Vestibular-Schwannoma-SEG

Contrast-enhanced T1 (ceT1) and high-resolution T2 (hrT2) brain MRI of 242 consecutive patients with a unilateral sporadic vestibular schwannoma treated with Gamma Knife stereotactic radiosurgery (Queen Square Radiosurgery Centre, London, 2012–2018), with the clinical treatment-planning tumor segmentation and (for a subset) the cochlea.

This is a segmentation-ready NIfTI conversion of the TCIA collection Vestibular-Schwannoma-SEG (DOI 10.7937/TCIA.9YTJ-5Q73).

⚠️ Benchmark non-independence: CrossMoDA

All 242 patients of this collection are the source of the crossMoDA 2021/2022/2023 challenge data (MedOtter/crossmoda2021): the 105 labeled ceT1 + 105 unlabeled hrT2 training cases and the 32 validation cases of crossMoDA 2021 are exactly these patients (cropped/preprocessed, re-indexed). No public ID mapping between the two releases exists. Never treat this dataset and any CrossMoDA release as independent benchmarks. The related TCIA collection Vestibular-Schwannoma-MC-RC also shares 58 patients with this one.

Composition

Per patient one pre-treatment session on a 1.5T Siemens Avanto: ceT1 (MPRAGE, ~0.4×0.4×1.0–1.5 mm, 512×512) and hrT2 (3D CISS/FIESTA, ~0.5×0.5×1.0–1.5 mm). The split is the official VS_Seg baseline split (176/20/46 patients; KCL-BMEIS/VS_Seg params/split_TCIA.csv).

Missing series. The paper describes 242 patients × 2 scans = 484 images, but the live TCIA archive no longer serves 8 of the 1,936 original series. ceT1 is missing for cases 95, 111, 117, 135, 154, 188, 199 and hrT2 for case 133 (all in train/validation — the test split is complete). Case IDs vs_gk_N correspond to TCIA patient VS-SEG-{N:03d}; numbers 39, 97, 130, 160, 168, 208, 219, 227 do not exist upstream (242 of 1..250).

Labels

• Label 1 is the clinical planning segmentation: manually contoured in consensus by the treating neurosurgeon, neuroradiologist and physicist in Leksell GammaPlan. Tumor and cochlea are disjoint everywhere except case vs_gk_128 (≤2 touching voxels per frame, tumor takes precedence in the encoding).
• Cochlea absence means "not annotated", not "no cochlea" — the cochlea was only contoured for some patients. Use the per-sample has_cochlea flag; evaluate the cochlea only on samples where it is true.
• The upstream structure naming is heterogeneous (TV, AN, tv, tumour, …). The tumor is always the first structure of the case's contours.json — the same rule the official VS_Seg conversion uses. For cases 63 and 73 the planning volume is named Men upstream; it is included as the tumor per that rule. The original name is preserved in tumor_structure_name. Historical re-treatment volumes (Vol2016, Vol 2y, …) and other extra structures are not rasterized; their names are listed in each sample's meta.json.

Conversion provenance

• Images: DICOM series → NIfTI with SimpleITK, native geometry, no resampling.
• Masks: rasterized from the original non-interpolated planning contours (contours.json, provided by TCIA in both the T1 and T2 frame) directly into each MR volume's voxel grid — pixel-center even-odd polygon fill, plane-to-nearest-slice extrusion.
• QC: every mask was cross-rasterized from the independent DICOM RTSTRUCT encoding of the same annotation. Agreement Dice: tumor 0.9997 in its native ceT1 frame / 0.9369 in the transformed hrT2 frame; cochlea 0.9993 in its native hrT2 frame / 0.7922 in the transformed ceT1 frame. Each structure is exact in the frame it was drawn in; the residual is slice-regridding ambiguity in the other frame, where the RTSTRUCT re-export is itself interpolated. Per-sample values in meta.json (rtstruct_dice_*).
• The ITK registration matrices between the two frames (from the Leksell frame fiducials) are included under aux/, as are the original contours and the official split file.

Structure

dataset/vs_gk_NNN/t1/image.nii.gz   # ceT1 volume (int16, native geometry)
dataset/vs_gk_NNN/t1/mask.nii.gz    # uint8 {0,1,2} in the same grid
dataset/vs_gk_NNN/t1/meta.json      # per-sample metadata + QC
dataset/vs_gk_NNN/t2/...            # same for hrT2
train.jsonl / validation.jsonl / test.jsonl   # one row per sample (image/mask paths + metadata)
aux/contours.zip                    # original TCIA planning contours (both frames)
aux/matrices.zip                    # T1↔T2 ITK registration matrices
aux/split_TCIA.csv                  # official VS_Seg patient split

import json, SimpleITK as sitk
from huggingface_hub import hf_hub_download

row = json.loads(open(hf_hub_download("MedOtter/Vestibular-Schwannoma-SEG",
                                      "test.jsonl", repo_type="dataset")).readline())
img = sitk.ReadImage(hf_hub_download("MedOtter/Vestibular-Schwannoma-SEG",
                                     row["image"], repo_type="dataset"))

The preview config renders a representative slice (image / colored mask / overlay) per sample in the Dataset Viewer.

License and citation

CC BY 4.0 (all components; unchanged by TCIA's 2025 access-policy update — this collection remains fully public). Please cite:

Shapey, J., Kujawa, A., Dorent, R., Wang, G., Dimitriadis, A., Grishchuk, D., Paddick, I., Kitchen, N., Bradford, R., Saeed, S. R., Bisdas, S., Ourselin, S., & Vercauteren, T. (2021). Segmentation of Vestibular Schwannoma from Magnetic Resonance Imaging: An Open Annotated Dataset and Baseline Algorithm (Vestibular-Schwannoma-SEG) (Version 2) [Data set]. The Cancer Imaging Archive. https://doi.org/10.7937/TCIA.9YTJ-5Q73

Shapey, J. et al. Segmentation of vestibular schwannoma from MRI, an open annotated dataset and baseline algorithm. Sci Data 8, 286 (2021). https://doi.org/10.1038/s41597-021-01064-w

Clark, K. et al. The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository. J Digit Imaging 26, 1045–1057 (2013).