README.md

metadata

license: cc-by-nc-sa-4.0
pretty_name: ISLES'24 (Ischemic Stroke Lesion Segmentation Challenge 2024)
task_categories:
  - image-segmentation
tags:
  - medical
  - medical-imaging
  - stroke
  - brain
  - ct
  - cta
  - ct-perfusion
  - mri
  - dwi
  - segmentation
  - isles
size_categories:
  - n<1K

ISLES'24 - Ischemic Stroke Lesion Segmentation Challenge 2024

A real-world, longitudinal, multimodal acute-stroke dataset: acute admission CT (NCCT, CTA, 4D CT-perfusion + derived perfusion maps), follow-up MRI (DWI/ADC), and clinical tabular data, with the final infarct lesion as the segmentation target.

This repository mirrors the public TRAINING set only (149 cases). The 98-case ISLES'24 test set is withheld by the organizers for challenge scoring on Grand Challenge and is not part of any public release.

Provenance

• Official source: Zenodo record 16748089 - "ISLES'24 - A Real-World Longitudinal Multimodal Stroke Dataset", train.7z, CC BY-NC-SA 4.0 (author-provided; not a third-party re-host).
• Challenge: https://isles-24.grand-challenge.org/
• Code / data-loading reference: https://github.com/ezequieldlrosa/isles24
• This mirror is an unmodified raw copy of the released volumes (no resampling, no intensity changes), with two documented exceptions below.

Counts & faithfulness notes

• 149 training cases (sub-stroke0001 ... sub-stroke0189, non-contiguous): University Hospital Munich (TUM) + University Hospital Zurich.
• The companion paper reports N=150 for training; the public Zenodo release ships 149 (one case excluded upstream). This mirror = the 149 released.
• Dropped duplicate: the upstream release contains one stray uncompressed file sub-stroke0142_ses-02_lesion-msk.nii (a native-space leftover). Subject 0142 also has the standard co-registered *_space-ncct_lesion-msk.nii.gz like every other case, so this mirror keeps the standard GT and omits the stray.

Structure (BIDS)

<sub>/                      sub-stroke0001 ... sub-stroke0189  (149)
  ses-01  = acute admission CT
  ses-02  = follow-up MRI (2-9 days later)

raw_data/<sub>/ses-01/                     # native acquisition space
  *_ncct.nii.gz   *_cta.nii.gz   *_ctp.nii.gz (4D, 55 timepoints)
  perfusion-maps/ *_{cbf,cbv,mtt,tmax}.nii.gz

derivatives/<sub>/                         # ALL co-registered to NCCT (space-ncct)
  ses-01/ *_space-ncct_cta.nii.gz  *_space-ncct_ctp.nii.gz (4D)
          *_space-ncct_lvo-msk.nii.gz      # large-vessel occlusion (CTA), binary
          *_space-ncct_cow-msk.nii.gz      # Circle-of-Willis anatomy (CTA), multi-class
          perfusion-maps/ *_space-ncct_{cbf,cbv,mtt,tmax}.nii.gz
  ses-02/ *_space-ncct_dwi.nii.gz  *_space-ncct_adc.nii.gz
          *_space-ncct_lesion-msk.nii.gz   # <-- GROUND TRUTH (final infarct), binary

phenotype/<sub>/ses-*/ *_demographic_baseline.csv  *_outcome.csv

Co-registration: within each subject, the raw NCCT and every derivatives/*space-ncct* volume (CTA, perfusion maps, DWI/ADC, all masks) share an identical voxel grid (shape, spacing, affine). The raw NCCT itself is the reference space (hence there is no space-ncct_ncct). So a fully aligned input+label stack = raw ncct + the derivatives space-ncct modalities + space-ncct_lesion-msk.

Ground truth

Mask	Space	Type	Role
space-ncct_lesion-msk (ses-02)	NCCT	binary {0,1}	Gold standard - final infarct lesion, the challenge target
space-ncct_lvo-msk (ses-01)	NCCT	binary {0,1}	Auxiliary - large-vessel occlusion (on CTA)
space-ncct_cow-msk (ses-01)	NCCT	multi-class	Auxiliary - Circle-of-Willis anatomy (on CTA)

The gold-standard lesion mask is derived from the follow-up DWI: masks were auto-generated by DeepISLES (the productized ISLES'22 winning ensemble) and then quality-controlled / manually corrected and verified by two neuroradiologists (>10 years' experience). The neuroradiologist-verified DWI lesion is the reference standard; it is provided co-registered into NCCT space.

Cross-dataset overlap (leakage note)

• No patient reuse from ISLES'22 (a different, MRI-only DWI cohort). ISLES'24 is a new acute-stroke CT cohort. Modality and patients differ.
• There is center-level overlap (Munich contributes to both editions) and a methodological dependency (ISLES'24 GT masks are produced by DeepISLES, which was trained on ISLES'22) - but neither is shared imaging.
• No shared TCIA / BraTS / Medical Segmentation Decathlon lineage. Identifiers are challenge-internal BIDS sub-stroke#### IDs only.

License

CC BY-NC-SA 4.0 (per the Zenodo deposit). Non-commercial, share-alike, attribution required.

Citation

@article{delarosa2024isles24,
  title   = {ISLES'24: Final Infarct Prediction with Multimodal Imaging and Clinical Data. Where Do We Stand?},
  author  = {de la Rosa, Ezequiel and Su, Ruisheng and Reyes, Mauricio and Riedel, Eda Otman and Baazaoui, Hakim and Wiest, Roland and Kofler, Florian and Kirschke, Jan S. and Wiestler, Benedikt and Menze, Bjoern},
  journal = {arXiv preprint arXiv:2408.10966},
  year    = {2024}
}

Data: Zenodo record 16748089 (de la Rosa et al., 2025). Please also credit the ISLES'24 challenge organizers and the contributing hospitals (TUM Munich, USZ Zurich).