README.md

metadata

license: other
license_name: mixed-cc-by-4.0-and-cc-by-sa-4.0
license_link: https://surfer.nmr.mgh.harvard.edu/docs/synthstrip/data/README
pretty_name: SynthStrip
tags:
  - medical
  - mri
  - brain
  - skull-stripping
  - brain-extraction
  - brain-segmentation
  - neuroimaging
  - multi-contrast
task_categories:
  - image-segmentation
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
dataset_info:
  features:
    - name: scan_id
      dtype: string
    - name: subset
      dtype: string
    - name: modality
      dtype: string
    - name: modality_full
      dtype: string
    - name: modality_type
      dtype: string
    - name: subject
      dtype: string
    - name: split
      dtype: string
    - name: source_archive
      dtype: string
    - name: license
      dtype: string
    - name: ixi_image_overlap
      dtype: bool
    - name: image_shape
      dtype: string
    - name: num_slices
      dtype: int32
    - name: voxel_spacing
      dtype: string
    - name: brain_voxels
      dtype: int64
    - name: slice_index
      dtype: int32
    - name: image
      dtype: image
    - name: mask
      dtype: image
    - name: overlay
      dtype: image
  splits:
    - name: validation
      num_bytes: 2630827
      num_examples: 24
    - name: test
      num_bytes: 62832682
      num_examples: 558
  download_size: 65436774
  dataset_size: 65463509

SynthStrip

The SynthStrip dataset is a collection of full-head brain scans paired with manually-refined binary brain masks, assembled by the FreeSurfer / Martinos Center team to evaluate the SynthStrip brain-extraction (skull-stripping) tool across a wide range of contrasts, resolutions, and subject populations (newborns to glioblastoma patients).

Scope of this mirror — read before use. The full SynthStrip release spans 622 MRI, PET, and CT scans. This repository contains the 582 redistributable image + brain-mask pairs, which are entirely MRI. The reason is specific: the only PET/CT scans in SynthStrip are the CIM (CERMEP-IDB-MRXFDG) cases, and the authors may not redistribute those images — the official release ships CIM as masks only. We therefore exclude all 40 cim_* cases (an image-less mask cannot be scored). We also exclude the anatomical labels.nii.gz maps (a different, SynthSeg-style parcellation task — not the binary brain mask). What remains is a clean multi-contrast brain-MRI skull-stripping benchmark.

Contents

582 scans, organized as synthstrip_data_v1.5/<SUBSET>_<MODALITY>_<SUBJECT>/ with image.nii.gz + mask.nii.gz. No train split (pure evaluation set); the authors' val/test split is preserved in scan_metadata.csv and synthstrip_data_v1.5/scans_*.txt.

Subset	Source archive	Modalities (count)	N	License
IXI	Imperial College London	T1w 50 · T2w 50 · PDw 50 · MRA 50 · DWI 32	232	CC BY-SA 4.0
FSM	FreeSurfer Maintenance (MGH)	T1w 38 · T2w 36 · PDw 32 · qT1 32	138	CC BY 4.0
ASL	HCP-Aging / MGH	T1w 43 · PCASL 43	86	CC BY 4.0
QIN	TCIA QIN-GBM-Treatment-Response	T1w 54 · T2-FLAIR 17 · T2w 39	110	CC BY 4.0
Infant	Boston Children's Hospital	T1w 16 (ages 0–18 mo)	16	CC BY 4.0
Total			582
(excluded) CIM	CERMEP-IDB-MRXFDG	FDG PET 20 · CT 20 — masks only, no image	40	CC BY 4.0

Split (authors' publication split): validation = 24 · test = 558 (the official scans_test.txt lists 598, of which 40 are the excluded cim_* cases → 558 here).

Per scan — synthstrip_data_v1.5/<scan_id>/:

File	Content	dtype
image.nii.gz	unmodified native-space image (orientation/resolution vary per scan)	int16 / float32 / uint8
mask.nii.gz	binary brain mask (0 = non-brain, 1 = brain) — same shape & affine as the image	uint8

scan_metadata.csv (repo root) carries one row per scan: scan_id, subset, modality, modality_full, modality_type, subject, split, source_archive, license, ixi_image_overlap.

Ground truth

Single tier — the manually-refined binary brain mask. The reference for each subject was built by running every classical skull-stripper on the T1w scan and taking the per-voxel majority vote, then manually refining it and propagating to the subject's other contrasts by rigid registration; v1.5 additionally cleaned masks via largest-connected-component + hole-filling. There are no competing rater/auto tiers to choose between. (This is not a SAMSEG output — SAMSEG/SynthSeg is only used for the excluded anatomical labels.nii.gz.)

Cross-dataset overlap (important)

• IXI (232 scans, ~40% of this mirror) is EXACT source-image overlap with Angelou0516/ixi-brain — the same IXI scans, here carrying skull-stripping brain masks instead of ixi-brain's target. They are legitimate, distinct benchmark samples (different task), but do not count IXI as unique images when aggregating SynthStrip + ixi-brain — de-duplicate on the IXI subject id (the ixi_* scan_id token / subject column is the join key; ixi_image_overlap = true flags every such row).
• QIN (110 scans) is sourced from TCIA QIN-GBM-Treatment-Response, a glioblastoma cohort — shared TCIA lineage with the BraTS-family / GBM datasets in the suite. Low leakage concern: the task here is skull-stripping (binary brain mask), not tumor segmentation. The qin_* token / subject column traces back to the TCIA subject.
• ASL (86 scans) draws in part from the Human Connectome Project–Aging.
• No shared lineage with the Medical Segmentation Decathlon.

Licensing (per subset)

The release uses two licenses, so this repo is tagged license: other rather than a single CC tag (tagging the whole repo plain CC BY 4.0 would mis-state the IXI images):

Component	License
IXI images and their brain masks (232 scans)	CC BY-SA 4.0 (share-alike)
All other images + masks (FSM, ASL, QIN, Infant)	CC BY 4.0

The share-alike clause propagates to any derivative of the IXI data — including the IXI rows of the data/ parquet preview, which are likewise CC BY-SA 4.0. See the official license notice.

Provenance & source

Author-hosted by FreeSurfer / Martinos Center (not a third-party re-host): persistent DOI https://w3id.org/synthstrip → https://surfer.nmr.mgh.harvard.edu/docs/synthstrip/ · 3D release v1.5 (2024-06-20). The full release reports 622 scans; 582 are redistributable and mirrored here (see Scope). Each subset must additionally be credited to its original archive (IXI; FreeSurfer/FSM; HCP-Aging; TCIA QIN-GBM; Boston Children's Hospital) per the README.

data/ preview

data/ holds a Parquet preview for the Hugging Face Dataset Viewer — for each scan, the brain-mask-centred slice of the image, the binary mask, and an overlay. The authoritative data are the raw .nii.gz files under synthstrip_data_v1.5/. IXI-derived preview rows are CC BY-SA 4.0.

Citation

Hoopes A, Mora JS, Dalca AV, Fischl B*, Hoffmann M*. SynthStrip: skull-stripping for any brain image. NeuroImage, 260:119474, 2022. doi:10.1016/j.neuroimage.2022.119474 (*equal contribution)

Please also cite the source archive for whichever subset(s) you use (see the bundled synthstrip_data_v1.5/README.txt for the full per-subset reference list).