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

ADMANI: Annotated Digital Mammograms and Associated Non-Image Datasets

https://doi.org/10.1148/ryai.220072

Helen.Frazer@svha.org.au

Supported in part by an Australian government grant (no. MRFAI000090), as part of the 2019 Medical Research Future Fund (MRFF) Applied Artificial Intelligence Research in Health grant opportunity, which funded the Transforming Breast Cancer Screening with AI (BRAIx) program. The BRAIx program also includes significant in-kind contributions from its grant partners: St Vincent’s Institute of Medical Research, St Vincent’s Hospital Melbourne, BreastScreen Victoria, University of Melbourne, and University of Adelaide. This work was also supported by a Ramaciotti Health Investment Grant from the Ramaciotti Foundation in Australia.

Use of the ADMANI datasets is governed under the executed BRAIx Multi-Institutional Agreement, with approvals by the human research ethics committee (approval nos. LNR/18/SVHM/162 and LNR/19/SVHM/123). All women sign a consent form at screening registration that provides for the use of the de-identified data for research purposes. A unique identifier is used for the purposes of the ADMANI datasets, with all image and non-image data de-identified.

The ADMANI datasets (ADMANI1, ADMANI2, and ADMANI3) are large-scale, multicenter, clinically curated breast screening mammographic datasets created for artificial intelligence algorithm development. They contain over 4.4 million images from 629,863 women and 1,048,345 screening episodes, enhanced with associated demographic and clinical non-image data.

To enable the development of AI-based algorithms to aid breast cancer detection in the mammographic screening population and support risk-based screening.

The ADMANI2 and ADMANI3 datasets are large-scale, population-based, longitudinal resources reflecting the real-world screened population. The RSNA Challenge subset was randomly selected from the dataset from a 3-year period.

The ADMANI datasets are structured into three temporal subsets (ADMANI1, ADMANI2, ADMANI3) covering different screening periods. A separate subset of 40,000 images from 10,000 episodes was randomly selected for the RSNA Breast Cancer Detection AI Challenge.

Interval cancer data from 2018 and 2019 are awaiting updates. Digital breast tomosynthesis images and breast ultrasound images are yet to be included.

The datasets are structured around individual screening episodes, defined as a single screening round including mammography, reading, assessment, and the subsequent 2-year screening interval.

All image and non-image data are de-identified, and a unique identifier is used for the ADMANI datasets.

It is not possible to identify individuals as all image and non-image data are de-identified and a unique identifier is used.

Patient data collected includes age, country of birth, risk category, symptoms (e.g., lump, nipple discharge), and use of hormone replacement therapy. Radiologist reading data includes lesion side and grade. Histopathologic data contains surgical specimen results, including lesion subtype. All data is de-identified.