GBM-struct-27

GBM-struct-27: a 10-case subset of UPENN GBM structural MRI dataset with multi-rater segmentations of 27 brain structures in the presence of glioblastoma
===========================================================================================

Overview
----------
The GBM-struct-27 dataset contains a subset of 10 glioblastoma (GBM) cases from
the UPENN-GBM collection hosted on The Cancer Imaging Archive (TCIA):
https://doi.org/10.7937/TCIA.709X-DN49

The dataset includes MRI NIfTI volumes and manual segmentations of
27 brain structures produced by four independent raters. Segmentations were
performed on selected axial slices of the T1-weighted volumes rather than across
the full 3D volume. Glioblastoma tumour segmentations are not included.

The MRI NIfTI volumes in this release were already pre-registered to the SRI24 atlas
space, resampled to a common 155-slice grid, defaced, and not skull-stripped. No
additional spatial registration was performed by the authors.

Dataset structure
-----------------
The primary data are organised under the derivatives/ directory:

GBM-struct-27/
├── derivatives/
│   ├── upenn-gbm-sri24/
│   │   ├── sub-UPENNGBM00035/
│   │   │   └── anat/
│   │   │       ├── sub-UPENNGBM00035_ce-gad_space-SRI24_desc-unstripped_T1w.nii.gz
│   │   │       ├── sub-UPENNGBM00035_space-SRI24_desc-unstripped_FLAIR.nii.gz
│   │   │       ├── sub-UPENNGBM00035_space-SRI24_desc-unstripped_T1w.nii.gz
│   │   │       └── sub-UPENNGBM00035_space-SRI24_desc-unstripped_T2w.nii.gz
│   │   ├── sub-UPENNGBM00041/
│   │   │   └── anat/ ...
│   │   └── dataset_description.json
│   └── upenn-gbm-sri24-seg/
│       ├── sub-UPENNGBM00035/
│       │   └── anat/
│       │       ├── sub-UPENNGBM00035_space-SRI24_label-structures_desc-raterA_T1w.nii.gz
│       │       ├── sub-UPENNGBM00035_space-SRI24_label-structures_desc-raterB_T1w.nii.gz
│       │       ├── sub-UPENNGBM00035_space-SRI24_label-structures_desc-raterC_T1w.nii.gz
│       │       └── sub-UPENNGBM00035_space-SRI24_label-structures_desc-raterD_T1w.nii.gz
│       ├── sub-UPENNGBM00041/
│       │   └── anat/ ...
│       └── dataset_description.json
├── annotation_protocol_version_1.pdf
├── annotation_protocol_version_2_revised.pdf
├── dataset_description.json
├── itksnap_label_map.label
└── labels.json

Subjects included in this release:

- sub-UPENNGBM00035
- sub-UPENNGBM00041
- sub-UPENNGBM00054
- sub-UPENNGBM00101
- sub-UPENNGBM00108
- sub-UPENNGBM00117
- sub-UPENNGBM00143
- sub-UPENNGBM00146
- sub-UPENNGBM00174
- sub-UPENNGBM00192

Preprocessed MRI data (derivatives)
-----------------------------------
The derivatives/upenn-gbm-sri24/ directory contains SRI24-aligned structural MRI
NIfTI volumes for each subject. Each subject directory contains an anat/ folder
with four MRI files:

- T1-weighted image:
  sub-<ID>_space-SRI24_desc-unstripped_T1w.nii.gz
- Post-contrast T1-weighted image:
  sub-<ID>_ce-gad_space-SRI24_desc-unstripped_T1w.nii.gz
- T2-weighted image:
  sub-<ID>_space-SRI24_desc-unstripped_T2w.nii.gz
- T2-FLAIR image:
  sub-<ID>_space-SRI24_desc-unstripped_FLAIR.nii.gz

These MRI volumes were used as the reference images for manual annotation.

Segmentations (derivatives)
----------------------------
The derivatives/upenn-gbm-sri24-seg/ directory contains manual segmentations of
27 anatomical brain structures delineated on the SRI24-aligned T1-weighted NIfTI
volumes.

For each subject, four independent raters provided segmentation labels stored as
integer-valued NIfTI label maps. All segmentation files share the same voxel grid
and spatial reference as the corresponding SRI24-aligned T1-weighted MRI.

For each subject sub-<ID>, segmentation files are stored in:

- derivatives/upenn-gbm-sri24-seg/sub-<ID>/anat/

with the following filename pattern:

- sub-<ID>_space-SRI24_label-structures_desc-raterA_T1w.nii.gz
- sub-<ID>_space-SRI24_label-structures_desc-raterB_T1w.nii.gz
- sub-<ID>_space-SRI24_label-structures_desc-raterC_T1w.nii.gz
- sub-<ID>_space-SRI24_label-structures_desc-raterD_T1w.nii.gz

Files are integer label maps in the same space as the corresponding T1w images.
Label indices and anatomical meanings are provided in the accompanying label
information distributed with this Zenodo record.

Important note on annotation dimensionality
-------------------------------------------
Although MRI data are provided as 3D NIfTI volumes, manual segmentations were
performed exclusively on four selected axial slices. Only these slices contain
non-zero labels in the segmentation volumes. Unlabelled slices should not be
interpreted as negative annotations.

Raters
------
- desc-raterA: Rater A (head author; over five years of combined research and
  clinical experience in radiology, including manual annotations of benign and
  malignant brain tumours)
- desc-raterB: Rater B (year 3 medical student with research experience in
  neurosurgery)
- desc-raterC: Rater C (junior clinical fellow with over five years of clinical
  experience in neurosurgery)
- desc-raterD: Rater D (junior clinical fellow with over three years of clinical
  experience in neurosurgery)

All raters used ITK-SNAP for manual annotation following a shared protocol.

Label definitions
-----------------
The labels.json file provides a mapping from integer values in the segmentation
NIfTIs to anatomical structures (e.g., caudate, putamen, thalamus, ventricles,
cerebellar hemispheres). This file was derived from the ITK-SNAP label map (itksnap_label_map.label)
used during annotation.

Annotation protocols
--------------------
The dataset is accompanied by two versions of the annotation protocol. The
protocol documents the procedures used to define, label, review, and
quality-control the annotations.

Version 1 (annotation_protocol_version_1.pdf) was used to create the manual segmentations for GBM-struct-27.
Version 2 (annotation_protocol_version_2_revised.pdf) is an improved annotation protocol based on quantitative and
qualitative analysis of inter- and intra-rater variability.

Both versions of the annotation protocol form part of the manuscript submitted
for publication. Please cite this Zenodo record when reusing the dataset or
annotation protocol, until the manuscript is published.

Provenance and citation
-----------------------
Original MRIs are from:

- Bakas et al. (2021). Multi-parametric magnetic resonance imaging (mpMRI)
  scans for de novo Glioblastoma (GBM) patients from the University of
  Pennsylvania Health System (UPENN-GBM). The Cancer Imaging Archive.
  https://doi.org/10.7937/TCIA.709X-DN49

Segmentation labels originate from:

- Marina Ivory et al. Manual 2D Segmentation Protocol for Selected Brain
  Structures from T1-weighted MRI in the Presence of Glioblastoma: Challenges,
  Variability and Recommendations. Brain and Spine, Volume 5, Supplement 2,
  2025, 105518, ISSN 2772-5294. https://doi.org/10.1016/j.bas.2025.105518

If you use this dataset, please cite both:

1. This Zenodo record: GBM-struct-27 dataset, doi:10.5281/zenodo.17524203
2. The UPENN-GBM TCIA dataset, doi:10.7937/TCIA.709X-DN49

License
-------
For the purpose of open access, the authors have applied a CC-BY public copyright license to any Author Accepted Manuscript version arising from this submission. Users must cite this Zenodo record and the original UPENN-GBM TCIA dataset when reusing the data.

Contact
-------
For questions about this dataset or segmentations, please contact:
k2479928@kcl.ac.uk