Comparison of Digital Mammography Plus Full ABUS Review and Digital Mammography Plus Selective ABUS Review Guided by ABUS Artificial Intelligence–Computer-Aided Diagnosis for Breast Cancer Screening
- Author:
Inyoung YOUN
1
;
Su Min HA
;
Myoung-jin JANG
;
Mi-ri KWON
;
Jung Min CHANG
Author Information
- Publication Type:Original Article
- From:Korean Journal of Radiology 2026;27(2):111-121
- CountryRepublic of Korea
- Language:English
-
Abstract:
Objective:To compare two breast cancer screening strategies, digital mammography (DM) plus radiologist-interpreted automated breast ultrasound (ABUS) and DM plus selective ABUS review, in which only examinations positive for DM or flagged by ABUS artificial intelligence–computer-aided diagnosis (AI-CAD) were reviewed by radiologists.
Materials and Methods:This retrospective study included asymptomatic women who underwent DM and ABUS screening for breast cancer between March 2022 and March 2023. The radiologists’ interpretations of DM and ABUS without AI assistance (DM + ABUS_radiologist) were collected from the clinical radiology reports. A selective DM plus ABUS reading strategy was simulated, in which only cases interpreted as positive in the radiologist’s DM report or flagged by retrospectively applied ABUS AI-CAD were triaged for further evaluation through a full review by radiologists (DM + ABUS_AI-CAD). The cancer detection rate (CDR), sensitivity, specificity, and abnormal interpretation rate (AIR) were calculated and compared between DM + ABUS_ radiologist and DM + ABUS_AI-CAD groups using the McNemar’s test.
Results:Among 2,275 women (mean age, 56.1 ± 8.6 years), 12 cancers were diagnosed. The sensitivity, CDR and AIR for DM + ABUS_radiologist was 83.3% (10/12; 95% confidence interval [CI]: 51.6–97.9), 4.4 (10/2,275; 95% CI: 2.1–8.1) per 1,000 screening examinations and 16.7% (379/2,275; 95% CI: 15.1–18.3), respectively. DM + ABUS_AI-CAD triaged 84.0% (1,910/2,275) of the examinations as negative in both DM reports and retrospectively applied ABUS AI-CAD, requiring radiologist reassessment in only 16.0% (365/2,275). This approach reduced the AIR to 7.3% (167/2,275) and improved the specificity from 83.7% (1,894/2,263) to 93.1% (2,107/2,263) (all P < 0.001), while maintaining a CDR of 4.8 per 1,000 and a sensitivity of 91.7% (11/12) (all P > 0.999), compared to the DM + ABUS-radiologist.
Conclusion:An AI-CAD-assisted selective ABUS reading strategy reduces unnecessary recalls and improves specificity, which may help optimize reading priorities and reduce the reading workload while maintaining cancer detection performance.
