README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- image-segmentation
tags:
- medical
- MRI
- segmentation
- CrossMoDA2022
- vestibular-schwannoma
size_categories:
- n<1K
pretty_name: crossMoDA 2022 (labeled source-training arm, ceT1)
dataset_info:
  features:
  - name: patient_id
    dtype: string
  - name: institution
    dtype: string
  - name: mri_sequence
    dtype: string
  - name: in_crossmoda2021
    dtype: bool
  - name: crossmoda2021_id
    dtype: string
  - name: koos
    dtype: int32
  - name: post_operative
    dtype: bool
  - name: num_slices
    dtype: int32
  - name: shape
    dtype: string
  - name: image
    dtype: image
  - name: mask
    dtype: image
  - name: overlay
    dtype: image
  splits:
  - name: train
    num_bytes: 53266199
    num_examples: 210
  download_size: 53263643
  dataset_size: 53266199
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# crossMoDA 2022 Dataset — labeled source-training arm

MRI from the MICCAI 2022 Cross-Modality Domain Adaptation (crossMoDA) challenge:
vestibular schwannoma (VS) and cochlea segmentation. This mirror contains the
**210 contrast-enhanced T1 (ceT1) scans with ground-truth masks** — the
challenge's labeled *source-domain training* arm — from both 2022 institutions:

| Institution | Cases | Filename stems | Native geometry |
|---|---|---|---|
| London (SC-GK) | 105 | `crossmoda2021_ldn_{1..105}` | 512x512x120 @ 0.41x0.41x1.5 mm (Siemens Avanto 1.5T) |
| Tilburg (ETZ) | 105 | `crossmoda2022_etz_{0..104}` | 256x256x50 @ 0.82x0.82x1.5 mm (Philips Ingenia 1.5T) |

Volumes are in native per-site geometry — not resampled, not co-registered.

## Not comparable to the challenge leaderboard (ceT1 != hrT2)

These are **source-domain ceT1** scans. The crossMoDA competition evaluates on
the **hrT2 target domain**, whose validation/test ground truth is withheld
(server-side scoring on grand-challenge.org). Scores computed on this repo are
valid measurements on ceT1 but are **not comparable** to crossMoDA leaderboard
numbers. Each row's `mri_sequence` column records this in the data itself.

## Overlap with crossMoDA 2021 — machine-readable

The London arm is a **voxel-identical duplicate** of
[MedOtter/crossmoda2021](https://huggingface.co/datasets/MedOtter/crossmoda2021):
all 210 London file pairs (105 ceT1 scans AND 105 label masks) match at
decompressed-content level (md5 of the full NIfTI stream, verified 2026-08-07).
The 2022 release only renamed `crossmoda_N_*` to `crossmoda2021_ldn_N_*`; the
outlines were **not** re-drawn. Every `train.jsonl` row therefore carries:

- `in_crossmoda2021` (bool) — `true` for all 105 London cases
- `crossmoda2021_id` (string|null) — the exact 2021 stem (`crossmoda_N`), `null` for Tilburg
- `institution` (`"London"` / `"Tilburg"`)

Do **not** evaluate this repo and `crossmoda2021` as independent benchmarks.
For a leak-free benchmark restricted to genuinely new data, filter
`institution == "Tilburg"`.

Lineage: the London cases descend from the public TCIA collection
*Vestibular-Schwannoma-SEG* (CC BY 4.0, DOI 10.7937/TCIA.9YTJ-5Q73);
`institution == "London"` is the collection-level marker for that lineage.
crossMoDA 2023 (Synapse `syn51236108`) is a superset of this material.

## Omitted from this mirror (by design)

The official Zenodo release ([record 6504722](https://zenodo.org/records/6504722))
additionally contains 210 unlabeled hrT2 training-target scans, 64 validation
hrT2 scans (GT withheld), and GIF auto-parcellations (algorithm outputs, not
ground truth); the 271-scan test set was never publicly released. None of those
carry usable public ground truth, so this mirror hosts only the labeled arm.
Fetch the omitted parts from Zenodo if you need them.

## Labels

`*_Label.nii.gz`: 0 = background, 1 = vestibular schwannoma,
2 = cochlea (**both cochleae share label 2**). VS was contoured by the treating
neurosurgeon and physicist (radiosurgery planning); cochleae by radiology
fellows. Single annotation tier — these masks are the challenge gold standard.

## train.jsonl columns

| Column | Type | Meaning |
|---|---|---|
| image / mask | string | repo-relative path prefixed `data/nii/CrossMoDA2022/` (loader convention) |
| label | list[str] | ["vestibular schwannoma", "cochlea"] |
| modality / dataset / official_split | string | "MRI" / "CrossMoDA2022" / "train" |
| patient_id | string | filename stem, e.g. `crossmoda2022_etz_0` |
| institution | string | "London" or "Tilburg" |
| mri_sequence | string | always "ceT1" (see leaderboard caveat) |
| in_crossmoda2021 | bool | duplicate-of-2021 marker |
| crossmoda2021_id | string/null | 2021 stem for London, null for Tilburg |
| koos | int/null | Koos grade 1-4 (challenge Task 2); null for post-operative cases |
| post_operative | bool | 28 London cases; excluded from challenge Task 2 |

Koos distribution: 1: 12, 2: 51, 3: 73, 4: 46, post-operative: 28
(`infos_source_training.csv` is the original organizer file).

## License

**CC BY-NC-SA 4.0**, per the organizers' challenge page and the Zenodo record
description ("all data"). Note: Zenodo's machine-readable license field
auto-tags `cc-by-4.0` — a known metadata inconsistency; the NC-SA statement
governs. The London subset is separately available upstream under CC BY 4.0
via TCIA (Vestibular-Schwannoma-SEG).

## Citation

- Wijethilake, Dorent, Ivory, Kujawa, Shapey, Cornelissen, Langenhuizen, et al.
  "crossMoDA Challenge: Evolution of Cross-Modality Domain Adaptation
  Techniques for Vestibular Schwannoma and Cochlea Segmentation from 2021 to
  2023." arXiv:2506.12006
- Dorent et al. "CrossMoDA 2021 challenge." Medical Image Analysis 83:102628,
  2023. DOI 10.1016/j.media.2022.102628
- Shapey et al. "Segmentation of vestibular schwannoma from MRI..."
  Scientific Data 8:286, 2021. Data DOI 10.7937/TCIA.9YTJ-5Q73