Angelou0516/isles2016

ISLES 2016 - Ischemic Stroke Lesion Segmentation Challenge 2016

Brain-MRI ischemic-stroke dataset from the MICCAI 2016 ISLES challenge. This is a lesion-OUTCOME-PREDICTION task, not conventional visible-lesion segmentation: the inputs are acute perfusion-derived MR maps, while the ground-truth mask is the final infarct traced on a ~90-day follow-up scan and mapped back into the acute imaging space.

Read before benchmarking. A model is asked to predict future tissue fate from acute perfusion imaging - the target lesion is generally not directly visible in any single input map. Scores are not comparable to visible-pathology segmentation sets (e.g. ISLES'22 DWI infarct segmentation).

This repository mirrors the public TRAINING set only (30 cases). The 19 ISLES 2016 test cases are withheld by the organizers (no public GT) and are not part of this release.

Provenance

• Official source: Zenodo record 17736412 - "ISLES (Ischemic Stroke Lesion Segmentation/Prediction) Challenge Datasets (2015, 2016, 2017, 2018)", files ISLES2016_Training_{Native,CoRegistered}.zip. Deposited by the original ISLES organizers (Reyes, de la Rosa, Menze), 2025 - author-provided, not a third-party re-host.
• The original host SMIR / virtualskeleton.ch is decommissioned; the Zenodo archive is the only live source.
• This mirror is an unmodified raw copy of the released volumes (no resampling, no intensity changes, original folder layout and case IDs).

Counts & faithfulness notes

• 30 training cases, in both spaces. Folder IDs are the original training_1 ... training_35 numbering with 5 cases withdrawn in the public V3 release - missing IDs: 3, 17, 25, 29, 34. The original IDs are preserved (they cross-reference the challenge paper and the ISLES 2017 superset).
• The peer-reviewed challenge paper (Winzeck et al. 2018) reports 35 training / 19 test; the public TrainingV3 release ships 30 training. This mirror = the 30 publicly released training cases.
• Each case keeps the per-folder original license notice (License_ODC_ODBL.txt) and, for 4D PWI, a timing .csv - both retained as-is.

Structure

TrainingV3_Native/        training_<n>/   # each modality in its native acquisition grid
TrainingV3_CoRegistered/  training_<n>/   # all modalities resampled to a common grid

training_<n>/VSD.Brain.XX.O.MR_<MOD>.<id>/VSD.Brain.XX.O.MR_<MOD>.<id>.nii
training_<n>/VSD.Brain.XX.O.OT.<id>/VSD.Brain.XX.O.OT.<id>.nii   # <-- GROUND TRUTH

The numeric <id> in each VSD folder name differs between the two spaces and between modalities, so load by the modality token (MR_ADC, OT, ...), not by ID.

Folder token	Modality	Dim	Notes
MR_4DPWI	Raw perfusion-weighted imaging	4D	Source time-series; large (~112-199 MB/case)
MR_ADC	Apparent diffusion coefficient	3D	Diffusion (lesion-core proxy)
MR_MTT	Mean transit time	3D	Perfusion map
MR_rCBF	Relative cerebral blood flow	3D	Perfusion map
MR_rCBV	Relative cerebral blood volume	3D	Perfusion map
MR_Tmax	Time-to-maximum	3D	Perfusion map (>6 s ~ hypoperfusion)
MR_TTP	Time-to-peak	3D	Perfusion map
OT	Ground-truth lesion	3D	Binary final-infarct mask

Native vs CoRegistered. For segmentation use the CoRegistered space: within each subject every map and the OT mask share one voxel grid, so inputs and label align directly. The Native space preserves each modality's own acquisition geometry.

Ground truth

Single tier: the OT mask - the final infarct lesion, manually delineated by neuroradiologists on the ~90-day follow-up imaging and provided in the acute imaging space. Binary {0, 1}. There are no multi-rater or partial-annotation tiers to choose between.

Cross-dataset overlap (leakage note)

• ISLES 2017 is a SUPERSET. Per Winzeck et al. 2018 the 2017 training cohort (43 cases) extends the 2016 cohort (35) with 8 new cases; these 30 publicly released 2016 cases are therefore a subset of ISLES 2017's training data. Concretely, in the sibling mirror Angelou0516/isles2017 the 30 VSD-prefixed cases carry the identical training_<n> numbering and the same withdrawn IDs {3, 17, 25, 29, 34} (and the same 5D 4D-PWI layout) as this set - i.e. they are these cases - while the 13 SMIR-prefixed cases are the 2017-only additions. Never benchmark ISLES 2016 and ISLES 2017 as independent sets; de-duplicate by matching training_<n>.
• vs ISLES'22 (MRI DWI infarct segmentation): different task, different cohort - no known patient reuse.
• vs ISLES'24 (CT-centric multimodal stroke): different modality and patients
  • no overlap.
• No BraTS / Medical Segmentation Decathlon / TCIA lineage.

License

Re-released on Zenodo under CC BY 4.0 (record 17736412). The original SMIR-era per-file notices embedded in each folder are ODC-ODbL (License_ODC_ODBL.txt), retained unmodified. Both permit redistribution with attribution.

Citation

@article{winzeck2018isles,
  title   = {ISLES 2016 and 2017-Benchmarking Ischemic Stroke Lesion Outcome Prediction Based on Multispectral MRI},
  author  = {Winzeck, Stefan and Hakim, Arsany and McKinley, Richard and Reyes, Mauricio and others},
  journal = {Frontiers in Neurology},
  volume  = {9},
  pages   = {679},
  year    = {2018},
  doi     = {10.3389/fneur.2018.00679}
}

Data: Zenodo record 17736412 (Reyes, de la Rosa, Menze, 2025). Please credit the ISLES 2016 challenge organizers.