Soft-tissue-Sarcoma (STS)

A TCIA collection of 51 patients with histologically proven soft-tissue sarcoma of the extremities, each imaged with joint pre-treatment FDG-PET/CT and MRI and contoured by an expert radiation oncologist. Collected at McGill University Health Centre (Montreal) and published with Vallières et al., Phys Med Biol 2015.

The original study built a radiomics model predicting lung metastases from joint PET/MRI texture features; 19 of the 51 patients developed lung metastases.

Dataset Details

Field	Value
Modalities	FDG-PET, CT, MRI (T1 + T2FS-or-STIR) + DICOM RTSTRUCT contours
Body part	EXTREMITY (all 612 series) — 28/51 thigh, remainder other limb/girdle sites
Task	3D tumour segmentation (single foreground structure)
Patients	51 — STS_001 … STS_051, contiguous, no gaps
Studies	102 — exactly 2 per patient (one MRI session, one PET/CT session)
Series	612 = 51 CT + 51 PT + 204 MR + 306 RTSTRUCT (exactly 12 per patient)
Images	38,283 DICOM
Size	9.19 GiB (TCIA quotes 9.87 GB)
Scanners	PET/CT: GE Discovery ST. MRI: GE / Philips / Siemens / Varian (multi-vendor)
License	CC BY 3.0 Unported — commercial use permitted
DOI	10.7937/K9/TCIA.2015.7GO2GSKS
Registration	None — fully public

There is no official train/val/test split and no named subsets. Splitting is left to the consumer. Natural strata that do exist: lung metastases 19 / 32, T2FS 26 / STIR 25, GTV_Edema present 32 / absent 19.

Per-patient structure — 6 image series + 6 RTSTRUCT

Every one of the 51 patients has exactly the same 12 series, in two frames of reference:

Frame	Image series	Paired RTSTRUCT	n
Native MRI	MR T1	RTstruct_T1	51
Native MRI	MR T2FS or STIR	RTstruct_T2FS / RTstruct_STIR	26 / 25
PET/CT	CT	RTstruct_CT	51
PET/CT	PT (FDG-PET)	RTstruct_PET	51
PET/CT (crop)	Aligned_T1toPET_BOX	RTstruct_AlignedT1toPET	51
PET/CT (crop)	Aligned_{T2FS,STIR}toPET_BOX	RTstruct_Aligned{T2FS,STIR}toPET	26 / 25

T2FS and STIR are mutually exclusive and partition all 51 patients. The paper treats STIR as a fallback under the "T2FS" umbrella: "When T2-weighted fat-saturated scans were not available, STIR scans were used (n = 25)."

Ground truth — one hand-drawn tier, five propagated

There is a single annotator (an expert radiation oncologist), so there is no inter-rater ambiguity. What varies is drawn versus derived:

• Gold / primary: RTstruct_T2FS (26) and RTstruct_STIR (25) — contours were manually drawn slice-by-slice on the T2FS/STIR scans. This is the tier the paper's features were extracted from.
• Propagated: the other four tiers (RTstruct_T1, RTstruct_CT, RTstruct_PET, RTstruct_Aligned*toPET) are MIM Software rigid-registration propagations of those contours onto the other grids.

⚠️ ROIGenerationAlgorithm = MANUAL on all 306 objects, including the propagated ones — the DICOM tag does not distinguish drawn from derived. Use the tier table above, not the tag.

⚠️ The PET/CT-frame masks cross a large time gap. The MRI and PET/CT sessions are a median 21 days apart (mean 20.5; range 0–63 days; only 3 same-day; 32/51 more than 14 days; 10/51 more than 30 days; the PET/CT session never precedes the MRI). The PET/CT-frame contours are rigid propagations across that much repositioning, so they are geometrically approximate. Neubauer et al. (MICCAI 2020) found it necessary to have a nuclear medicine physician re-delineate on PET rather than trust them.

Two ROIs, and they are nested

ROI	Patients	Meaning
GTV_Mass	51 (all)	tumour mass, excluding peritumoural edema — the paper's reference target
GTV_Edema	32	tumour including peritumoural edema — secondary, used in the paper only to probe segmentation-uncertainty

⚠️ GTV_Edema is a nested superset of GTV_Mass, not a disjoint class. Across all 32 patients the in-plane area ratio is ≥ 1.031 (median 1.29, max 8.74), and the edema slice set is a superset of the mass slice set for 31/32 — the sole exception, STS_021, has exactly one mass slice with no edema contour.

Build these as two independent binary targets, never as a {0,1,2} label map. A label map assigns "edema" only to the rim annulus GTV_Edema \ GTV_Mass, which is not a structure anyone annotated and scores meaninglessly.

Loader gotchas — verified against all 612 series and all 306 RTSTRUCT objects

These are easy to get silently wrong. series_to_patient.json and pairs.json (shipped at repo root) pre-resolve all of them.

1. SeriesDescription is typo-ridden. RTSTRUCT carries RTStruct_CT, RTStruct_PET, RTstruct_AlignedT!toPET (bang instead of a 1), RTstruct_Aligned_T1toPET / _STIRtoPET (stray underscore), plus stray _BOX suffixes; the MR side has Alligned_T1toPET_BOX (double L). After case-insensitive normalisation the counts land exactly on 51/51/51/26/25/51/26/25 = 306.

2. Native-MR T1-vs-T2FS/STIR is NOT readable from the description. The 102 native MR series carry free-text clinical protocol names — AX STIR, Axial FSE/T2 Fatsat, KNEE *AXT2SP, eT1W_TSE_ax SENSE W PICT-PLUS, 2. AXIAL T1 BOTH LEGS - RESEARCH … Resolve the sequence by following the normalised RTSTRUCT back through ReferencedFrameOfReferenceSequence to the series it references. pairs.json has already done this.

3. The other 102 MR series are MIM-derived crops, distinguishable only by Manufacturer containing MIM Software (e.g. SIEMENS / MIM Software). The Aligned_*_BOX volumes are tumour-centred crops, not full FOV (e.g. 153×123×60), though self-consistent — 0% out-of-bounds contour vertices.

4. ROI order is not stable. 12 of the 192 two-ROI objects list GTV_Edema first. Map by ROIName, never by ROINumber or sequence index.

5. STS_046 renames its ROIs. On its PET/CT-frame RTSTRUCTs the names are GTV_Research (= mass) and GTV_Res+edema (= edema). Hard-matching the literal string "GTV_Mass" yields a silently empty mask for that patient. Match with an alias set.

6. Never group contours by z — the gold scans are not all axial. Of the 51 annotated T2FS/STIR scans, 45 are axial, 4 coronal and 2 sagittal (for the 32 edema patients specifically: 27 / 4 / 1), and contours are slightly oblique. Use ReferencedSOPInstanceUID — it is present on every contour, and every referenced SOP exists in its series.

7. Foreground is sparse on the PET/CT grid, but not on the gold tier. GTV_Mass covers a median of only 13.6% of CT slices (min 5.6%, max 36.3%; 14/51 patients under 10%), so a debug_n_samples: 1 smoke test against the PET/CT frame will look degenerate through no fault of the loader. The gold T2FS/STIR tier is not sparse — median 50.0% slice coverage (min 20.0%, max 90.0%, none under 15%), so single-sample debug runs on the default target behave normally.

8. Frames of reference. T1 and T2FS/STIR share one FrameOfReferenceUID (same MRI session), so T1↔T2FS propagation is geometrically trivial; the PET/CT frame is separate.

9. Dates are shifted for de-identification (intervals preserved), so StudyDate values are not real. 84 of the 612 series carry DateReleased = 2020-05-21 — a partial resubmission over the 2015 release.

Cross-dataset overlap

• ⚠️ SAROS shares 6 patients: STS_001, STS_007, STS_008, STS_026, STS_040, STS_044 → SAROS case_815, case_636, case_678, case_714, case_703, case_753. The cross-reference lives in TCIA's Segmentation-Info CSV under tcia_case_id. SAROS labels body regions, not tumour, so the risk is CT-image contamination rather than tumour-label leakage.
• ⚠️ MSTT-199 uses all 51 of these patients as its external test set (reported Dice 0.79 U-Net / 0.80 LiteMedSAM). Do not train on STS and evaluate on MSTT-199.
• No overlap with MSD (no sarcoma task), the BraTS family, MedSAM's training corpus, SA-Med2D-20M, autoPET, HECKTOR (same PI and lab, different patients and disease), QIN-SARCOMA (OHSU, DCE-MRI only, no segmentations), or TCGA-SARC (disjoint cohort — TCGA-SARC has no McGill/Montreal site, and its TCIA arm is 5 TCGA-QQ-* patients).
• ⚠️ INFOclinical_STS.xlsx has no cross-reference ID column. Its MSKCC type column is a histological classification scheme (Liposarcoma / Leiomyosarcoma / MFH / Other) — not a link to MSKCC or TCGA.

Provenance

Official, author-deposited TCIA collection. All 51 patients from the paper are present — no count mismatch. Third-party re-hosts exist and all of them misstate the license or the content; prefer this mirror or TCIA directly:

Re-host	Problem
Kaggle 4quant/soft-tissue-sarcoma	HDF5 at 5 mm isotropic, PET/CT only — no MRI; ~4% of the original
reasat/sts-reg	NIfTI, T1→T2 registered; claims CC0 (understates CC BY 3.0)
husnainrasool/tcia-sarcoma-ds	verbatim re-upload but claims MIT (wrong license class)
hyper.ai listing	labels it Non-Commercial — incorrect; CC BY 3.0 permits commercial use

Faithful-naming disclosures for this mirror:

• 25 of the 51 "T2FS" scans are actually STIR — the paper's own umbrella term, not a mirroring error.
• Half the MR series (102/204) are derived MIM registered crops, not raw acquisitions. This is a raw + derived mix, exactly as TCIA publishes it.
• Body part is tagged EXTREMITY throughout, but roughly 10/51 primary sites are girdle/trunk (buttock, pelvis, groin, parascapular) rather than true limb.

Structure

images/<PatientID>/<SeriesInstanceUID>/*.dcm         # CT, PT, MR — 306 series
segmentations/<PatientID>/<SeriesInstanceUID>/*.dcm  # RTSTRUCT — 306 objects
clinical/INFOclinical_STS.xlsx                       # demographics + outcome vector
clinical/Soft-tissue-Sarcoma-nbia-digest.xlsx        # official NBIA series digest
series_to_patient.json                               # all 612 series, normalised `role`
pairs.json                                           # RTSTRUCT -> image series + ROI names
LICENSE.txt

series_to_patient.json keys each SeriesInstanceUID to PatientID, StudyInstanceUID, Modality, SeriesDescription, a normalised role (ct, pet, mr_t1, mr_t2fs, mr_stir, aligned_t1_to_pet, aligned_t2fs_to_pet, aligned_stir_to_pet), Manufacturer, ImageCount, FileSize, license/DOI and the relative path.

pairs.json keys each RTSTRUCT SeriesInstanceUID to the image_series_uid it is drawn on (resolved from ReferencedFrameOfReferenceSequence, not from the typo-ridden description), frame_of_reference_uid, the literal roi_names, roi_names_with_contours, n_referenced_sop, and the boolean flags has_gtv_mass, has_gtv_edema and uses_alias_roi_names. Pairing therefore needs no TCIA round-trip and no description parsing.

Benchmark reference point

Neubauer et al., Soft Tissue Sarcoma Co-Segmentation in Combined MRI and PET/CT Data (MICCAI MMMI 2020, arXiv:2008.12544): 39/51 patients after cropping to the leg region, 5-fold CV — best T2 Dice 77.2 ± 16.5%, PET Dice 74.6 ± 19.0%, T2-only baseline 65.6 ± 24.0%. They resampled in-plane to 0.75 mm and kept the native T2 slice distance to avoid interpolation artifacts.

Source & Citation

• TCIA collection: https://www.cancerimagingarchive.net/collection/soft-tissue-sarcoma/
• DOI: 10.7937/K9/TCIA.2015.7GO2GSKS

@article{vallieres2015sts,
  author  = {Valli{\`e}res, Martin and Freeman, Carolyn R. and
             Skamene, Sonia R. and El Naqa, Issam},
  title   = {A radiomics model from joint {FDG-PET} and {MRI} texture features
             for the prediction of lung metastases in soft-tissue sarcomas of
             the extremities},
  journal = {Physics in Medicine and Biology},
  volume  = {60},
  number  = {14},
  pages   = {5471--5496},
  year    = {2015},
  doi     = {10.1088/0031-9155/60/14/5471}
}

@misc{vallieres2015stsdata,
  author    = {Valli{\`e}res, M. and Freeman, C. R. and Skamene, S. R. and
               El Naqa, I.},
  title     = {A radiomics model from joint {FDG-PET} and {MRI} texture features
               for the prediction of lung metastases in soft-tissue sarcomas of
               the extremities [Data set]},
  year      = {2015},
  publisher = {The Cancer Imaging Archive},
  doi       = {10.7937/K9/TCIA.2015.7GO2GSKS}
}

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