MedOtter/isles2015

ISLES 2015 (SISS + SPES) — training splits

Raw mirror of the MICCAI ISLES 2015 (Ischemic Stroke Lesion Segmentation) challenge training data, from the challenge organizers' Zenodo re-archive (record 19135955, v3 of concept DOI 10.5281/zenodo.17736411; Reyes / de la Rosa / Menze).

Scope (read this): this mirror contains the 58 training cases with public ground truth. The 56 test cases (SISS 36, SPES 20) are not included: their ground truth was never released and the original SMIR online evaluation service is dead, so they cannot be scored.

Composition — two sub-challenges, not mere splits

Sub-challenge	Task	Cases here	Modalities	Voxels	Center
SISS	Sub-acute ischemic stroke lesion segmentation	28 train	T1, T2, DWI (b=1000), FLAIR	1×1×1 mm (~230×230×154)	UMC Schleswig-Holstein, Lübeck
SPES	Acute stroke penumbra estimation	30 train	T1c, T2, DWI, CBF, CBV, TTP, Tmax	2×2×2 mm (96×110×71-72)	University Hospital Bern (2005–2013)

Per case, all volumes are skull-stripped and rigidly co-registered to a common grid (SISS: to FLAIR, resampled 1 mm³; SPES: to T1c). Uncompressed NIfTI (.nii).

Ground truth

• SISS — OT folder per case: binary lesion mask by one experienced MD, delineated on FLAIR (lesion counted only if pathologic in both FLAIR and DWI). Single public tier (GT01). A second test-set rater existed (inter-observer Dice 0.70±0.20 — the human ceiling reported by the paper) but test GT is withheld.
• SPES — three label tiers per case:
  • corelabel — binary infarct core (ADC < 600×10⁻⁶ mm²/s threshold).
  • penumbralabel — the designated gold standard: complete hypoperfused lesion (semi-manual, Tmax > 6 s + manual correction). This is the target the 2015 challenge actually evaluated.
  • mergedlabels — disjoint decomposition: 1 = penumbra − core, 2 = core. Verified voxel-exact: merged==2 ≡ core and merged>0 ≡ core ∪ penumbra.
  • ⚠️ Core ⊄ penumbra exactly: a few core voxels (~14–205 per case) fall outside penumbralabel. Treat core and penumbra as two overlapping binary targets (fan-out), or use mergedlabels for a clean 3-class map — do not assume strict nesting.

No empty masks: every training case has non-empty GT (all tiers).

Structure

ISLES2015_SISS/training/<1..28>/VSD.Brain.XX.O.<MR_DWI|MR_Flair|MR_T1|MR_T2|OT>.<id>/<same>.nii
ISLES2015_SPES/training/<1..30>/VSD.Brain.XX.O.<MR_CBF|MR_CBV|MR_DWI|MR_T1c|MR_T2|MR_Tmax|MR_TTP>.<id>/<same>.nii
ISLES2015_SPES/training/<n>/{corelabel,penumbralabel,mergedlabels}/VSD.Brain*.XX.O.OT.<id>/<same>.nii
train.jsonl          # one row per case: relative paths per modality/label tier + shape/spacing

• The numeric <id> suffixes are SMIR-internal object IDs, unique per file and release-specific — match files by modality token (.MR_Flair., .OT.), never by ID.
• mergedlabels files use a VSD.Brain_1more.* prefix.
• Each file folder retains the original SMIR License_ODC_ODBL.txt.
• The parquet under data/ is a browsing preview only (one representative lesion slice rendered per case); train on the raw NIfTI volumes.

Benchmark / leakage warnings

• SPES ↔ ISLES 2016/2017 (MedOtter/isles2016, MedOtter/isles2017): plausible patient overlap — same institution (Bern) and overlapping era (2005–2013 vs 2005–2015), no public crosswalk exists. Do not treat SPES and ISLES 2016/17 as independent benchmarks and do not put one in train and the other in eval.
• SISS ↔ ISLES 2022: institutional lineage only; the 8 Munich SISS cases are all in the (unmirrored) test split, so the 28 mirrored SISS cases are all Lübeck.
• SISS test-set GT does not exist publicly → supervised evaluation uses the training split only.

License & provenance

Zenodo re-archive: CC BY 4.0 (organizer-deposited). The still-live official challenge page additionally states the post-workshop re-release under ODbL 1.0 (contents: DbCL 1.0) — the per-folder License_ODC_ODBL.txt files are that original grant. Both licenses permit redistribution with attribution. Original host SMIR / virtualskeleton.ch is defunct.

Citation

Maier O, Menze BH, von der Gablentz J, et al. ISLES 2015 — A public evaluation benchmark for ischemic stroke lesion segmentation from multispectral MRI. Medical Image Analysis 35 (2017): 250–269. doi:10.1016/j.media.2016.07.009

Plus the Zenodo re-archive: doi:10.5281/zenodo.17736411.