Duke Liver Dataset (DLDS)
Duke Liver Dataset (MRI)
2146 abdominal MRI series of 105 patients scanned for suspected liver cirrhosis at Duke, with series-type labels in 17 classes and manual whole-liver masks for 310 series, for MRI series classification and liver segmentation.
Overview
The Duke Liver Dataset (DLDS) is a collection of routine clinical liver MRI examinations from Duke University, released on Zenodo and described in a 2023 data resource article in Radiology: Artificial Intelligence. It was built to train and test two kinds of models: classifiers that recognize the type of an abdominal MRI series from its pixels alone, and liver segmentation models.
Composition
The dataset holds 2146 de-identified image series (113 280 DICOM images) from 105 adult patients, 76 men and 29 women aged 30 to 80. Every series carries one of 17 series-type labels, coded as letters A to Q so that users can blind themselves; a key file maps the letters to the types. The types cover diffusion and ADC, axial and coronal T2w, MRCP, localizers, in-phase and opposed-phase imaging, precontrast fat-suppressed T1w, several arterial, transitional, late dynamic and portal venous phases, and a catch-all "anything else" class. A segmentation subset contains 310 series from 95 patients with liver masks. Version 2.0.0 restored labels and masks for 26 patients that version 1.0.0 had withheld; the record lists their ids for users who want them as a test set.
Acquisition
Patients had contrast-enhanced abdominal MRI for a clinical indication of cirrhosis at one of three centers, and 87 showed imaging signs of cirrhosis. Contrast agents were gadobenate dimeglumine or gadoxetate disodium. Scanners were Siemens (96) and GE (9), at 1.5 T (54) or 3 T (51). The article reports ranges of sequence parameters per series type.
Annotations
Series types were labelled by humans. Liver masks were drawn slice by slice in OsiriX on axial portal venous and precontrast fat-suppressed T1w, in-phase and opposed-phase series, by an abdominal radiologist or one of two research fellows whose work the radiologist checked and corrected. Major portal vein branches and the inferior vena cava are excluded. Masks are binary DICOM stacks matching the image series.
Known limitations
- Each liver was segmented by a single reader, so no inter-rater estimate exists.
- Clinical images include motion and susceptibility artifacts; cirrhotic liver shapes differ from healthy livers.
- No lesion labels, and MRI is the only modality.
- The CC BY-NC-ND 4.0 option rules out commercial use and sharing of derived data; a separate Duke license is needed for commercial use.
Cohort
Aggregate numbers from the sources below. Bars are relative to the 105 subjects.
Sex
- Male 76 72%
- Female 29 28%
Contrast / sequence
scans, values can overlap
- T1-weighted, contrast enhanced 668 31%
- T2-weighted 428 20%
- T1-weighted 220 10%
- Diffusion-weighted 113 5%
- ADC map 99 5%
Condition
subjects, values can overlap
- Liver cirrhosis 87 83%
Scanner vendor
subjects
- Siemens Healthineers 96 91%
- GE HealthCare 9 9%
Field strength
subjects
- 1.5 T 54 51%
- 3 T 51 49%
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
A research proposal is reviewed and approved
Files are restricted on Zenodo. Request access through the record's form, giving your affiliation and a short description of the intended use; Duke University reviews the request.
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.
Requesters choose between CC BY-NC-ND 4.0 and a custom license negotiated with the Duke Office for Translation & Commercialization, which can allow commercial use. Modifications contributed back must have their copyright assigned to Duke University.
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
Macdonald JA, Zhu Z, Konkel B, Mazurowski MA, Wiggins WF, Bashir MR. Duke Liver Dataset: A Publicly Available Liver MRI Dataset with Liver Segmentation Masks and Series Labels. Radiology: Artificial Intelligence 5(5):e220275 (2023). doi:10.1148/ryai.220275
All numbers
Every number on this page, as stored in stats.csv, with its source.
| Measure | Breakdown | Value | Source |
|---|---|---|---|
| Subjects | total | 105 | macdonald2023 Patients and Image Acquisition |
| Subjects | sex=male | 76 | macdonald2023 Patients and Image Acquisition |
| Subjects | sex=female | 29 | macdonald2023 Patients and Image Acquisition |
| Minimum age | total age range 30-80 years | 30 | macdonald2023 Patients and Image Acquisition |
| Maximum age | total age range 30-80 years | 80 | macdonald2023 Patients and Image Acquisition |
| Subjects | condition=liver_cirrhosis imaging findings suggestive of cirrhosis; all 105 had a clinical indication of cirrhosis | 87 | macdonald2023 Patients and Image Acquisition |
| Subjects | vendor=siemens counted as machines in the paper | 96 | macdonald2023 Patients and Image Acquisition |
| Subjects | vendor=ge counted as machines in the paper | 9 | macdonald2023 Patients and Image Acquisition |
| Subjects | field_strength=1.5 | 54 | macdonald2023 Patients and Image Acquisition |
| Subjects | field_strength=3 | 51 | macdonald2023 Patients and Image Acquisition |
| Scans | total image series in 17 series types | 2,146 | macdonald2023 Table |
| Images | total unique DICOM images across all series | 113,280 | macdonald2023 Key Points |
| Scans | contrast=ADC axial ADC | 99 | macdonald2023 Table |
| Scans | contrast=dwi axial DWI | 113 | macdonald2023 Table |
| Scans | contrast=T2w sum of axial T2w (318) and coronal T2w (110); MRCP (159) not counted | 428 | macdonald2023 Table |
| Scans | contrast=T1w axial precontrast fat-suppressed T1w; in-phase (115) and opposed-phase (113) series not counted | 220 | macdonald2023 Table |
| Scans | contrast=T1w_ce sum of the seven contrast-enhanced T1w series types (early, mid and late arterial, transitional, axial and coronal late dynamic, portal venous) | 668 | macdonald2023 Table |
Sources
The keys used in the table above.
- macdonald2023 Macdonald et al. 2023, Radiology Artificial Intelligence (data resource article) paper
- zenodo-dlds-v2 Zenodo record Duke Liver Dataset (MRI) v2 (description and access conditions) website