VALDO — "Where is VALDO?" VAscular Lesions DetectiOn and segmentatiOn challenge (MICCAI 2021)

⚠️ This repository is the publicly-released training set only of VALDO 2021. The challenge had 431 cases total (152 train + 279 test); the 279-case test set was never publicly distributed (the organisers evaluate submitted Docker containers on a held-out server). What you see here is the 152-case training release from Zenodo — the only portion the authors made openly downloadable.

⚠️ Preprocessed-for-challenge variant, not raw acquisitions. All images are defaced, and within each task the non-reference sequences are rigidly/affinely co-registered and resampled to the task's reference space (T1 for Tasks 1 & 3, T2* for Task 2).

VALDO targets three imaging markers of cerebral small vessel disease (CSVD) on brain MRI, organised as three independent tasks. Data is drawn from three population cohorts: SABRE (London), RSS = Rotterdam Scan Study (part of the Rotterdam Study), and ALFA (BarcelonaBeta, Task 2 only).

Tasks & contents

Layout is BIDS-like: Task{1,2,3}/sub-XXX/ with files named sub-{ID}_space-{Space}_desc-{Description}_{Suffix}.nii.gz.

Task 1 — Enlarged Perivascular Spaces (EPVS) · 40 subjects · T1 space

Suffix	Role
desc-masked_T1 / desc-masked_T2 / desc-masked_FLAIR	images (T2/FLAIR co-registered to T1, defaced)
desc-masked_Regions	region mask: each region a distinct integer, keyed to Count.csv
Count.csv	per-region EPVS counts (value, count, min_slice, max_slice; slices 0-indexed)
desc-Rater{1,2}_PVSSeg / PVSSeg	EPVS segmentation — present for only 12 of 40 subjects

Heterogeneous / weak-label task. Of 40 subjects: 6 (SABRE, sub-101…106) have voxel segmentations from two raters (desc-Rater1/Rater2_PVSSeg); 6 (RSS) have a single PVSSeg; the remaining 28 (RSS) have counts only (Count.csv + Regions, no segmentation mask). A segmentation loader must filter to the 12 subjects with a PVSSeg file; the 28 counts-only cases are a deliberate weak-supervision tier.

Task 2 — Cerebral Microbleeds (CMB) · 72 subjects · T2* space

Suffix	Role
desc-masked_T2S / desc-masked_T2 / desc-masked_T1	images (T2/T1 co-registered to T2*, defaced)
CMB	microbleed segmentation — 70 of 72 subjects carry a mask

2 subjects have no CMB file (treat as zero-microbleed / empty-mask cases). Cohorts: SABRE, RSS, ALFA.

Task 3 — Lacunes · 40 subjects · T1 space

Suffix	Role
desc-masked_T1 / desc-masked_T2 / desc-masked_FLAIR	images (T2/FLAIR co-registered to T1, defaced)
desc-Rater{1,2,3,4}_Lacunes	lacune segmentation — two raters per subject

Every subject has exactly two rater masks, but the rater identity differs by cohort: SABRE (sub-101…106) uses Rater1+Rater2; RSS (sub-201+) uses Rater3+Rater4. A loader must pick whichever two *_Lacunes files exist per subject — do not hardcode Rater1/Rater2. One subject (sub-105) additionally carries a sub-105_space-T1_possible.nii.gz ("possible lacunes") auxiliary mask.

Gold-standard ground truth (paper's consensus rule)

Where two rater masks exist (Task 3 all cases; Task 1 SABRE sub-101…106), the reference standard is the average of the two raters' masks, thresholded at 0.5 (binary). Where a single mask exists (Task 1 RSS single PVSSeg; Task 2 CMB), that mask is the ground truth. All rater masks are preserved here so this rule can be applied — or revised — downstream.

⚠️ Cross-cohort lineage (leakage note)

VALDO subjects come from SABRE, RSS (Rotterdam Scan Study), and ALFA population cohorts. There is no overlap with tumor/oncology benchmarks (BraTS family, Medical Segmentation Decathlon, TCIA collections). The one lineage to watch is RSS / the Rotterdam Study, which contributes to many neuroimaging datasets — de-duplicate before combining VALDO with any other Rotterdam-derived set. No cross-reference ID column is exposed: subject IDs are challenge-local (sub-###) and original cohort IDs are not released, so subject-level linking is not directly possible from these files.

License

The Zenodo record's metadata states CC BY-NC 4.0, while the paper and the original README state CC BY-NC-SA (NonCommercial-ShareAlike). We adopt the more restrictive non-commercial + share-alike interpretation (cc-by-nc-sa-4.0). Either way the data is non-commercial use only.

Provenance

Official author release — Zenodo record 10.5281/zenodo.4520773 (the challenge organisers' own deposit, linked from https://valdo.grand-challenge.org/). Public training counts (Task1=40, Task2=72, Task3=40 → 152) match the paper's Table 2 exactly. Not a third-party re-host. Re-distributed here under the source non-commercial license.

Citation

Sudre, C.H., Van Wijnen, K., Dubost, F., et al. "Where is VALDO? VAscular Lesions Detection and segmentatiOn challenge at MICCAI 2021." Medical Image Analysis, 91:103029, 2024. https://doi.org/10.1016/j.media.2023.103029 (preprint arXiv:2208.07167). Challenge: https://valdo.grand-challenge.org/ · Data: https://zenodo.org/records/4520773