Mammography AI triage across age groups, datasets and reader assessments

Objective. To determine whether a mammography AI policy can identify benign breasts eligible to forgo review while limiting the fraction of cancers set aside, and whether the same rule remains useful after transfer to another population. Methods. We evaluated a frozen five-member mammography ensemble in 1,359 biopsy-selected CMMD participants. The original study separated tuning (270), certification (406) and evaluation (683). Follow-up experiments reused the evaluation group to test age-specific calibration, score fusion, agreement and whole-breast texture. External analyses used CBIS-DDSM mass and calcification cohorts and a positive-only stress test in 117 NLBS examination directories. We evaluated simulated dismissal separately from allocating a second read. Results. The original ungated cutoff set aside 57 of 683 cases, including 32 of 529 cancers. Both age groups failed the 5% certification target. Expanded age calibration set aside 19 cancers and 19 benign breasts. Requiring two scores to agree reduced these counts to five and six, respectively. The transferred malignancy-only rule set aside 82 of 117 NLBS cancer directories. In 96 CBIS-DDSM mass participants the transferred contrast rule set aside nobody; local recalibration subsequently set aside two of 23 cancers in its 49-person evaluation split. In CMMD, adding whole-breast texture to a two-score model set aside 11 additional cancers for one additional benign case. At a fixed 10% set-aside volume, the hybrid placed six of 44 calcification cancers in the ten lowest-ranked cases. At a separate 25% second-read budget, grade followed by AI flagged seven of 14 low-assessment mass cancers and neither of two in a test-release subset. Conclusions. More conservative rules reduced omissions at the cost of benign yield. Recalibration restored some yield after transfer, but could not remove cancers already ranked among the least suspicious cases. The evidence supports distinct tests of calibration, ranking and reader workload before a triage policy is used in screening. Third-party images retain the licenses credited in the figure captions.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-19
DOI
https://doi.org/10.5281/zenodo.24369354
Primary Topic
AI in cancer detection
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Mammography AI triage across age groups, datasets and reader assessments

S. M. Abdulla Al Mamun
Zenodo (CERN European Organization for Nuclear Research)
AI in cancer detection
preprint

Mammography AI triage across age groups, datasets and reader assessments

S. M. Abdulla Al Mamun
preprint en

Abstract

Objective. To determine whether a mammography AI policy can identify benign breasts eligible to forgo review while limiting the fraction of cancers set aside, and whether the same rule remains useful after transfer to another population. Methods. We evaluated a frozen five-member mammography ensemble in 1,359 biopsy-selected CMMD participants. The original study separated tuning (270), certification (406) and evaluation (683). Follow-up experiments reused the evaluation group to test age-specific calibration, score fusion, agreement and whole-breast texture. External analyses used CBIS-DDSM mass and calcification cohorts and a positive-only stress test in 117 NLBS examination directories. We evaluated simulated dismissal separately from allocating a second read. Results. The original ungated cutoff set aside 57 of 683 cases, including 32 of 529 cancers. Both age groups failed the 5% certification target. Expanded age calibration set aside 19 cancers and 19 benign breasts. Requiring two scores to agree reduced these counts to five and six, respectively. The transferred malignancy-only rule set aside 82 of 117 NLBS cancer directories. In 96 CBIS-DDSM mass participants the transferred contrast rule set aside nobody; local recalibration subsequently set aside two of 23 cancers in its 49-person evaluation split. In CMMD, adding whole-breast texture to a two-score model set aside 11 additional cancers for one additional benign case. At a fixed 10% set-aside volume, the hybrid placed six of 44 calcification cancers in the ten lowest-ranked cases. At a separate 25% second-read budget, grade followed by AI flagged seven of 14 low-assessment mass cancers and neither of two in a test-release subset. Conclusions. More conservative rules reduced omissions at the cost of benign yield. Recalibration restored some yield after transfer, but could not remove cancers already ranked among the least suspicious cases. The evidence supports distinct tests of calibration, ranking and reader workload before a triage policy is used in screening. Third-party images retain the licenses credited in the figure captions.

Zenodo (CERN European Organization for Nuclear Research)
University of Manchester (GB)
AI in cancer detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon