---
license: cc-by-3.0
tags:
- medical
- brain-mri
- glioma
- tumor-segmentation
- tcia
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: patient_id
    dtype: string
  - name: cohort
    dtype: string
  - name: gt_tier
    dtype: string
  - name: num_slices
    dtype: int32
  - name: shape
    dtype: string
  - name: image
    dtype: image
  - name: mask
    dtype: image
  - name: overlay
    dtype: image
  splits:
  - name: train
    num_bytes: 6186343
    num_examples: 167
  download_size: 6186175
  dataset_size: 6186343
---

# BraTS-TCGA (BraTS-TCGA-GBM + BraTS-TCGA-LGG)

Expert segmentation labels for the pre-operative TCGA glioma MRI cohorts
(Bakas et al. 2017), combining the two TCIA analysis-result collections
**BraTS-TCGA-GBM** (102 glioblastoma patients) and **BraTS-TCGA-LGG**
(65 lower-grade glioma patients) = **167 cases**.

**What this is (faithful-naming note):** the publicly released *training*
half of the pre-operative subset of TCGA-GBM / TCGA-LGG, already
co-registered to a T1 template, resampled to 1 mm³, and skull-stripped
(NIfTI). The 33 GBM + 43 LGG challenge *test* subjects are withheld by
TCIA (controlled access) and are NOT included. The raw DICOM collections
(TCGA-GBM / TCGA-LGG) are separate, NIH-controlled, and not mirrored here.

## Structure

```
dataset/{gbm|lgg}/TCGA-XX-XXXX/
  TCGA-XX-XXXX_<date>_t1.nii.gz
  TCGA-XX-XXXX_<date>_t1Gd.nii.gz
  TCGA-XX-XXXX_<date>_t2.nii.gz
  TCGA-XX-XXXX_<date>_flair.nii.gz
  TCGA-XX-XXXX_<date>_GlistrBoost.nii.gz                    (automated)
  TCGA-XX-XXXX_<date>_GlistrBoost_ManuallyCorrected.nii.gz  (when present)
train.jsonl              # one record per case; `mask` = recommended GT
TCGA_GBM_radiomicFeatures.csv
TCGA_LGG_radiomicFeatures.csv
```

## Labels (BraTS convention)

| value | structure |
|---|---|
| 1 | necrotic + non-enhancing tumor core (NCR/NET) |
| 2 | peritumoral edema (ED) |
| 4 | GD-enhancing tumor (ET) |

Some LGG tumors do not enhance — label 4 legitimately absent in those
cases (class absent, not an empty/broken mask).

## Ground truth tier

Two masks per case: `GlistrBoost` (automated, BraTS'15-winning method) and
`GlistrBoost_ManuallyCorrected` (revised and approved by a board-certified
neuroradiologist). **Recommended GT = ManuallyCorrected when present**
(97/102 GBM, 62/65 LGG); for the 8 cases without it the automated mask was
accepted as-is. `train.jsonl` field `mask` already applies this rule
(`gt_tier` records which file was chosen). Known quirk: the corrected
files can carry slightly different NIfTI headers than the images — take
geometry from the image volume.

## Splits

No internal split — this release is the BraTS 2017 training portion only
(single split: `train`).

## ⚠️ Benchmark overlap

All 167 subjects were folded into the **BraTS challenge training data
(2017 onward)** — do not treat this set as independent of models trained
on BraTS (e.g. BraTS2023-GLI). The same TCGA-LGG patients also appear in
the 2D `TCGA-LGG-Mask` dataset. Folder names are TCGA patient barcodes
(`TCGA-XX-XXXX`) — use them for cross-referencing/deduplication; the
BraTS name-mapping CSV distributed with BraTS'17–'20 training archives
links barcodes to BraTS subject IDs.

## License & citation

CC BY 3.0. Cite:

1. Bakas S, et al. "Advancing The Cancer Genome Atlas glioma MRI
   collections with expert segmentation labels and radiomic features."
   Nature Scientific Data 4:170117 (2017). DOI: 10.1038/sdata.2017.117
2. Data DOIs: GBM 10.7937/K9/TCIA.2017.KLXWJJ1Q,
   LGG 10.7937/K9/TCIA.2017.GJQ7R0EF
3. Clark K, et al. "The Cancer Imaging Archive (TCIA)." J Digit Imaging
   26(6):1045-1057 (2013).