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

Mia (version 2.0.1) – Kheiron Medical Technologies

https://pmc.ncbi.nlm.nih.gov/articles/PMC10245180

2.0.1

Corresponding author: Clarisse F. de Vries, email: ku.ca.ndba@seirved.essiralc (as presented in article)

Supported by the Industrial Centre for Artificial Intelligence Research in Digital Diagnostics (iCAIRD), funded by Innovate UK on behalf of UK Research and Innovation (UKRI) (project no. 104690).

Approved by Proportionate Review Subcommittee of the London-Bloomsbury Research Ethics Committee (20/LO/0563); Public Benefit and Privacy Panel approval (1920–0258). Secondary use of de-identified data; individual consent not required.

Ensemble of deep learning algorithms producing a continuous malignancy prediction value (0–1) from four standard mammography views.

Commercial product (Mia), Conformité Européenne (CE) marked per article.

Assists breast cancer screening by identifying exams to recall; detected more cancers than single human reader in study setting and flagged a subset of interval cancers.

Clinical decision support within population breast screening (triage/decision support alongside human readers).

Vendor-prespecified threshold ≥0.1117 indicates recall; site-specific thresholds calibrated for local site and mammography unit software versions (0.2712 pre-upgrade; 0.4319 post-upgrade).

Decision support; outputs a recall/no-recall recommendation based on threshold and a malignancy score; recommended to be combined with human reader input.

Assessment of routine screening full-field digital mammography in women approximately 50–70 years attending the U.K. breast screening program in a clinical screening environment.

Four standard mammography images per exam (bilateral CC and MLO full-field digital mammography).

Validate and calibrate decision thresholds locally before deployment; monitor performance through quality assurance; consider per–mammography software-version thresholds.

Performance and recall rates were affected by mammography equipment/software version; prespecified threshold did not transfer, causing high recall rates; evaluated on a single-center, predominantly White population; higher false-positive rate than routine dual reading; generalizability requires validation.

Perform local retrospective evaluation and threshold calibration; implement ongoing quality assurance and monitor for consistency across equipment/software versions.