VinDr-Mammo
VinDr-Mammo: A large-scale benchmark dataset for computer-aided diagnosis in full-field digital mammography
5,000 four-view full-field digital mammography exams (20,000 DICOM images) from two Hanoi hospitals, double read by radiologists with breast-level BI-RADS and density, bounding boxes for findings and a fixed train/test split.
Overview
VinDr-Mammo is a Vietnamese collection of full-field digital mammograms released by VinBigData on PhysioNet in 2022. It was built as a benchmark for computer-aided detection and diagnosis of breast findings and for predicting BI-RADS assessment and breast density, and it is one of the larger public digital mammography sets with radiologist annotations.
Composition
The dataset contains 5,000 exams with four images each: craniocaudal and mediolateral oblique views of both breasts, 20,000 DICOM images in total. The creators split the exams into 4,000 for training and 1,000 for testing with iterative stratification, so that BI-RADS categories, density levels and finding types have similar frequencies in both parts. Two CSV files hold the breast-level labels and the finding boxes, and a third keeps age and scanner model from the DICOM headers. The number of distinct women is not published.
Acquisition
Exams were sampled at random from the PACS of Hanoi Medical University Hospital and Hospital 108, covering 2018 to 2020, so screening and diagnostic exams are mixed. Images are "for presentation" mammograms from Siemens, IMS and Planmed units. Identifying text burned into image corners was blacked out, and only age and device information were kept in the headers.
Annotations
Three radiologists with 14 to 22 years of experience took part. Each exam was read independently by two of them, and a third, more senior reader settled disagreements. Each breast received a BI-RADS category (1 to 5) and a density category (A to D). Findings that needed follow-up (BI-RADS 3 or higher) were boxed and typed: mass, suspicious calcification, asymmetry (global or focal), architectural distortion, skin thickening or retraction, nipple retraction and suspicious lymph node. Benign BI-RADS 2 findings were not boxed.
Known limitations
- No pathology confirmation; labels are radiologist consensus only.
- Some finding types have fewer than 40 examples.
- The authors note that the files are not fully DICOM-compliant.
- Single country and two hospitals.
Cohort
Aggregate numbers from the sources below. Bars are relative to the largest value.
Anatomy
studies
- Breast 5,000 100%
Split
images
- Training 16,000 80%
- Test 4,000 20%
View
images
- Craniocaudal 10,000 50%
- Mediolateral oblique 10,000 50%
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
PhysioNet Restricted Health Data License 1.5.0
Research-only license for registered PhysioNet users who sign the matching data use agreement online. Unlike the credentialed license it asks for no identity check or human-subjects training. You must not share the data or try to identify people, and you must release the code behind your publications.
PhysioNet lists the PhysioNet Restricted Health Data License 1.5.0 and the matching data use agreement. The Scientific Data paper instead names the PhysioNet Credentialed Health Data License 1.5.0; the PhysioNet project page, which grants access, is taken as authoritative.
What you can do
- Not stated
- Not stated
- Create derived data Not stated
- Conditional
What you can share
- No
- Not stated
- Share trained models Not stated
What you must do
- Not stated
- Share alike No
- Yes
- Ethics approval No
- Manuscript review No
- Yes
- No
- No
Limits
- Yes
- Location limits No
Citation
Nguyen HT, Nguyen HQ, Pham HH, et al. VinDr-Mammo: A large-scale benchmark dataset for computer-aided diagnosis in full-field digital mammography. Sci Data 10, 277 (2023). Pham HH, Nguyen Trung H, Nguyen HQ. VinDr-Mammo (version 1.0.0). PhysioNet (2022). https://doi.org/10.13026/br2v-7517
All numbers
Every number on this page, as stored in stats.csv, with its source.
| Measure | Breakdown | Value | Source |
|---|---|---|---|
| Studies | total Exams; the number of distinct women is not reported | 5,000 | nguyen2023 Methods: Data acquisition |
| Studies | split=train | 4,000 | nguyen2023 Methods: Data stratification |
| Studies | split=test | 1,000 | nguyen2023 Methods: Data stratification |
| Studies | anatomy=breast | 5,000 | nguyen2023 Methods: Data acquisition |
| Images | total Four images per exam | 20,000 | nguyen2023 Methods: Data acquisition |
| Images | split=train Four images per exam | 16,000 | nguyen2023 Table 3 (8000 breasts x 2 views) |
| Images | split=test Four images per exam | 4,000 | nguyen2023 Table 3 (2000 breasts x 2 views) |
| Images | view=CC One CC and one MLO image per breast | 10,000 | nguyen2023 Data Records |
| Images | view=MLO One CC and one MLO image per breast | 10,000 | nguyen2023 Data Records |
Sources
The keys used in the table above.
- physionet VinDr-Mammo 1.0.0 project page on PhysioNet website
- nguyen2023 Nguyen et al. 2023, Scientific Data 10, 277 (Tables 3 to 5) paper