IvyGAP-Radiomics (SRI-atlas subset)

Multi-parametric brain-MRI glioblastoma (GBM) segmentation dataset: 4 co-registered, skull-stripped MRI sequences per subject with expert tumor sub-compartment masks from two independent institutions, plus precomputed radiomic features. From the TCIA Analysis Result collection IvyGAP-Radiomics (Pati et al., 2020).

⚠️ Scope — this is a preprocessed subset of the full Ivy GAP project

"IvyGAP-Radiomics" is not the complete Ivy Glioblastoma Atlas Project (Ivy GAP). It is a derived analysis result containing only the pre-operative scans, skull-stripped and co-registered to the SRI24 atlas, converted to NIfTI, with expert tumor masks and radiomic features added. It does not include the original DICOM imaging (~140 GB), histology (ISH), or gene-expression / genomic data of the parent Ivy GAP collection. Only the SRI-atlas package is mirrored here (the alternate MNI-atlas / CWRU-only package is not included).

🔴 Cross-dataset overlap with BraTS (evaluation-integrity hazard)

Pre-operative IvyGAP data is included in the BraTS challenge training set (together with TCGA-GBM, TCGA-LGG, CPTAC-GBM), and these masks were produced with the BraTS preprocessing pipeline (SRI24 atlas, skull-strip) and follow the BraTS labeling convention. Do not benchmark these subjects against any BraTS-trained model, or alongside Angelou0516/brats2023-*, without treating the results as non-held-out — there is likely train/test leakage. Subjects are identified by Ivy GAP IDs (W1…W55); there is no published BraTS-ID cross-reference column, so any mapping to BraTS subjects must be done externally via the Ivy GAP → BraTS name tables.

Two expert raters — no single gold standard (by design)

The dataset's purpose is inter-rater reproducibility. Each tumor was segmented independently by board-certified neuroradiologists at two institutions:

• UPenn (Hospital of the University of Pennsylvania) — 34 subjects
• CWRU (Case Western Reserve University) — 31 subjects

31 subjects are paired (annotated by both). The paper anoints neither rater as "the" ground truth — both are equally-valid expert annotations. Both mask sets are provided here. For a single-GT workflow, the recommended default is UPenn (broader coverage: 34 vs 31). Subjects without a CWRU mask: W6, W30, W54.

Labels (BraTS convention)

value	sub-compartment
0	background
1	NCR/NET — necrotic & non-enhancing tumor core
2	ED — peritumoral edema / invaded tissue
4	ET — enhancing tumor

Composite regions: TC (tumor core) = {1,4}; WT (whole tumor) = {1,2,4}; ET = {4}. Mask dtypes as released: UPenn uint16, CWRU float32 — both encode the integer labels above (content is preserved byte-for-byte; cast to integer when loading).

Geometry

All volumes are 240 × 240 × 155 @ 1.0 mm isotropic in SRI24 space, skull-stripped, LPS orientation. One pre-operative study per subject.

Structure

W{n}/                           # one folder per subject, at the repo root
  W{n}_t1.nii.gz                # native T1
  W{n}_t1c.nii.gz               # post-contrast T1 (T1-Gd; original token "t1gd")
  W{n}_t2.nii.gz                # T2
  W{n}_flair.nii.gz             # T2-FLAIR
  W{n}_seg-UPenn.nii.gz         # UPenn expert mask (all 34 subjects)
  W{n}_seg-CWRU.nii.gz          # CWRU expert mask (31 subjects)
atlas/spgr_unstrip_lps.nii.gz   # SRI24 reference template
data/test-*.parquet             # Dataset Viewer preview only (rendered slices + metadata)
radiomic_features/              # CaPTk/IBSI features per rater, feature parameters, reproducibility correlations
ivygap_metadata.csv             # per-subject clinical & molecular metadata (see below)
subject_study_dates.csv         # subject -> original study date + which rater masks exist

ivygap_metadata.csv

Per-subject metadata keyed by Patient (W-ID). Columns include Included_Upenn, Included_CWRU, 4_Modalities (inclusion flags), Age, gender, Histopathology, location, EoR (extent of resection), Surgery, survival_days, Molecular_subtype, IDH1, 1p19q_deletion, MGMT/MGMT PCR, EGFR/EGFR vIII, PTEN, KPS(initial), time-to-progression / last-follow-up, and cause of death. (The CSV also lists a few subjects that lack 4-modality imaging and are therefore not present under data/.)

Provenance, license & citation

• Provenance: Official TCIA Analysis Result (author-provided), DOI 10.7937/9j41-7d44. SRI-atlas package downloaded via TCIA IBM-Aspera faspex5.
• License: CC BY 3.0 (TCIA Analysis Results) + the TCIA Data Usage Policy.
• Data citation: Pati, S., Verma, R., Akbari, H., Bilello, M., Hill, V.B., Sako, C., Correa, R., Beig, N., Venet, L., Thakur, S., Serai, P., Ha, S.M., Blake, G.D., Shinohara, R.T., Tiwari, P., Bakas, S. (2020). Data from the Multi-Institutional Paired Expert Segmentations and Radiomic Features of the Ivy GAP Dataset. The Cancer Imaging Archive. DOI 10.7937/9j41-7d44.
• Publication: Pati S, et al. Reproducibility analysis of multi-institutional paired expert annotations and radiomic features of the Ivy GAP dataset. Medical Physics 2020;47(12):6039–6052. DOI 10.1002/mp.14556.