Overview

Schema Version

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

Name

AI-integrated Screening to Replace Double Reading of Mammograms: A Population-wide Accuracy and Feasibility Study

Link

https://pubs.rsna.org/doi/10.1148/ryai.230529

Indexing

Keywords: Mammography, Breast, Neoplasms-Primary, Screening, Epidemiology, Diagnosis, Convolutional Neural Network (CNN)

Content: BR, OI

Author(s)

Mohammad T. Elhakim

Sarah W. Stougaard

Ole Graumann

Mads Nielsen

Oke Gerke

Lisbet B. Larsen

Benjamin S. B. Rasmussen

Organization(s)

Odense University Hospital, Department of Radiology

Odense University Hospital, Department of Nuclear Medicine

CAI-X–Centre for Clinical Artificial Intelligence, Odense University Hospital

University of Southern Denmark, Department of Clinical Research, Research and Innovation Unit of Radiology

Aarhus University Hospital, Department of Radiology

Aarhus University, Department of Clinical Research

University of Copenhagen, Department of Computer Science

Version

1.1.7.1

License

Text: CC BY 4.0

URL: https://creativecommons.org/licenses/by/4.0/

Contact

Mohammad T. Elhakim; email: kd.dysr@etm

Funding

Supported by the Innovation Fund by the Region of Southern Denmark (grant 10240300).

Ethical review

Retrospective study approved by the national ethics committee (identifier D1763009); informed consent waived.

Date

Updated: 2024-08-23

Published: 2024-09-04

Created: 2023-11-17

References

[1] Elhakim MT, Stougaard SW, Graumann O, Nielsen M, Gerke O, Larsen LB, Rasmussen BSB. "AI-integrated Screening to Replace Double Reading of Mammograms: A Population-wide Accuracy and Feasibility Study". Radiology: Artificial Intelligence. 2024;6(6):e230529. 2024-09-04. doi:10.1148/ryai.230529. PMID: 39230423. PMCID: PMC11605135.

Model

Architecture

Deep learning; Convolutional Neural Network (commercial system).

Availability

Commercial, CE-marked and U.S. FDA–approved product: Lunit INSIGHT MMG version 1.1.7.1 (used in this study).

Clinical benefit

Potential to reduce screen-reading workload by about 50% without reducing cancer detection accuracy, depending on deployment scenario.

Clinical workflow phase

Screening workflow decision support and triage within double-reading mammography programs.

Decision threshold

Scenarios 1 and 2: AI abnormality score threshold 80.99% (AIspec). Scenario 3 triage: low-risk <3.36% (no recall), high-risk ≥95.29% (recall), remaining 50% routed to standard combined reading.

Degree of automation

Simulated replacement of one reader (first or second) or triage replacement of both readers for low- and high-risk cases; moderate-risk cases read by humans.

Indications for use

Concurrent reading aid for breast cancer detection at mammography (screening population).

Input

Screening mammograms (standard screening views) from a population-wide program.

Instructions

Examination-level abnormality score is the maximum per-breast score (0–100%). Dichotomize using validated threshold(s) depending on workflow scenario. For scenarios requiring arbitration when AI and human decisions disagree and no original arbitration exists, simulate arbitration to approximate original arbitrator accuracy.

Limitations

All scenarios were simulations; real-life effects and potential automation bias are unknown. Reference standard validity may be affected by differing follow-up tests, correlation between radiologist decisions and reference standard, and lack of verification for AI-recalled cases. Thresholds were validated and selected on the same sample, potentially overestimating AI accuracy. Image types unsupported by AI were excluded.

Output

CDEs: RDE784, RDE926, RDE1586.0

Description: Per-view lesion scores combined into per-breast abnormality scores (0–100%); examination-level score is the maximum of either breast; dichotomized to recall/no recall per scenario thresholds; triage categorization into low, moderate, and high risk.

Recommendation

Triage-based approach achieved best overall performance metrics, but current guidelines preclude replacing both readers; replacing the first reader with AI is a feasible implementation pathway.

Regulatory information

Comment: Commercial AI system Lunit INSIGHT MMG version 1.1.7.1 used.

Authorization status: Conformité Européenne–marked and U.S. Food and Drug Administration–approved (as stated in article).

Reproducibility

Arbitration decisions in cases without original arbitration were simulated to match original arbitrator accuracy; thresholds chosen to target ~50% workload reduction across scenarios.

Use

Intended: Detection and diagnosis

User

Intended: Subspecialist diagnostic radiologist, Administrator