TCGA-LGG-Mask — LGG Segmentation Dataset (Brain MRI)

Mirror of the LGG Segmentation Dataset ("Brain MRI segmentation"), the canonical release by Mateusz Buda on Kaggle (mateuszbuda/lgg-mri-segmentation), associated with:

Buda M., Saha A., Mazurowski M.A. Association of genomic subtypes of lower-grade gliomas with shape features automatically extracted by a deep learning algorithm. Computers in Biology and Medicine 109:218-225, 2019.

Brain MRI of 110 patients from the TCIA TCGA-LGG (lower-grade glioma) collection, with manual FLAIR-abnormality (tumor) segmentation masks created by the Mazurowski lab (Duke) and radiologist-verified.

⚠️ Naming note

Despite "Mask" in the repository name, this is NOT a masks-only set — it is the full image + mask paired segmentation dataset (3,929 image slices and 3,929 corresponding binary masks).

⚠️ Cross-dataset overlap (evaluation hazard)

These ~110 TCGA-LGG patients overlap with the BraTS family (BraTS 2017+ training data was built from the TCIA TCGA-GBM/LGG collections; ~108 pre-operative TCGA-LGG subjects were folded into BraTS). If you benchmark against any BraTS-derived model or split, exclude the TCGA-LGG intersection first to avoid leakage (same pattern as the documented UCSF-PDGM ⊃ BraTS hazard).

• Cross-reference key: the tcga_id column (e.g. TCGA_HT_8018). BraTS TCGA mapping files (TCIA / CBICA) link these to BraTS subject IDs.
• Labels differ: here the mask is a single binary whole-FLAIR-abnormality region; BraTS uses multi-class enhancing / edema / necrosis labels. The masks are therefore not interchangeable even for shared subjects.

Composition

• Patients (volumes): 110 — sites: DU 45, HT 34, CS 16, FG 14, EZ 1
• Slices: 3,929 (20-88 per patient, mean 35.7), all 256×256
• Tumor-containing slices: 1,373 · empty slices: 2,556
• Image channels: 3-channel RGB TIFF stacking pre-contrast / FLAIR / post-contrast sequences (re-encoded here as lossless PNG)
• Mask: single tier — manual binary FLAIR-abnormality mask (no multi-rater tiers)

Splits

The LGG dataset has no official train/val/test split (the paper used cross-validation). This release ships a single train split; the EasyMedSeg dataloader maps val/test → train. patient_id + slice_idx are preserved so the 110 per-patient volumes can be reconstructed for 3D/sequence evaluation.

Schema

Column	Type	Description
image	Image	3-channel slice, RGB = pre-contrast / FLAIR / post-contrast
mask	Image	Binary FLAIR-abnormality mask (L mode, 0/255)
image_id	string	Full slice id, e.g. TCGA_HT_8018_19970411_14
patient_id	string	Source folder (carries acquisition date)
tcga_id	string	TCGA patient id, e.g. TCGA_HT_8018 — join key + BraTS cross-ref
institution	string	TCGA site code (CS / DU / EZ / FG / HT)
slice_idx	int32	Slice number parsed from the filename
num_slices	int32	Total slices for this patient (volume length)
has_tumor	bool	True iff the mask has any non-zero pixel

Genomic / clinical metadata

data.csv (at the repository root) carries the per-patient genomic-cluster and clinical table from the original release (RNASeq / Methylation / miRNA / CN / RPPA / Oncosign / COC clusters, histological grade, age, etc.). Join it to the parquet rows on tcga_id (its Patient column uses the same TCGA_XX_NNNN key).

Provenance

Raw imaging originates from TCIA's TCGA-LGG collection; the FLAIR-abnormality masks were produced by the Mazurowski lab (Duke) and distributed via Kaggle and GitHub (mateuszbuda/brain-segmentation-pytorch). Patient count (110) matches the paper. This mirror re-encodes the lossless source TIFFs as PNG inside parquet and does not otherwise modify the data.

License

CC BY-NC-SA 4.0 (per the Kaggle release metadata). Research, non-commercial, share-alike with attribution.

Citation

@article{Buda2019LGG,
  title   = {Association of genomic subtypes of lower-grade gliomas with shape
             features automatically extracted by a deep learning algorithm},
  author  = {Buda, Mateusz and Saha, Ashirbani and Mazurowski, Maciej A.},
  journal = {Computers in Biology and Medicine},
  volume  = {109},
  pages   = {218--225},
  year    = {2019},
  doi     = {10.1016/j.compbiomed.2019.05.002}
}