AIMS-TBI
Automated Identification of Moderate-Severe Traumatic Brain Injury Lesions (AIMS-TBI) challenge dataset
MICCAI challenge data from the ENIGMA paediatric and adult moderate-severe TBI working groups: 892 defaced T1-weighted brain MRI scans from 13 sites (2025 release) with manual binary lesion masks, shared after signing a University of Utah data use agreement.
Overview
AIMS-TBI is a lesion segmentation challenge on brain MRI after moderate to severe traumatic brain injury, run at MICCAI in 2024, 2025 and 2026 by researchers at the University of Utah. The images come from the paediatric and adult moderate-severe TBI working groups of the ENIGMA consortium. The task is to detect and segment lesions on a single T1-weighted scan, the sequence most sites have.
Composition
The 2025 dataset holds 892 T1-weighted images: 553 for training, 100 for validation and 239 for a hidden test set, drawn from 13 sites with a similar mix of sites in each split. A case is one scan timepoint, and some patients have longitudinal scans, so the number of people is lower than the number of images and is not reported. The 2024 edition used 764 images (388 training, 101 validation, 275 test). For 2026 the same data were reused, and some training cases also come with diffusion MRI, T2-weighted, FLAIR or SWI scans, without separate lesion masks. Age, sex and time since injury come with the images. The source does not say whether the working groups contributed control subjects.
Acquisition
Scans were made on GE, Siemens and Philips scanners at 1.5 T and 3 T with differing protocols. Most T1-weighted images have 1 mm isotropic voxels. Contributing sites include Baylor College of Medicine, UCLA, Penn State, Kennedy Krieger Institute, Loma Linda, Nationwide Children's Hospital, Murdoch Children's Research Institute, Deakin University, UT Houston, Kessler Foundation, VA Palo Alto and the University of Oslo. Images are defaced with pydeface and carry no other preprocessing.
Annotations
Lesions are binary masks that merge all visible TBI damage, such as contusions, haemorrhage, haematoma, encephalomalacia, gliosis, white matter lesions and drainage tracts. Each mask went through three steps: a trained rater edited an automated segmentation, a second rater reviewed it, and one of five expert raters approved it. Raters had to reach a Dice of 0.6 on training cases before working on real data. Masks are released for training cases only.
Known limitations
- Access requires a data use agreement with the University of Utah; registration for the 2026 edition is closed.
- One lesion class: haemorrhage is not separated from other lesion types.
- Subject counts, age and sex distributions per split are not published.
- Two terms apply: the signed agreement and a CC BY-NC-ND license named in the challenge design document.
Cohort
Aggregate numbers from the sources below. Bars are relative to the largest value.
Contrast / sequence
scans, values can overlap
- T1-weighted 892 100%
Split
scans
- Training 553 62%
- Test 239 27%
- Validation 100 11%
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
Sign a data use agreement, often reviewed by the provider
Different parts of the data carry different licenses. The summary on the right shows the most restrictive answer per rule.
AIMS-TBI Data Use Agreement (University of Utah)
Signed through the challenge data access form with the University of Utah. Use is limited to research under the University of Utah IRB 138479 or other purposes agreed in writing; the data may not be passed on, publications must credit the University of Utah, and the data must be returned or destroyed when the permitted use ends.
What you can do
- Conditional
- Not stated
- Create derived data Not stated
- Conditional
What you can share
- No
- Not stated
- Share trained models Not stated
What you must do
- Yes
- Share alike No
- Yes
- No
- Manuscript review No
- Release code No
- Return results No
- Yes
Limits
- No re-identification No
- No
Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International
Share the unchanged data with credit for non-commercial purposes. You may not share anything derived from it, and Creative Commons advises against training models on it.
What you can do
- No
- No
- Conditional
- Yes
What you can share
- Conditional
- No
- No
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
Deutscher E, Tustison N, Onicas A, Wilde EA, Pease MW, Bakas S, Dennis EL. Automated Identification of Moderate-to-Severe Traumatic Brain Injury Lesions (AIMS-TBI) 2025 MICCAI Challenge. Lecture Notes in Computer Science, Springer (2026). doi:10.1007/978-3-032-16370-7_24
All numbers
Every number on this page, as stored in stats.csv, with its source.
| Measure | Breakdown | Value | Source |
|---|---|---|---|
| Scans | total final AIMS-TBI 2025 dataset; one case is one MRI timepoint and some patients have several | 892 | aims-tbi-design-2026 Training and test case characteristics (b) |
| Scans | split=train | 553 | aims-tbi-design-2026 Training and test case characteristics (b) |
| Scans | split=validation | 100 | aims-tbi-design-2026 Training and test case characteristics (b) |
| Scans | split=test hidden test set | 239 | aims-tbi-design-2026 Training and test case characteristics (b) |
| Scans | contrast=T1w all T1w data are annotated | 892 | aims-tbi-design-2026 Training and test case characteristics (c) |
Sources
The keys used in the table above.
- aims-tbi-design-2026 Dennis et al. 2026, AIMS-TBI structured challenge design document (Zenodo, CC BY 4.0) paper
- aims-tbi25-site AIMS-TBI 2025 challenge website (grand-challenge.org) website