README.md

metadata

license: cc-by-3.0
task_categories:
  - image-segmentation
tags:
  - medical
  - mri
  - dce-mri
  - breast
  - breast-cancer
  - tumor-segmentation
  - tcia
  - tcga
pretty_name: TCGA-Breast-Radiogenomics
size_categories:
  - n<1K
configs:
  - config_name: preview
    data_files:
      - split: train
        path: preview/train-*
dataset_info:
  config_name: preview
  features:
    - name: patient_id
      dtype: string
    - name: image
      dtype: image
    - name: mask
      dtype: image
    - name: overlay
      dtype: image
    - name: split
      dtype: string
    - name: slice_index
      dtype: int32
    - name: num_slices
      dtype: int32
    - name: rows
      dtype: int32
    - name: cols
      dtype: int32
    - name: tumor_voxels
      dtype: int32
    - name: tumor_voxels_in_slice
      dtype: int32
    - name: series_description
      dtype: string
    - name: n_phases
      dtype: int32
    - name: breast_side
      dtype: string
    - name: tissue_source_site
      dtype: string
    - name: in_bcss
      dtype: bool
    - name: series_reassigned
      dtype: bool
  splits:
    - name: train
      num_bytes: 19932079
      num_examples: 91
  download_size: 19936139
  dataset_size: 19932079

TCGA-Breast-Radiogenomics

Whole-lesion breast tumour segmentation on dynamic contrast-enhanced (DCE) MRI. 91 patients from the TCGA-BRCA cohort, each with one binary mask of the primary invasive carcinoma, paired with its post-contrast source volume.

The upstream TCIA product is an analysis result: a 105 KB archive of masks in an undocumented .les format with no image reference of any kind, plus a set of spreadsheets. This mirror decodes those masks, resolves each one to its owning DICOM series, reconstructs the spatial volume, and ships both as voxel-aligned NIfTI — so none of that has to be repeated downstream.

Dataset Details

Field	Value
Modality	MRI, breast DCE (T1 post-contrast; GE VIBRANT / SAG 3D / BRAVA)
Body part	Breast
Task	Binary 3D segmentation (primary tumour)
Patients	91 (one study, one series, one lesion each)
Classes	1 + background
In-plane grid	512x512 (46 cases), 256x256 (45)
Slices per volume	41-172 (median 94)
In-plane spacing	0.55-0.94 mm
Field strength	1.5 T (GE Medical Systems, 90/91)
Format	NIfTI (.nii.gz), converted from DICOM + .les
License	CC BY 3.0 Unported (commercial use permitted)

Labels

Value	Structure
0	background
1	primary invasive breast carcinoma (whole lesion)

One lesion per patient — the masks are binary, with no multi-focal or multi-class structure. Lesion size ranges from 311 to 54,188 voxels (median 2,428), spread over 4-30 slices. This is a small-target segmentation problem: the median lesion occupies well under 0.1% of its volume.

Layout

volumes/TCGA-AO-A03M/image.nii.gz   # post-contrast DCE volume
volumes/TCGA-AO-A03M/mask.nii.gz    # uint8 binary mask, identical geometry
...
metadata.jsonl                      # one record per patient
label_map.json                      # label -> name

mask.nii.gz carries the same affine as its image.nii.gz, so the pair is voxel-aligned with no resampling.

Each metadata.jsonl record holds patient_id, image, mask, split, num_slices, rows, cols, spacing_xyz, tumor_voxels, tumor_slices, series_description, series_uid, n_phases, phase_index, acquisition parameters, the BI-RADS consensus reads (breast_side, birads_shape, birads_margin, birads_internal_enhancement, birads_fibroglandular, birads_background_enhancement), tissue_source_site, and the two provenance flags described below (in_bcss, series_reassigned).

Splits

There is no official train/val/test split. All 91 cases are published as a flat pool and are labelled train in metadata.jsonl. Any split is your own construction — group on patient_id.

How this mirror was built (and why you want it)

The upstream release cannot be consumed directly. Three steps were required, each verified against all 91 cases.

1. .les is Fortran-ordered — silently

Each .les file is a 12-byte header (6 x uint16 LE = an inclusive bounding box y_start, x_start, z_start, y_end, x_end, z_end) followed by (dy)(dx)(dz) uint8 voxels in {0, 1}. The body is column-major, because TCIA's reference reader is MATLAB.

A C-order reshape does not raise — it just returns a shredded mask:

reshape order	mean 3-D connected components	masks that are a single blob
C-order (numpy default)	182.0	0 / 91
Fortran order	1.04	87 / 91

2. The mask names no series — a separate spreadsheet does

.les files carry no SeriesInstanceUID, affine, spacing or frame of reference. Matching on slice count alone is hopeless: 0 of 91 patients have a unique candidate series (median 9 candidates each). The owning series is recoverable only from the SERIES_UID column of tcga-breast-radiologist-reads.xls, which the segmenting team used to pick "the sequence that corresponded to the one on which the radiologists annotated the truth". Across the 3 reviewers per patient: 68 unanimous, 19 resolved by 2-of-3 majority, 4 three-way ties.

3. DCE series are temporally interleaved

The designated series are multiphase acquisitions with up to 7 temporal phases in a single series (e.g. 410 files over 82 distinct slice positions = 5 phases). The spatial volume is recovered by grouping on ImagePositionPatient; only then does the mask's z index address the right slice. This mirror ships the first post-contrast phase (phase_index / n_phases record which was taken).

Voxel convention, resolved empirically

TCIA never documents whether .les axis 0 is the DICOM row or column, nor the slice direction. All 16 transpose/flip combinations were scored by whether they place the mask on enhancing voxels — a breast tumour on post-contrast DCE is bright relative to its immediate surroundings. The result is unambiguous:

convention	median lesion-vs-surround contrast
axis0=row, axis1=col, axis2=slice ascending, no flips	+0.895
next-best of the other 15	+0.073

The winner is positive in 25/25 probed patients; every alternative is indistinguishable from noise. Independently confirmed by laterality: the mask centroid falls in the breast recorded by the radiologists in 89/89 patients with a recorded side (2 patients are Not Applicable), at 64-114 mm off midline.

Four patients were reassigned to a different series

For 4 patients the reads-designated series is demonstrably wrong — the mask lands on negative contrast and/or in the wrong breast. All 4 were 2-of-3 majority votes, and 3 of them designate a 512x512 BRAVA while the patient also has a 256x256 VIBRANT DCE series that the mask fits perfectly. Every plausible alternative series was scored with the same test, excluding subtraction series (whose post-minus-pre construction makes any lesion trivially bright and would not be comparable to the other 87) and non-post-contrast sequences:

patient	designated (rejected)	contrast	reassigned to	contrast
TCGA-AO-A03V	T1 Axial post fat sat	-0.145	T1 Sagittal post fat sat	+0.617
TCGA-AR-A1AX	brava (1 Min.)	-0.284	VIBRANT	+0.932
TCGA-AR-A24S	BRAVA isotropic	+0.048	vibrant	+1.203
TCGA-AR-A24X	brava (1 Min.)	-0.029	VIBRANT	+0.700

These carry series_reassigned: true. The 4 three-way ties are unaffected — all 4 validated cleanly on their majority pick.

After reassignment: 91/91 cases have positive lesion contrast (median +0.955, minimum +0.337) and 89/89 agree on laterality.

Provenance and integrity notes

Official source, counts verified. Downloaded from TCIA directly (masks and spreadsheets over plain HTTPS; images via the unauthenticated NBIA REST API). No registration and no re-host is involved. All TCIA-stated figures reproduce exactly from the NBIA digest.

This is a 91-patient subset of TCGA-BRCA, which has 139 imaging patients; the other 48 have no segmentation. The 91 mask barcodes are a strict subset (91 inside, 0 outside). Note also that the associated papers analyse 84 cases (after a gene-expression filter) while TCIA released 91 masks.

"Radiogenomics" names the study, not the contents. The genomic data lives at the NCI GDC and is not included here; this package is imaging + masks. Join on the TCGA patient barcode to recover it.

Only the mask-bearing series are mirrored. The analysis result spans 1,129 series (36 GB) across these 91 patients, but only one series per patient carries a lesion. The other sequences (T2, pre-contrast, subtraction, DWI) are available from TCIA under the same barcode.

⚠️ Patient overlap with TCGA histopathology datasets

The cross-reference key is the TCGA patient barcode (TCGA-XX-XXXX), which is also the DICOM PatientID and the GDC submitter_id, and is preserved verbatim as patient_id / tcga_patient_barcode.

10 of these 91 patients also appear in BCSS, and the same slide pool feeds NuCLS and Pan-Cancer-Nuclei-Seg:

TCGA-AO-A12F, TCGA-AR-A1AQ, TCGA-BH-A0B3, TCGA-BH-A0BG, TCGA-BH-A0E0, TCGA-BH-A0RX, TCGA-E2-A150, TCGA-E2-A159, TCGA-E2-A1B6, TCGA-E2-A1L7

These carry in_bcss: true. No pixels are shared — that is H&E histopathology and this is MRI — so it is not a mask conflict, but the same humans appear on both sides of any multimodal split. Group on the barcode.

No overlap with Duke-Breast-Cancer-MRI (922 patients, Breast_MRI_### IDs, zero TCGA barcodes), I-SPY1/I-SPY2, ACRIN-6698, QIN-Breast, Breast-MRI-NACT-Pilot, BreastDM, RIDER, or the Medical Segmentation Decathlon (which has no breast task).

⚠️ MAMA-MIA is not a zero-shot baseline here. MAMA-MIA excludes TCGA-BRCA from its data, but its segmentation model was trained on 331 cases including 80 sagittal TCGA-BRCA cases. Those masks were never released, so MAMA-MIA is also not an alternative mask source for this cohort.

Citation

@misc{morris2014tcgabreastradiogenomics,
  title     = {Using Computer-extracted Image Phenotypes from Tumors on Breast
               {MRI} to Predict Stage},
  author    = {Morris, Elizabeth and Burnside, Elizabeth and Whitman, Gary and
               Zuley, Margarita and Bonaccio, Ermelinda and Ganott, Marie and
               Sutton, Elizabeth and Net, Jose and Brandt, Kathleen and
               Li, Hui and Drukker, Karen and Perou, Charles and Giger, Maryellen L.},
  year      = {2014},
  publisher = {The Cancer Imaging Archive},
  doi       = {10.7937/K9/TCIA.2014.8SIPIY6G}
}

@article{burnside2016usingcomputer,
  title   = {Using computer-extracted image phenotypes from tumors on breast
             magnetic resonance imaging to predict breast cancer pathologic stage},
  author  = {Burnside, Elizabeth S. and Drukker, Karen and Li, Hui and
             Bonaccio, Ermelinda and Zuley, Margarita and Ganott, Marie and
             Net, Jose M. and Sutton, Elizabeth J. and Brandt, Kathleen R. and
             Whitman, Gary J. and Conzen, Suzanne D. and Lan, Li and
             Ljung, Britt-Marie and Morris, Elizabeth A. and Perou, Charles M. and
             Giger, Maryellen L.},
  journal = {Cancer},
  volume  = {122},
  number  = {5},
  pages   = {748--757},
  year    = {2016},
  doi     = {10.1002/cncr.29791}
}

@misc{lingle2016tcgabrca,
  title     = {The Cancer Genome Atlas Breast Invasive Carcinoma Collection
               ({TCGA-BRCA})},
  author    = {Lingle, W. and Erickson, B. J. and Zuley, M. L. and Jarosz, R. and
               Bonaccio, E. and Filippini, J. and Net, J. M. and Levi, L. and
               Morris, E. A. and Figler, G. G. and Elnajjar, P. and Kirk, S. and
               Lee, Y. and Giger, M. and Gruszauskas, N.},
  year      = {2016},
  publisher = {The Cancer Imaging Archive},
  doi       = {10.7937/K9/TCIA.2016.AB2NAZRP}
}

@article{clark2013tcia,
  title   = {The Cancer Imaging Archive ({TCIA}): Maintaining and Operating a
             Public Information Repository},
  author  = {Clark, Kenneth and Vendt, Bruce and Smith, Kirk and others},
  journal = {Journal of Digital Imaging},
  volume  = {26},
  number  = {6},
  pages   = {1045--1057},
  year    = {2013},
  doi     = {10.1007/s10278-013-9622-7}
}