RSNA Knee Abnormality Detection (2026)

Knee MRI with 12 study-level binary abnormality labels, and the original free-text radiology report on each training study

Non-commercialVerifiedAccount required
AREAMSKMODMRFORMATDICOMNORMALSPartialLICKaggle competition rules (non-commercial)CHECKED2026-11-11

Open the source page ↗ See it in the directory

At a glance

What it is
Knee MRI with 12 study-level binary abnormality labels, and the original free-text radiology report on each training study
Area
MSK
Modality
MR
Format
DICOM
Size
4,407 training studies and 24,371 series (about 570 GB of DICOM). The RSNA challenge page describes more than 5,000 knee MRI exams from 19 sites
Normals included
Partial (mixed or unlabeled)
License
Kaggle competition rules (non-commercial)
Commercial use
Not allowed under the free terms
Access
Free account and data-use terms required
Status
Verified: source page reviewed by Hounsfield Lab
Added
2026-11-04
Last checked
2026-11-11
Source
www.kaggle.com/competitions/rsna-knee-abnormality-detection

How to access

Create a free Kaggle account, open this competition, accept the rules, then download from the Data tab. Do not pass the files to anyone who has not accepted the rules.

Limitations

Competition Data may be used only for the competition, academic research, and education. Commercial use and redistribution are prohibited. These competitions use Kaggle's standard rules template; re-read this competition's rules page before you rely on it. RSNA's challenge page calls this the first RSNA challenge that trains on both the MRI and the text of the radiology reports. The reports are in a dozen languages. The 12 labels are ACL, MCL, medial meniscus, lateral meniscus, medial OA, lateral OA, patellofemoral OA, effusion, synovitis, Baker's cyst, contusion, and fracture. Only 58 training studies have a complete set of those expert labels; the other training studies have the report and not the structured labels. The RSNA pages do not say how many exams are negative for every finding, so this is not listed as a normal-control set. A study can still be negative for a given label. The challenge page says 19 sites; an August 2026 RSNA news post said 16. The challenge page does not print the training counts, the 58-study figure, or the byte size.

Citation required?

Yes. Cite the RSNA 2026 Knee Abnormality Detection AI Challenge. Check the rules page for the requested wording. Challenge page: https://www.rsna.org/artificial-intelligence/ai-image-challenge/knee-mri-ai-challenge

Related open-source tools

More msk datasets

Hounsfield Lab links to the source and does not host images. The license and access summary reflects the source page on the last-checked date; terms can change, so read the source before you download. Not legal advice.

← All DICOM datasets