MRNet
MRNet: A Dataset of Knee MRIs
Clinical knee MRI exams with labels for any abnormality, ACL tear and meniscal tear taken from reports, built to train and benchmark whole-exam knee classifiers. 1,370 exams from Stanford, 120 in a hidden test set. Non-commercial research use only.
Overview
MRNet is a collection of 1,370 clinical knee MRI exams from Stanford University Medical Center, acquired between 2001 and 2012. The Stanford Machine Learning Group assembled it to train and evaluate MRNet, a convolutional network that classifies whole knee exams, and released it together with a public benchmark. Each exam carries three binary labels: any abnormality, anterior cruciate ligament (ACL) tear and meniscal tear. The data are now distributed by the Stanford AIMI center through Redivis.
Composition
Of the 1,370 exams, 1,104 are abnormal, 319 show an ACL tear and 508 a meniscal tear; 194 have both tears. The exams are divided into a training set (1,130 exams, 1,088 patients with an identifier), a validation set (120 exams, 111 patients) and a test set (120 exams, 113 patients). The paper calls the validation set the tuning set and the test set the validation set. All exams of one patient fall into the same split, and the two smaller sets were sampled so that each holds at least 50 positive cases per label. The test set is held back for the benchmark and is not part of the download. The mean patient age is 38.0 years, and 569 exams belong to female patients.
Acquisition
All exams were done on GE scanners with a standard knee coil and a routine protocol without contrast: coronal T1-weighted, coronal T2-weighted with fat saturation, sagittal proton density, sagittal T2-weighted with fat saturation and axial proton density with fat saturation. 775 exams were acquired at 3 T and the rest at 1.5 T. The paper used three series per exam (sagittal T2-weighted, coronal T1-weighted and axial proton density), resampled to 256 × 256 pixels.
Annotations
Training and validation labels were extracted by hand from the clinical radiology reports. For the test set, the reference labels are the majority vote of three musculoskeletal radiologists who had the images, reports, clinical history and follow-up exams.
Known limitations
Released labels come from reports, not surgery, and are exam-level only, with no localization. The data come from one institution and one scanner vendor. The research use agreement forbids commercial use, redistribution and derivative works.
Cohort
Aggregate numbers from the sources below. Bars are relative to the largest value.
Sex
Studies- Female 569 100%
Covers 569 of 1,370 studies.
Scanner vendor
Studies- GE HealthCare 1,370 100%
Methods - Dataset in Bien et al. 2018, Deep-learning-assisted diagnosis for knee MRI - development and retrospective validation of MRNet (PLoS Med 15(11))
Field strength
Studies- 3 T 775 57%
- 1.5 T 595 43%
Methods - Dataset in Bien et al. 2018, Deep-learning-assisted diagnosis for knee MRI - development and retrospective validation of MRNet (PLoS Med 15(11))
Split by condition
StudiesReported cross table. A dot marks a cell the source does not give.
| Meniscal tear | Anterior cruciate ligament tear | |
|---|---|---|
| Training | 397 | 208 |
| Test | 59 | 57 |
| Validation | 52 | 54 |
Age
mean 38
No age bins reported.
Condition
StudiesGroups can overlap
- Meniscal tear 508 37%
- Anterior cruciate ligament tear 319 23%
Methods - Dataset in Bien et al. 2018, Deep-learning-assisted diagnosis for knee MRI - development and retrospective validation of MRNet (PLoS Med 15(11))
Split
- Training 1,088
- Test 113
- Validation 111
Table 1 in Bien et al. 2018, Deep-learning-assisted diagnosis for knee MRI - development and retrospective validation of MRNet (PLoS Med 15(11)); Splits in Stanford ML Group MRNet dataset and competition page
Split by sex
StudiesReported cross table. A dot marks a cell the source does not give.
| Female | |
|---|---|
| Training | 480 |
| Validation | 50 |
| Test | 39 |
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
Create an account and accept terms online
Register and accept the Stanford research use agreement. The data are downloaded through Redivis. The test set is not released.
Stanford University School of Medicine Dataset Research Use Agreement (AIMI)
Free for personal, non-commercial research only. You may not share the data or the download link, may not create derivative works, and must not try to re-identify patients. Stanford sells a separate commercial license.
The MRNet page publishes the Stanford University School of Medicine Research Use Agreement with "MRNet Dataset" as the dataset name. It also states that the dataset is for non-clinical research use only and must not be relied upon in patient care.
What you can do
- No
- Not stated
- Not stated
- No
- Not stated
What you can share
- No
- No
- Share trained models Not stated
What you must do
- Yes
- Share alike No
- Yes
- Ethics approval No
- Manuscript review No
- Release code No
- Return results No
- No
Limits
- Yes
- Location limits No
Commercial license: Yes Companies can apply for a one-year license per dataset, reviewed by a Stanford committee. The page lists an annual fee of USD 70,000 per dataset for agreements contracted in FY25. Details
Citation
Bien N, Rajpurkar P, Ball RL, et al. Deep-learning-assisted diagnosis for knee magnetic resonance imaging: Development and retrospective validation of MRNet. PLoS Med 15(11): e1002699 (2018). https://doi.org/10.1371/journal.pmed.1002699. Dataset DOI: https://doi.org/10.71718/rcbp-8c35
Sources
Every number on this page comes from one of these documents. Each chart names the table or page it is taken from. The raw numbers are in stats.csv.