MedOtter/isles2018

ISLES 2018 — Ischemic Stroke Lesion Segmentation Challenge 2018

CT-perfusion (CTP) infarct-core segmentation in acute anterior-circulation ischemic stroke. 103 patients / 156 CTP acquisitions ("cases") pooled from the DEFUSE 2 study and a companion prospective cohort (4 centers: 3 US, 1 Australian; Cereda et al. 2016). Because CTP z-coverage was limited (4.4-16 cm), some patients were scanned as two separate slabs (A/B) — each slab is its own case_N folder.

Split	Cases	Patients	Two-slab patients	Ground truth
TRAINING	94	63	31	yes (OT)
TESTING	62	40	22	withheld (never released; SMIR leaderboard is dead)

Grouping warning: case_N is NOT a patient. Group on the patient_id column of cases.jsonl (derived from the SMIR sidecar descriptions, e.g. Train_40_A = case_64) to avoid slab-pair leakage across evaluation splits.

Per-case channels

Folder	Content
CT	baseline CT of the CTP study (~2 MB)
CT_4DPWI	raw 4D CTP source series, motion-corrected, ~1 s temporal resolution (22-336 MB)
CT_CBF, CT_CBV, CT_MTT, CT_Tmax	derived perfusion parameter maps
OT	binary infarct-core mask (TRAINING only)

The OT masks were manually drawn on MR DWI trace images acquired within 3 h after CTP (median CT-to-MRI 36 min) by a single stroke neurologist with group review, blinded to CTP, then registered to CTP space. DWI itself is NOT distributed — the task is cross-modality: CTP inputs, DWI-defined target (challenge-winner Dice ~0.51).

Layout

TRAINING/case_<1..94>/SMIR.Brain.XX.O.<MOD>.<smir_id>/SMIR.Brain.XX.O.<MOD>.<smir_id>.nii  (+ .json sidecar)
TESTING/case_<1..62>/...   (same, no OT)
cases.jsonl                 one row per case: case_id, split, smir_case_label, patient_id,
                            slab, images{mod:path}, mask, has_mask, smir_ids

Provenance & integrity

• Mirrored from the organizers' own re-release: Zenodo DOI 10.5281/zenodo.17736412 (published 2025-11-27 by M. Reyes, E. de la Rosa, B. Menze after smir.ch went offline). Zip md5: Training b6d16c456e173743e30a3fbdb3bd29a5, Testing 932a493c105024942700fe589cd441a9.
• Case counts verified against Hakim et al. 2021 (94/62 cases, 63/40 patients).
• Note: the Zenodo record's prose lumps 2018 under "multispectral MRI" — that is a boilerplate error; ISLES 2018 is CT perfusion.
• Modifications vs the archives: the 1,030 identical 285-byte License_ODC_ODBL.txt copies are deduplicated to the root LICENSE_ODC_ODBL.txt; .DS_Store dropped. All NIfTI volumes and JSON sidecars are byte-identical to the Zenodo zips.
• Cohort is disjoint from ISLES 2015/2016/2017 (Bern/Freiburg MRI), ISLES 2022 and ISLES 2024 — no cross-edition patient overlap is documented.

License

Every SMIR data folder embeds the Open Database License (ODbL) 1.0 (contents: Database Contents License 1.0) — see LICENSE_ODC_ODBL.txt. The organizers' Zenodo re-release is additionally tagged CC BY 4.0. Please attribute the ISLES 2018 organizers and cite the papers below.

Citations

@article{cereda2016benchmarking,
  title={A benchmarking tool to evaluate computer tomography perfusion infarct core predictions against a DWI standard},
  author={Cereda, Carlo W and Christensen, S{\o}ren and Campbell, Bruce CV and others},
  journal={Journal of Cerebral Blood Flow \& Metabolism},
  volume={36}, number={10}, pages={1780--1789}, year={2016},
  doi={10.1177/0271678X15610586}
}
@article{hakim2021predicting,
  title={Predicting infarct core from computed tomography perfusion in acute ischemia with machine learning: Lessons from the ISLES challenge},
  author={Hakim, Arsany and Christensen, S{\o}ren and Winzeck, Stefan and others},
  journal={Stroke}, volume={52}, number={7}, pages={2328--2337}, year={2021},
  doi={10.1161/STROKEAHA.120.030696}
}

Dataset re-release: Reyes M., de la Rosa E., Menze B. ISLES Challenge Datasets (2015, 2016, 2017, 2018). Zenodo, 2025. doi:10.5281/zenodo.17736412