ISLES 2022
ISLES 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset
Acute to subacute ischemic stroke MRI (FLAIR, DWI and ADC) with expert infarct masks from Munich, Bern and Hamburg: 400 cases for the ISLES 2022 challenge, of which the 250 training cases are on Zenodo under CC BY 4.0.
Overview
ISLES 2022 is the dataset behind the 2022 edition of the Ischemic Stroke Lesion Segmentation challenge. It pairs routine clinical stroke MRI from three European stroke centers with expert infarct masks, and was built to benchmark automatic segmentation of acute and subacute infarcts, from large territorial strokes to scattered small embolic lesions.
Composition
The collection has 400 cases, one per patient. 250 form the training set, which is public on Zenodo; the other 150 are a hidden test set used only to score challenge entries. The training cases come from the Technical University of Munich (198) and the University Hospital of Bern (52), per the center list published with Zenodo version 2.3.1. The test set adds a third center, the University Medical Center Hamburg-Eppendorf, that does not appear in training. Each case has a FLAIR image, a DWI trace image at b=1000 and its ADC map, plus a lesion mask. Three training cases were scanned for suspected stroke but show no infarct. The authors deliberately included more posterior circulation and infratentorial infarcts than random sampling would give. Only adults (18 or older) were included; no age or sex figures are published.
Acquisition
Scans were taken during clinical care on 3T Philips (Achieva, Ingenia), 3T Siemens (Verio) and 1.5T Siemens (Avanto, Aera) systems. Munich and Hamburg cases were acquired after revascularization therapy, Bern cases before it. Images are released in native space as NIfTI in a BIDS layout, with DICOM header fields as JSON where available. For de-identification, all images were skull-stripped with HD-BET before release.
Annotations
A 3D U-Net produced first drafts, which trained medical students corrected or redrew. A neuroradiology resident then revised every mask and one of three senior neuroradiologists approved it, using DWI, ADC and FLAIR together. In a 10-case check by two further neuroradiologists, the released masks agreed better with each expert than the experts agreed with each other.
Known limitations
- Demographics, onset-to-scan times and clinical outcomes are not released.
- The test set and the entire Hamburg center are hidden, so held-out evaluation needs the challenge platform.
- Case selection over-represents posterior circulation strokes.
- Annotation started from algorithm drafts, which may carry a bias toward that model's output.
Cohort
Aggregate numbers from the sources below. Bars are relative to the 400 subjects.
Contrast / sequence
subjects, values can overlap
- FLAIR 400 100%
- Diffusion-weighted 400 100%
- ADC map 400 100%
Split
subjects
- Training 250 63%
- Test 150 38%
License and access
Our reading of the license, not legal advice. Before you use the data, read the original license and confirm that your use is allowed. We take no responsibility for how you use a dataset. Full disclaimer
Download without an account
Only the 250 training cases are on Zenodo. The 150 test cases are used for challenge evaluation and are not released.
Creative Commons Attribution 4.0 International
Use, share and adapt the data for any purpose, including commercial use, as long as you credit the creators.
What you can do
- Yes
- Yes
- Yes
- Yes
What you can share
- Yes
- Yes
- Yes
What you must do
- Yes
- Share alike No
- No
- No
- Manuscript review No
- Release code No
- Return results No
- Delete after use No
Limits
- No
- Location limits No
Citation
Hernandez Petzsche MR, de la Rosa E, Hanning U, et al. ISLES 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. Scientific Data 9, 762 (2022). doi:10.1038/s41597-022-01875-5
All numbers
Every number on this page, as stored in stats.csv, with its source.
| Measure | Breakdown | Value | Source |
|---|---|---|---|
| Subjects | total one MRI case per subject | 400 | hernandezpetzsche2022 Abstract |
| Subjects | split=train released on Zenodo with lesion masks | 250 | hernandezpetzsche2022 Data Records, Data structure and file formats |
| Subjects | split=test not released; equal parts from the three centers | 150 | hernandezpetzsche2022 Data Records, Data structure and file formats |
| Subjects | contrast=FLAIR the protocol required at least FLAIR and DWI (with its ADC map) | 400 | hernandezpetzsche2022 Methods, Subject selection |
| Subjects | contrast=dwi | 400 | hernandezpetzsche2022 Methods, Subject selection |
| Subjects | contrast=ADC | 400 | hernandezpetzsche2022 Methods, Subject selection |
| Subjects | split=train;contrast_set=ADC+dwi+FLAIR each released training case has FLAIR, DWI (b=1000) and the ADC map | 250 | zenodo-7960856 Description |
| Subjects | split=train;condition=ischemic_stroke 250 training cases minus the 3 in which no infarct was found | 247 | hernandezpetzsche2022 Methods, Subject selection |
| Subjects | split=train;country=DE | 198 | zenodo-7960856-center-ids columns case, center (center 1 = TUM Munich) |
| Subjects | split=train;country=CH | 52 | zenodo-7960856-center-ids columns case, center (center 2 = University Hospital of Bern) |
Sources
The keys used in the table above.
- hernandezpetzsche2022 Hernandez Petzsche et al. 2022, ISLES 2022 data descriptor (Scientific Data) paper
- zenodo-7960856 Zenodo record 7960856, ISLES 2022 training dataset version 2.3.1 (description and file list) data
- zenodo-7960856-center-ids center_ids.xlsx from Zenodo record 7960856 (CC BY 4.0), rows counted per value of the center column computed