Source-linked AI summary
VinDr-Mammo: A large-scale benchmark dataset for computer-aided diagnosis in full-field digital mammography
Hieu T. Nguyen, Ha Q. Nguyen, Hieu H. Pham, Khanh Lam, Linh T. Le, Minh Dao, Van Vu
TL;DR
Existing mammography datasets are often limited in scale, access, or use of screen-film images, despite the need for FFDM resources for CADe and CADx research. VinDr-Mammo introduces a public FFDM benchmark with structured breast- and lesion-level annotations, created through radiologist rereading and arbitration. It provides 5,000 four-view examinations for screening-oriented research, while lacking pathology-confirmed ground truth and some clinical information.
Problem
Many mammography datasets are limited in size or based on screen-film mammography, restricting available FFDM resources for CADe and CADx development.
Method
The authors construct VinDr-Mammo from FFDM examinations with BI-RADS breast-level labels, density assessments, lesion categories and locations, independent double reading, and arbitration of disagreements.
Results
VinDr-Mammo contains 5,000 four-view FFDM examinations and is publicly released on PhysioNet with breast-level and lesion-level annotations.
Takeaways & Limitations
The dataset provides an open imaging resource intended to support development of CADe and CADx tools for breast cancer screening.
Takeaways & Limitations
The dataset lacks pathology-confirmed ground truth and essential clinical information such as molecular and histology data.
Abstract
from arXiv · showhide
Mammography, or breast X-ray, is the most widely used imaging modality to detect cancer and other breast diseases. Recent studies have shown that deep learning-based computer-assisted detection and diagnosis (CADe or CADx) tools have been developed to support physicians and improve the accuracy of interpreting mammography. However, most published datasets of mammography are either limited on sample size or digitalized from screen-film mammography (SFM), hindering the development of CADe and CADx tools which are developed based on full-field digital mammography (FFDM). To overcome this challenge, we introduce VinDr-Mammo - a new benchmark dataset of FFDM for detecting and diagnosing breast cancer and other diseases in mammography. The dataset consists of 5,000 mammography exams, each of which has four standard views and is double read with disagreement (if any) being resolved by arbitration. It is created for the assessment of Breast Imaging Reporting and Data System (BI-RADS) and density at the breast level. In addition, the dataset also provides the category, location, and BI-RADS assessment of non-benign findings. We make VinDr-Mammo publicly available on PhysioNet as a new imaging resource to promote advances in developing CADe and CADx tools for breast cancer screening.
Background & Summary
VinDr-Mammo addresses limitations in existing mammography datasets by providing a large, open FFDM resource with breast-level assessments and extensive lesion-level annotations. The dataset is intended to support more diverse and interpretable CADe and CADx research.
- Motivation: Breast cancer screening remains challenging: recall is around 11%, sensitivity is 86.9%, and cancer detection is 5.1 per 1,000 screens.Many recalled cases ultimately are non-cancer, motivating improved screening support.
- Related datasets: Existing public mammography datasets vary in access, scale, and imaging type, with early resources largely based on digitalized screen-film mammograms.Examples include MIAS with 161 studies, DDSM with 2,620 exams, and INbreast with 115 exams, while the much larger NYU dataset has restricted access.
- Contribution: VinDr-Mammo provides 5,000 four-view FFDM exams with breast-level assessment and extensive lesion-level annotations.The dataset includes breast-level BI-RADS assessment and density, plus lesion category and localization information.
- Dataset design: Mammograms were acquired from two Hanoi hospitals, read independently by two radiologists, and sent to arbitration when assessments disagreed.The dataset follows BI-RADS reporting and marks abnormalities requiring short-term follow-up or suspected malignancy with bounding rectangles.
Methods
The dataset was constructed from retrospectively sampled hospital mammograms, re-read using BI-RADS-based breast and finding annotations, and quality-controlled through independent double reading with arbitration. A predefined stratification process supports consistent training and evaluation.
- Data stratification: The data creation workflow ends with separate training and test sets containing 4,000 and 1,000 examinations, respectively.A predefined split is intended to improve reproducibility and comparison across studies.
- Data acquisition: The dataset comprises 20,000 DICOM images from 5,000 examinations sampled through hospital PACS between 2018 and 2020.Random sampling included both screening and diagnostic examinations and retained the hospitals’ real patient-cohort distribution.
- Mammography reading: Radiologists re-read all 5,000 examinations because existing reports did not specify exact finding locations.This enabled both breast-level assessments and local-level annotations for CADe and CADx development.
- Mammography reading: Annotations follow BI-RADS and include breast-level assessment and density, while findings such as masses, calcifications, asymmetries, and architectural distortion receive bounding boxes and finding assessments.BI-RADS 6 is absent because biopsy results were unavailable.
- Mammography reading: Each examination was independently read by two radiologists, with discordant cases resolved by a third, more senior radiologist.The web-based VinDr Lab tool supported remote viewing and annotation by three experienced radiologists.
Data Records
VinDr-Mammo distributes images and annotations in a study-organized structure with CSV records for breast-level and finding-level information. Records encode image identity, breast and view metadata, labels, coordinates, and train/test membership.
- File organization: Images are organized into study-identifier subfolders, each containing four exam views, while annotations are stored in breast-level_annotations.csv and finding_annotations.csv.Both DICOM images and radiologists’ annotations are publicly available on PhysioNet.
- Finding records: Finding annotations represent abnormalities in specific images using bounding boxes, although cross-view linking for the same finding was not acquired.Each record includes identifiers for the image, study, and series, plus laterality and view position.
- Spatial metadata: Image metadata includes dimensions and bounding-box coordinates defined by xmin, ymin, xmax, and ymax.The coordinate fields specify the box’s left, top, right, and bottom boundaries.
- Labels: Breast-level records encode BI-RADS assessment and density, while finding records include finding categories and finding-level BI-RADS assessment.Finding categories can include combinations such as a mass with skin retraction.
- Dataset metadata: The split field identifies whether an image belongs to the training or test set, and standard projections are CC and MLO.Laterality is encoded as L or R.
Technical Validation
Technical validation combined privacy review with manual inspection of image content and rule-based checks on radiologist-generated labels. The dataset also reports a patient-age distribution across exams with available age information.
- Privacy validation: All metadata was manually reviewed to remove personally identifiable health information and satisfy privacy regulations including HIPAA and GDPR.Image content was also reviewed case-by-case to ensure that patient information did not remain visible.
- Label validation: Rule-based checks automatically verified radiologist-generated labels to reduce mechanical annotation mistakes.The rules checked issues such as missing global labels or lesion markings inconsistent with selected labels.
- Dataset statistics: Figure 4 reports the distribution of patient age across examinations for which age was available.The statistic is calculated over all exams with recorded patient age.
Usage Notes
VinDr-Mammo supports development and evaluation of computer-aided algorithms, but its lack of pathology-confirmed ground truth limits direct diagnostic evaluation.
- Usage Notes: VinDr-Mammo was created to develop and evaluate computer-aided detection and diagnosis algorithms based on full-field digital mammography.It also supports computer-vision tasks including object detection and multilabel image classification.
- Usage Notes: The dataset uses radiologist consensus as ground truth, but lacks pathology-confirmed labels and essential molecular and histology information.Consequently, the annotations rely heavily on radiologists’ expertise.
- Usage Notes: VinDr-Mammo should not directly evaluate CAD for diagnosis, but can be used for training because pathology reports provide incomplete support.For screening evaluation, its five BI-RADS categories can be converted to a three-category system.
Competing interests
The study was funded by Vingroup JSC, which had no role in the study’s design, data collection, analysis, publication decision, or manuscript preparation.
- Competing interests: Vingroup JSC funded the work but had no role in study design, data collection, analysis, publication decisions, or manuscript preparation.