MedOtter/Shifts-2022-Part-2

Shifts 2022 — Part 2 (MS White-Matter Lesion Segmentation, open CC subset)

Brain MRI multiple-sclerosis (MS) white-matter lesion segmentation dataset from the Shifts 2.0 benchmark (Malinin et al., NeurIPS 2022 Datasets & Benchmarks).

⚠️ What "Part 2" means (faithful-naming note). The Shifts MS lesion dataset is split across two Zenodo archives purely by license/access, not by task — both parts are the same MS lesion segmentation task:

• Part 1 (Zenodo 7051658) = the MSSEG-1 / OFSEP cohorts (Rennes, Bordeaux, Lyon, 52 cases). It is released only under an OFSEP Data Usage Agreement and cannot be redistributed — it is NOT included here.
• Part 2 (Zenodo 7051692, this repo) = the openly CC-licensed cohorts: Best (ISBI 2015) + Ljubljana (PubMRI) = 46 released cases.
• The shifted evaluation cohort Lausanne (74 cases, eval_out) is never released publicly (Grand-Challenge Docker leaderboard only) and is likewise not here.

This repo is therefore the open, redistributable 46-case subset of the Shifts MS benchmark, faithful to Zenodo record 7051692.

Contents

Cohort	Source dataset	Shifts split	Cases	Modalities
best	ISBI 2015 Longitudinal MS	train	10	FLAIR, T1, T2, PD
best	ISBI 2015 Longitudinal MS	dev_in	2	FLAIR, T1, T2, PD
best	ISBI 2015 Longitudinal MS	eval_in	9	FLAIR, T1, T2, PD
ljubljana	PubMRI (Lesjak 2018)	dev_out	25	FLAIR, T1, T2, T1ce
		Total	46

• best = in-domain data (ISBI 2015); ljubljana = the distribution-shifted dev set (PubMRI).
• All volumes preprocessed to 1 mm isotropic: denoised, registered to FLAIR space, skull-stripped (brain mask from T1), N4 bias-field corrected. Images linearly interpolated; masks nearest-neighbour interpolated.

Ground truth

• gt/ — the challenge gold-standard consensus lesion mask (binary {0,1}), one per case.
  • Ljubljana / PubMRI: consensus of 3 expert raters.
  • Best / ISBI: the release designates annotator 2 (the more experienced rater) as GT; both raters' individual masks are also provided under individual_annotators/.
• fg_mask/ — brain/foreground mask (binary {0,1}), used to restrict error-retention curves to brain tissue.
• Challenge metric is the lesion-load-decorrelated normalized Dice (nDSC).

Repository layout

(repo root)
  best/{train,dev_in,eval_in}/
    flair/  t1/  t2/  pd/  gt/  fg_mask/
    individual_annotators/{annotator1,annotator2}/
  ljubljana/dev_out/
    flair/  t1/  t2/  t1ce/  gt/  fg_mask/
  index/
    {train,dev_in,eval_in,dev_out,all}.jsonl   # per-case path index (added by this mirror)
  README.txt   # original author README
  LICENSE.md   # original author license notice

All index/*.jsonl paths are repo-root relative (e.g. best/train/flair/24_FLAIR_isovox.nii.gz). Files are named {subject}_{MODALITY}_isovox.nii.gz (GT: {subject}_gt_isovox.nii.gz, foreground: {subject}_isovox_fg_mask.nii.gz). Subject IDs are unique within a cohort/split but collide across cohorts — use the case_id ({cohort}/{split}/{subject}) from the index as the global key.

Index records (index/*.jsonl)

Each line: case_id, cohort, split, subject, source_dataset, flair, t1, t2, pd, t1ce, gt, fg_mask [, individual_annotators]. Modality fields are null where absent (PD only in best, T1ce only in ljubljana).

⚠️ Cross-dataset overlap (leakage risk)

Part 2 physically reuses two public MS datasets. If any of these enter a benchmark suite separately, dedup against the Shifts cohort using the source_dataset / cohort tag before co-evaluating:

• Best cohort ≡ ISBI 2015 Longitudinal MS Lesion Segmentation Challenge (Carass et al., 2017).
• Ljubljana cohort ≡ PubMRI (Lesjak et al., Neuroinformatics 2018).
• (Part 1, not here) reuses MSSEG-1 / MICCAI 2016.

License

CC BY-NC-SA 4.0 (verified on Zenodo record 7051692 and in the paper). Non-commercial; share-alike; attribution required.

Citation

@inproceedings{malinin2022shifts2,
  title     = {Shifts 2.0: Extending The Dataset of Real Distributional Shifts},
  author    = {Malinin, Andrey and Athanasopoulos, Andreas and Barakovic, Muhamed and
               Bach Cuadra, Meritxell and Gales, Mark J. F. and Granziera, Cristina and others},
  booktitle = {NeurIPS Datasets and Benchmarks},
  year      = {2022},
  note      = {arXiv:2206.15407}
}

Source: Zenodo record 7051692 · Project: shifts.ai · Paper: arXiv:2206.15407