---
license: cc-by-4.0
task_categories:
- image-segmentation
tags:
- medical
- mri
- mpmri
- brain
- glioblastoma
- glioma
- brain-tumor
- tumor-segmentation
- brats
- nifti
- tcia
- upenn-gbm
pretty_name: UPENN-GBM (mpMRI + Tumor Segmentation)
size_categories:
- n<1K
dataset_info:
  features:
  - name: subject_id
    dtype: string
  - name: patient_id
    dtype: string
  - name: timepoint
    dtype: string
  - name: gt_source
    dtype: string
  - name: num_slices
    dtype: int32
  - name: slice_index
    dtype: int32
  - name: tumor_voxels
    dtype: int64
  - name: modality
    dtype: string
  - name: image
    dtype: image
  - name: mask
    dtype: image
  - name: overlay
    dtype: image
  splits:
  - name: manual
    num_bytes: 5989524
    num_examples: 147
  - name: automated
    num_bytes: 18932953
    num_examples: 464
  - name: unsegmented
    num_bytes: 856502
    num_examples: 60
  download_size: 25763705
  dataset_size: 25778979
configs:
- config_name: default
  data_files:
  - split: manual
    path: data/manual-*
  - split: automated
    path: data/automated-*
  - split: unsegmented
    path: data/unsegmented-*
---

# UPENN-GBM — mpMRI + Tumor Segmentation

The University of Pennsylvania glioblastoma (UPenn-GBM) cohort: structural
multi-parametric MRI (mpMRI) of de novo glioblastoma patients with tumor
sub-region segmentations. This is the **NIfTI** release from TCIA — images are
skull-stripped and co-registered to the SRI24 atlas, and the segmentations are
aligned to them.

> **This HuggingFace mirror is a segmentation-focused subset** of the full TCIA
> collection. It contains the structural mpMRI sequences and the tumor
> segmentations only. The DICOM package (139 GB), the DSC/DTI derivative maps,
> the skull-**unstripped** images, the histopathology WSIs (149 GB), and the
> radiomic-feature tables are **not** included here — obtain those from TCIA.

## Dataset Details

| Field | Value |
|---|---|
| Modality | Brain mpMRI — T1, T1-Gd (T1CE), T2, T2-FLAIR |
| Body part | Brain (de novo glioblastoma) |
| Task | 3D multi-class tumor sub-region segmentation |
| Structural scans | 671 (630 patients; `_21` = follow-up timepoints) |
| Manual/expert masks | 147 (`images_segm`) — **recommended ground truth** |
| Automated masks | 611 (`automated_segm`) |
| Segmentable scans | 611 (have a manual and/or automated mask) |
| Volume geometry | 240 × 240 × 155, 1 mm isotropic, SRI24 atlas space |
| Format | NIfTI (`.nii.gz`) |
| License | CC BY 4.0 |

## Label Scheme (BraTS convention)

| Value | Tumor sub-region |
|---|---|
| 0 | Background |
| 1 | Necrotic / non-enhancing tumor core (NCR/NET) |
| 2 | Peritumoral edematous / infiltrated tissue (ED) |
| 4 | GD-enhancing tumor (ET) |

Evaluation regions: **WT** (whole tumor) = 1+2+4, **TC** (tumor core) = 1+4,
**ET** (enhancing tumor) = 4. Note the enhancing-tumor label is **4** (native
BraTS/TCIA encoding), *not* 3 — verified across the released masks.

## Mask Sources (two)

1. **`images_segm` — manually-corrected expert segmentation (147 scans).**
   Automated labels reviewed and corrected/approved by board-certified
   neuroradiologists. **This is the recommended ground truth.**
2. **`automated_segm` — automated segmentation (611 scans).** Label fusion
   (STAPLE) of an ensemble of top BraTS-ranked deep models (DeepMedic,
   DeepSCAN, nnU-Net). Silver/weak standard.

All 147 manually-corrected scans also have an automated mask. 60 scans
(follow-up `_21` timepoints) have neither and are image-only.

**Recommended GT policy (used by `subjects_manifest.json`):** prefer the manual
mask in `images_segm`; fall back to `automated_segm`; skip scans with neither.

## Structure

```
images_structural/<subject>/<subject>_T1.nii.gz
images_structural/<subject>/<subject>_T1GD.nii.gz
images_structural/<subject>/<subject>_T2.nii.gz
images_structural/<subject>/<subject>_FLAIR.nii.gz
images_segm/<subject>_segm.nii.gz                          # manual/expert GT (147)
automated_segm/<subject>_automated_approx_segm.nii.gz      # automated (611)
metadata/UPENN-GBM_clinical_info_v2.1.csv                  # clinical + genomic per subject
subjects_manifest.json                                     # per-scan paths, mask availability, GT policy
```

`<subject>` = `UPENN-GBM-NNNNN_TT`, where `TT` is the timepoint (`11` = baseline,
`21` = follow-up). `subjects_manifest.json` lists, for every structural scan, the
four modality paths, the manual/automated mask paths (if present), and the chosen
`gt_path`/`gt_source` — so loaders need not re-derive availability.

## Notes for Loaders

- **Images and masks share an identical grid** (240×240×155, 1 mm iso, SRI24) —
  no resampling or axis permutation is needed between a scan and its mask.
- The NIfTI images are SRI-registered and **do not** align with the TCIA DICOM
  package by design.
- Multi-channel input: stack T1/T1GD/T2/FLAIR as channels (BraTS-style).

## Source

- TCIA collection: https://www.cancerimagingarchive.net/collection/upenn-gbm/
- DOI: `10.7937/TCIA.709X-DN49`
- Public, no registration required (TCIA fully public since 2025-07-07).

## Citation

```bibtex
@article{bakas2022upenngbm,
  author  = {Bakas, Spyridon and Sako, Chiharu and Akbari, Hamed and Bilello, Michel
             and Sotiras, Aristeidis and Shukla, Gaurav and Rudie, Jeffrey D. and
             Flores Santamar\'ia, Nadina and Fathi Kazerooni, Anahita and Pati, Sarthak
             and others},
  title   = {The University of Pennsylvania glioblastoma (UPenn-GBM) cohort:
             advanced MRI, clinical, genomics, \& radiomics},
  journal = {Scientific Data},
  volume  = {9},
  number  = {1},
  pages   = {453},
  year    = {2022},
  doi     = {10.1038/s41597-022-01560-7}
}
```