Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/dataset.json

Name

BreastScreen South Australia (BSSA) screening mammography dataset

Link

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11294949/

Indexing

Keywords: Screening mammography, Deep learning, Transfer learning, Breast cancer, Australia, BreastScreen SA, NYU1, NYU2

Content: BR, OI, RS

RadLex: RID45682, RID10357

SNOMED: 254837009, 1162814007

Author(s)

James J. J. Condon

Vincent Trinh

Kelly A. Hall

Michelle Reintals

Andrew S. Holmes

Lauren Oakden-Rayner

Lyle J. Palmer

Organization(s)

Australian Institute for Machine Learning, University of Adelaide

School of Public Health, University of Adelaide

BreastScreen South Australia (BSSA)

Contact

Lyle J. Palmer (Corresponding author): ua.ude.edialeda@remlap.elyl

Funding

Authors declared no funding for this work.

Ethical review

Approved by local institutional review board (HREC/16/229, R20160601) with waiver of consent for retrospective use of de-identified clinical data.

Comments

Retrospective dataset from a South Australian public mammography screening program used to evaluate and retrain NYU deep learning models (NYU1 and NYU2) for breast cancer screening.

Date

Published: 2024-05-08

References

[1] Condon JJJ, Trinh V, Hall KA, Reintals M, Holmes AS, Oakden-Rayner L, Palmer LJ. "Impact of Transfer Learning Using Local Data on Performance of a Deep Learning Model for Screening Mammography". Radiology: Artificial Intelligence. 2024-05-01. doi:10.1148/ryai.230383. PMID: 38717291. PMCID: PMC11294949.

Dataset

Motivation

Assess generalizability/replication of NYU mammography deep learning models on an Australian screening dataset and evaluate impact of local retraining with transfer learning.

Sampling

All individuals with biopsy or surgical pathology–proven lesions (n=3160) and age-matched controls (n=3240) identified from all individuals screened 2010–2016; controls required no concerning abnormalities at index and at least one subsequent round.

Partitioning scheme

Study sample randomly split 70% training, 15% validation, 15% test, stratified by dominant finding; two local test sets were defined: LTS1 (primary) and LTS2 (lower prevalence).

Missing information

Breast density not routinely collected in BSSA; interval cancer data not available.

Relationships between instances

Analyses conducted per-individual; each screening round comprised at least four standard views (bilateral CC and MLO). Controls were age-matched on average within 1 year.

Noise

Subtle visual differences in image contrast/opacity and differences in image sizes across datasets; different acquisition vendors between NYU and BSSA.

External data

NYU models and patch-level heatmap network trained on NYU radiologists’ hand-drawn benign (n=3158) and malignant (n=855) segmentations were used for inference; these segmentations are publicly available per cited resource.

Confidentiality

De-identified clinical data with IRB approval and waiver of consent.

Sensitive data

Health imaging data from a public screening program; de-identified.