AI-assisted low-dose breast MRI can achieve noninferior cancer detection compared to standard full-dose contrast, according to the findings of an RSNA 2025 trainee research prize-winning study.
Specifically, a 75% gadolinium dosage reduction retained diagnostic sensitivity at 83.3% (p = 1) across all study groups and preserved lesion detection, said presenter Hanna Terhaar, MD, from the University of Alabama (UAB) at Birmingham Heersink School of Medicine. She shared the results December 2 at the recent RSNA meeting.
"Studies have shown that gadolinium accumulates in the brain as well as other major organs after repeated contrast-enhanced MRIs," Terhaar explained. "Patients are increasingly aware of the gadolinium retention and express a strong preference for strategies to minimize contrast exposure. Aside from safety, gadolinium also carries a high acquisition cost and environmental disposal concern."
Using simulated low-dose AI enhancement in neuroimaging and breast MRI has demonstrated diagnostic potential, Terhaar said. However, prior research has neither exclusively nor prospectively compared full-dose standard of care against true low-dose acquisitions in a cohort restricted to biopsy-proven malignancies, she noted during her talk.
To that end, Terhaar initiated a single-center study that recruited 20 female participants from breast health clinics at UAB. Those enrolled had prior breast MRI findings, were newly diagnosed, or were undergoing imaging for screening or BI-RADS 3 follow-up.
The study included several aims. The first was to determine if AI-assisted low-dose breast MRI would achieve noninferior cancer detection compared to standard full-dose protocols. The second was to compare lesion characteristics and overall exam quality, assess interreader agreement, and evaluate additional diagnostic metrics (sensitivity, specificity, positive predictive value [PPV], and negative predictive value [NPV]).
Participants underwent two MRI exams (1.5-tesla TwinSpeed, GE Healthcare): a standard dose (0.1 mmol/kg) gadolinium-enhanced MRI followed by a low-dose exam (0.025 mmol/kg).
AI was applied to enhance image quality and lesion conspicuity on both datasets, according to Terhaar, who divided the images into three groups: low dose with AI (Group 1), standard dose without AI (Group 2), and standard dose plus AI (Group 3).
She and her colleagues reported that 98 lesions were detected, leading to 12 biopsy-proven malignancies. Group 1 (low dose with AI) demonstrated the highest diagnostic accuracy, with no statistically significant differences in accuracy between the three groups.
Diagnostic accuracy of three MRI exams at varying contrast doses, with and without AI | |
Group | Percentage accuracy |
| Group 1 (low-dose gadolinium with AI) | 74.5% |
| Group 2 (standard dose without AI) | 69.5% |
| Group 3 (standard dose plus AI) | 65.5% |
Group 1 also demonstrated the highest PPV, at 50%. All three groups showed consistently high NPV at 93%, 94%, and 93%, respectively.
Finding type demonstrated the highest interreader agreement (kappa = 0.65 to 0.74) across all groups, achieving substantial reliability regardless of the protocol, according to Terhaar.
Malignancy likelihood (kappa = 0.4 to 0.53) was strong across all groups, but lesion conspicuity assessments showed greater variability, particularly in AI-enhanced groups, she said.
Terhaar said the results of the study highlight the potential for AI-enhanced low-dose breast MRI to serve as a safe and clinically viable alternative to standard gadolinium protocols.
Implementing standardized AI interpretation training could improve consistency among radiologists, she added. However, she acknowledged that the study was small, and there were no longitudinal outcomes to assess long-term diagnostic impact.
The study received the Dr. Tapan K. Chaudhuri Award and was funded by Bracco.
