[!IMPORTANT] All data-related questions must be posted in the Forum. We do not provide technical support or answer data questions via email.

📂 Dataset Release: Whole Training Data (Batches 1 and 2) has been released (N = 1453)

The full ISLES'26 training dataset is now available for model training and local development. This release consists of the 1) Training and (previously labelled as) Validation sets from the ATLAS v2.0 archive (N=955), 2) N=169 cases sourced from the Stroke Outcome Optimization Project (SOOP) [1], and 3) N=329 newly added cases.

[!NOTE] For the purposes of ISLES'26, ALL scans in this training set are intended for training.

🔗 Data Download

Participants can access the files through the following link:

NITRC: ATLAS v3.0 Download Portal

Further information

⚠️ Critical Requirement: Use the RAW Data

The ISLES'26 challenge focuses on algorithms operating in native space. * Mandatory: You must download and use the "Raw Data" folder. * Note: Do not use standardized or normalized versions of the dataset, as the official challenge evaluation will occur exclusively on raw, native-space MRI.


📝 Key Information

This release contains significant updates from the original ATLAS v2.0 publication [2]:

  • Native Format: All images are provided in their raw native coordinate system.
  • Release of Validation Masks: Unlike the original ATLAS v2.0 publication, this release includes lesion masks for the old validation set, making them available for your training pipelines.
  • Skull-Stripped: All images have already been skull-stripped; no additional brain extraction is required.


📁 Training set Summary

Feature Details
Total Scans 1453
Challenge Phase Training Data (Includes ATLAS v2.0 Train + Validation, SOOP scans (N=169) and 329 new cases)
Image Space Native / RAW
Preprocessing Skull-stripped
Annotations Lesion masks included for all subjects

References:

[1] Absher, John, et al. "The stroke outcome optimization project: Acute ischemic strokes from a comprehensive stroke center." Scientific Data 11.1 (2024): 839.

[2] Liew, Sook-Lei, et al. "A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms." Scientific data 9.1 (2022): 320.

[!TIP] Ensure your processing pipeline is designed to handle varying orientations and resolutions inherent to native-space (RAW) clinical data.