---
license: cc-by-4.0
task_categories:
- image-segmentation
language:
- en
tags:
- medical-imaging
- mri
- brain
- stroke
- ischemic-stroke
- lesion-segmentation
- perfusion
- isles
pretty_name: ISLES 2017 (Ischemic Stroke Lesion Segmentation / Outcome Prediction)
size_categories:
- n<1K
---

# ISLES 2017 — Ischemic Stroke Lesion Segmentation Challenge (Outcome Prediction)

Multispectral brain **MRI** from the **MICCAI 2017** edition of the Ischemic
Stroke Lesion Segmentation (ISLES) challenge. The 2017 task is **lesion outcome
prediction**: estimate the *final* infarct (segmented on a ~90-day follow-up
scan) from *acute* multispectral MRI.

This repository contains the **public training split only (43 cases)**, every
case carrying its expert ground-truth lesion mask. The 32-case test split is
**not** included because its ground-truth masks were withheld by the challenge
organizers for leaderboard scoring (image-only, not useful for segmentation
evaluation).

## Contents

- **43 training cases** (`training_1` … `training_48`, **non-contiguous** — IDs
  3, 17, 25, 29, 34 are absent).
- **8 NIfTI (`.nii`) volumes per case**, each in its own sub-folder:

  | Token | Description | Dim |
  |---|---|---|
  | `MR_4DPWI` | Raw 4-D perfusion-weighted source series (40–80 timepoints) | 4-D / 5-D |
  | `MR_ADC`   | Apparent diffusion coefficient map | 3-D |
  | `MR_MTT`   | Mean transit time | 3-D |
  | `MR_rCBF`  | Relative cerebral blood flow | 3-D |
  | `MR_rCBV`  | Relative cerebral blood volume | 3-D |
  | `MR_Tmax`  | Time-to-maximum of the residue function | 3-D |
  | `MR_TTP`   | Time to peak | 3-D |
  | `OT`       | **Ground-truth lesion mask** (binary 0/1) | 3-D |

- Perfusion maps were computed by the organizers with Olea Sphere (bSVD
  deconvolution). All volumes within a case share the same spatial grid and are
  co-registered to the mask. Spatial grids vary across cases (e.g. 192×192×19,
  256×256×24, 128×128×25, 192×192×30).
- Note: the raw `MR_4DPWI` series is stored **5-D** `(X, Y, Z, 1, T)` for cases
  using the `VSD.*` naming and **4-D** `(X, Y, Z, T)` for cases using the
  `SMIR.*` naming (squeeze the singleton when present).

## Folder / naming convention

```
training_<N>/<PREFIX>.Brain.XX.O.<TOKEN>.<objectID>/<PREFIX>.Brain.XX.O.<TOKEN>.<objectID>.nii
```

`<PREFIX>` is **`VSD`** for 30 cases and **`SMIR`** for 13 cases
(`training_36`–`training_48`). A dataloader must match on `<TOKEN>` (e.g.
`*.MR_ADC.*`, `*.OT.*`) **regardless of prefix**, and must enumerate the actual
directories rather than assuming a contiguous `1..43` range. The `<objectID>`
in each folder name is the original **SMIR/VSD object ID** and is preserved
here — it is the only cross-year linkage handle (see Overlap below).

## Ground truth

The `OT` volume is the final infarct delineated on the follow-up T2 image by a
board-certified neuroradiologist in 3D Slicer. **Binary** (0 = background,
1 = lesion). All 43 training masks are non-empty (lesion sizes range from a few
dozen to ~24 k voxels). This is the single, gold-standard annotation tier; there
is no alternative rater set in the public release.

## Provenance & license

- **Source:** organizers' Zenodo deposit
  [DOI 10.5281/zenodo.17736412](https://doi.org/10.5281/zenodo.17736412)
  (Reyes, de la Rosa, Menze; Nov 2025), released **CC BY 4.0**. The original
  official host (SMIR / `smir.ch`, `isles-challenge.org/ISLES2017`) is offline
  as of 2026, so this organizer-authored re-host is the working source.
- The per-folder `License_ODC_ODBL.txt` files shipped inside the original
  archive carry an Open Data Commons ODbL notice; the umbrella Zenodo deposit
  re-licenses the posted files under CC BY 4.0. Attribute the challenge when
  reusing.
- Case counts match the benchmark paper exactly (43 train / 32 test = 75).

## Cross-dataset overlap (evaluation-leakage note)

- **ISLES 2017 fully contains ISLES 2016** (it is the 2016 cohort extended by
  8 train + 13 test cases). Never benchmark 2016 and 2017 as independent sets.
- **Independent of `Angelou0516/isles2022` and `Angelou0516/isles2024`** — those
  are newer, separate multi-center cohorts with no documented patient overlap.
- There is no official cross-year patient-ID crosswalk; the SMIR/VSD object ID
  embedded in folder names is the only handle and is shared only between the
  2016 and 2017 editions.

## Citation

> Winzeck S, Hakim A, McKinley R, et al. *ISLES 2016 and 2017 — Benchmarking
> Ischemic Stroke Lesion Outcome Prediction Based on Multispectral MRI.*
> Frontiers in Neurology 2018;9:679. doi:10.3389/fneur.2018.00679

Please also cite the data via its Zenodo DOI `10.5281/zenodo.17736412`.