ribbon → ~95 % of voxels are background (label 0); expect strong class imbalance.
The reduced **DKT25** scheme and the FreeSurfer-`aseg` whole-brain variants exist only in the
surface / `+aseg` products and are **not** in this volumetric mirror.
## Cross-dataset overlap (important)
Every Mindboggle-101 subject is **drawn from another public dataset** (OASIS, NKI-RS, NKI-TRT,
MMRR, the MNI Colin27 template). Before combining this with any **OASIS- or NKI-derived** benchmark,
**de-duplicate** using **`subject_sources_Mindboggle101.txt`** (preserved at the repo root), which
maps each Mindboggle ID (e.g. `OASIS-TRT-20-1`) back to its original-archive identifier. The DKT
atlas derived from these brains also ships inside FreeSurfer (`aparc.DKTatlas`), and CerebrA is a
corrected MNI-registered derivative. No shared lineage with the BraTS family or the Medical
Segmentation Decathlon.
## Licensing (per component)
This mirror follows the **authors' own copyright notices shipped inside the data**, which are more
authoritative than third-party host tags:
| Component | License | Where stated |
|---|---|---|
| T1 images + manual DKT cortical labels (the 101 subjects) | **CC BY-NC-SA 3.0** | `subject_sources_Mindboggle101.txt` (Mindboggle team, 2012) |
| Labeling-protocol doc + helper code (`label_definitions.txt`, `code_*`) | Apache-2.0 | bundled file headers |
| *(excluded)* OASIS-TRT-20 whole-brain cortex+subcortex labels | CC BY-NC-ND 3.0 | Neuromorphometrics BrainCOLOR-derived |
- The repo `license:` tag is **`cc-by-nc-sa-3.0`**, reflecting the image+label data (the primary
content): **non-commercial**, **attribution** required, derivatives **shared alike**.
- Harvard Dataverse tags the primary image/label files **CC0**; we intentionally adopt the **more
restrictive** CC BY-NC-SA 3.0 from the authors' bundled notice as the safer reading of the
conflicting signals (NITRC also lists CC BY-NC-SA 3.0).
- The CC BY-NC-ND **subcortical** whole-brain labels were dropped so this repository can be cleanly
redistributed and previewed; obtain them from the Dataverse DOI below if required.
## Provenance & source
Author-maintained: <https://mindboggle.info/data.html> · OSF [`nhtur`](https://osf.io/nhtur/) ·
Harvard Dataverse — individual brains [`doi:10.7910/DVN/HMQKCK`](https://doi.org/10.7910/DVN/HMQKCK)
(**this mirror's source**), atlases `doi:10.7910/DVN/XCCE9Q`, templates `doi:10.7910/DVN/WDIYB5`.
Subject count verified against the paper: **101/101**.
## `data/` preview
The `data/` directory holds a Parquet preview (middle-axial-slice PNG + cortical-label overlay per
volume) for the Hugging Face Dataset Viewer. Previews are rendered **only** from the CC BY-NC-SA
cortical labels. The authoritative data are the raw `.nii.gz` files in the `*_volumes/` folders.
## Citation
> Klein A, Tourville J. *101 labeled brain images and a consistent human cortical labeling
> protocol.* Frontiers in Neuroscience 6:171, 2012. doi:10.3389/fnins.2012.00171
Please also acknowledge the source archives (OASIS, NKI / Rockland, MMRR) per their terms when
using the corresponding cohorts.