Notes by Lex Toumbourou

Summary

VinDr-Mammo is a large-scale Vietnamese full-field digital Mammography (FFDM), which consists of 5,000 exams (20,000 images) created to support Computer-Aided Detection (CADe) and Computer-Aided Diagnosis (CADx) research.

Contains 4 standard views for each patient (L/R CC and MLO) and provides:

• Breast-level BI-RADS assessment.
• Breast density (ACR categories A-D)
• Extensive lesion-level bounding-box annotations for masses, calcifications, asymmetries, distortions and other suspicious features.
• Double-reading with abritration. That is, two radiologists read each exam, if they disagree, a third one resolve the conflict.

One of the largest, publically accessible FFDM datasets.

Dataset

• Time window: 2018–2020
• Equipment vendors: Siemens, IMS, Planmed
• All images are for-presentation DICOMs (no raw for-processing images saved by hospitals).
• Images de-identified:
  • DICOM PHI scrubbed
  • Text burned into corners masked out
• Source:
  • Images sourced from two major hospitals in Hanoi, Vietnam:
    • Hospital 108 (H108)
    • Hanoi Medical University Hospital (HMUH)
• Format: DICOM
• Views: CC & MLO (left & right)
• Total: 20,000 images / 5,000 exams
• Stored in per-study folders with encoded IDs.

Labels and Annotations

• Breast-level (per image)
  • BI-RADS scale:
    • 1, 2, 3, 4, 5 (no 6; pathology unavailable)
  • Breast density: A-D
  • Metadata also has: laterality, view, image size, split.
• Lesion level:
  • Bounding box coordinates
  • Categories:
    • mass
    • suspicious calcification
    • asymmetry (global / focal)
    • architectual distortion
    • associated findings: skin thickening, skin retraction, nipple retraction, suspicious lymph node
  • Each lession includes own BI-RADS assessment.
• Annotation protocol:
  • Double-reading by experienced radiologists.
  • Arbitraty third read if disagreement.
  • Readers average 19 years clinical experience and interpret ~10k-15k mammograms/year
  • Web-based tool: VinDR Lab (OHIF-based viewer)
• Stratified Data Split
  • 4,000 training exams
  • 1,000 test exams
  • Iterative stratification preserving distributions across:
    • BI-RADS
    • density
    • lesion categories

Technical Validation

• Privacy & Data Quality
  • DICOM metadata reviewed manually post-scrubbing
  • Image inspection to ensure absence of PHI
  • Annotation errors prevented by built-in QA rules (e.g., cannot mark a lesion while selecting BI-RADS 1)
• Dataset Characteristics (Selected)
  • BI-RADS distribution (overall 10,000 breasts):
    • BI-RADS 1: 67.0%
    • BI-RADS 2: 23.4%
    • BI-RADS 3: 4.7%
    • BI-RADS 4: 3.8%
    • BI-RADS 5: 1.1%
• Density distribution:
  • A: 0.5%
  • B: 9.5%
  • C: 76.5% (dominant)
  • D: 13.5%
• Lesion counts (entire dataset):
  • Masses: 1,226
  • Suspicious calcifications: 543
  • Asymmetries (global, focal): 392
  • Architectural distortions: 119
  • Lymph nodes / retractions: <60 each

Significance

VinDr-Mammo addresses several key gaps in mammography AI research: 1. Scale & diversity:
* Largest public FFDM dataset with rich annotation detail. 2. Population representation:
* Introduces Southeast Asian data—critical given known domain shift across ethnicities, vendors, and clinical pipelines. 3. Realistic screening distribution:
* Mix of diagnostic + screening exams, and clinically realistic prevalence. 4. Bounding-box–level lesion annotations:
* Suitable for CADe, object detection, and weakly supervised methods. 5. Standardized split ensures reproducibility across publications.

Limitations:
- No pathology-proven labels → BI-RADS 4/5 are only suspicious, not confirmed malignancies.
- Some findings have very low sample counts (<40).
- DICOMs are not fully compliant with all processing libraries.

Comparisons

Moreira et al. (2012) - INbreast: Toward a Full-field Digital Mammographic Database

Feature	INbreast	VinDr-Mammo
Size	115 cases	5,000 exams — ~40× larger
Population	Portuguese	Vietnamese (underrepresented population)
Annotation precision	Pixel-accurate contours	Bounding boxes (faster + scalable)
Pathology labels	Included for many lesions	Not available
Imaging	Siemens FFDM	Siemens / IMS / Planmed
Use cases	CADx, segmentation	CADe, detection, multi-label classification

Key distinction:
INbreast remains the gold standard for pixel-level precision, but is tiny; VinDr-Mammo is designed for scale, population diversity, and bounding-box detection workflows.

CBIS-DDSM Mammography Dataset

Feature	CBIS-DDSM	VinDr-Mammo
Imaging	Digitised film	Full-field digital mammography
Size	~1,644 cases	5,000 exams
Lesion types	Mass, calcification	Mass, calcification, distortion, multiple asymmetries, associated features
Annotations	Updated segmentations	Bounding boxes
Pathology labels	Yes	No
Image quality	Variable (film artefacts, scanner noise)	High-quality clinical FFDM

CBIS-DDSM is rich and pathology-linked but suffers from film-scanner limitations.
VinDr-Mammo provides modern FFDM, higher consistency, vastly larger scale, and more comprehensive lesion categories.

Reference

• Nguyen, H. T., Nguyen, H. Q., Pham, H. H., et al. (2023). VinDr-Mammo: A large-scale benchmark dataset for computer-aided diagnosis in full-field digital mammography. Scientific Data, 10(1), 277. https://doi.org/10.1038/s41597-023-02100-7