NotesByLex

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

Lex Toumbourou · · revised

BreastCancerDetectionDeepLearningDatasetMammography

Summary

Introduces InBreast a dataset designed for the development and evalation of computer-aided detection (CADe) and computer-aided diagnosis (CADx) algorithm.

Acquired at the Breast Centre Hospitalar de São João (CHSJ), Porto, it contains precisely annotated digital Mammograms with XML-based contours of lesions and associated metadata such as BI-RADS categories, breast density, patient age, and pathology results.

Device used was MammoNovation Siemens full-field digital mammography (as opposed to digitised film images) in the datasets prior like CBIS-DDSM Mammography Dataset, with a solid-state detector of amorphous selenium was used.

Dataset Details

  • Developed by the University of Porto / INESC Porto under hospital and ethics committee approval.
  • Acquisition system: Siemens MammoNovation FFDM (amorphous selenium detector).
  • Image specs:
    • Pixel size: 70 μm.
    • Bit depth: 14-bit grayscale.
    • Resolution: 2560 × 3328 or 3328 × 4084 (depending on compression plate).
    • Format: DICOM.
  • Dataset size:
    • 115 cases, 410 images.
    • 90 cases with both breasts (4 images per case: CC + MLO).
    • 25 single-breast (mastectomy) cases (2 images per case).
  • Contains:
    • Normal, benign, and malignant images.
    • Multiple lesion types: masses, calcifications, architectural distortions, asymmetries, and spiculated regions.
  • Annotations:
    • Detailed XML contours for all regions of interest.
    • Labels for lesion type and location.
    • Seven annotation types: asymmetry, calcification, cluster (of microcalcifications), mass, distortion, spiculated region, pectoral muscle.
    • Annotation tool: OsiriX (Mac) PACS workstation.
    • Dual-specialist validation; disagreements resolved by consensus.
  • Metadata:
    • BI-RADS categories (0–6).
    • ACR breast density classification (1–4).
    • Patient age, family history, and pathology results.
    • Biopsy results included for BI-RADS 3–6.
  • Distribution:
    • Hosted at INESC Porto.
      • Which is a broken link. Found it on Google drive: https://drive.google.com/file/d/19n-p9p9C0eCQA1ybm6wkMo-bbeccT_62/view
    • Suggested train/test split included.
    • Available for public research use with ethics compliance.

Technical Validation

  • 56 biopsy-confirmed cases: 45 malignant, 11 benign.
  • Strong inter-radiologist consensus ensured annotation precision.
  • The high-resolution contours enable pixel-level segmentation evaluation, superior to DDSM’s coarse ROIs.
  • Provides distributions by BI-RADS class, lesion type, breast density, and age.
  • Demonstrates natural clinical diversity (including difficult cases and post-surgery exams), increasing model robustness potential.

Findings Breakdown

  • Lesion distribution:
  • 116 masses across 107 images (~1.1 per image).
  • 6,880 individual calcifications annotated.
  • 27 clusters of microcalcifications (MCCs).
  • Prominent representation of calcifications (reflecting clinical prevalence).
  • Lesion size: average 479 mm² (SD = 619 mm²; range = 15–3689 mm²).
  • Age distribution: broad, with representative real-world variability.
  • Density distribution: includes ACR categories 1–4.

Significance

  • First public FFDM dataset available to researchers (as of 2012).
  • Designed to overcome DDSM’s limitations in precision, compression, and clinical completeness.
  • Enables:
  • Benchmarking of deep learning models for segmentation and classification.
  • Shape-based malignancy analysis and microcalcification clustering studies.
  • Evaluation of CAD systems under realistic imaging conditions.
  • Provides an important bridge from digitized to natively digital mammography research.

Reference

  • Moreira, I. C., Amaral, I., Domingues, I., Cardoso, A., Cardoso, M. J., & Cardoso, J. S. (2012). INbreast: Toward a full-field digital mammographic database. Academic Radiology, 19(2), 236–248. https://doi.org/10.1016/j.acra.2011.09.014