CHICAGO -- An image-based breast cancer risk prediction model achieved success within a real-world clinical screening environment, suggest preliminary findings presented November 30 at RSNA 2025.
Georgia Spear, MD, from Northwestern Medicine in Chicago, IL, presented research she and colleagues led, showing how the Mirai deep-learning model demonstrated robust performance across different clinical scenarios and imaging modalities. These included full-field digital mammography and 2D synthetic mammography images.
Georgia Spear, MD, discusses her team's preliminary findings showing the effectiveness of the Mirai breast cancer prediction model.
“Adopting image-based breast cancer risk prediction models could significantly streamline screening workflows,” Spear said.
While major guidelines stress the need for risk-based breast cancer assessment, they don’t agree on how risk is interpreted and applied. Consequences from this divergence include creating inconsistencies in identifying who is at high risk of disease and who receives supplemental imaging. This can also translate into inequitable care, Spear said.
The one-size-fits-all approach persists in breast imaging, though large-scale trials are assessing the effectiveness and feasibility of risk-based screening approaches.
Trained on full-field digital mammograms, Mirai has shown promise in predicting breast cancer and detecting interval cancers in recent studies. Spear said real-world evaluation is needed to determine its clinical utility and generalizability.
Spear and colleagues evaluated Mirai’s clinical performance within a routine mammography screening program. It assessed Mirai’s discriminatory accuracy to determine its performance in clinical practice with a mix of mammography unit models.
Preliminary analyses focused on evaluating discrimination. The researchers compared results from 1,763 breast cancer cases to those of a control group (n = 56,953).
The Mirai model achieved strong overall discriminatory performance, having consistent accuracy across various imaging types. For women who developed cancer following a negative screening mammogram, Mirai outperformed the Tyrer-Cuzick model for predicting cancer (23.6% vs. 18%, respectively). It performed even better within six months of a negative screening mammogram at 36.3% for predicting cancer development.
The two-year area under the curve (AUC) values for the risk model were statistically significantly different between the machine types (p = 0.002). And Mirai showed comparable efficacy when applied to synthetic 2D images.
Spear said the results support Mirai’s integration into clinical workflows. This could make way for breast cancer screening strategies that are more personalized, efficient, and scalable.
“Future work includes stakeholder engagement to determine how best this could be used in our practice and what further research is required before widespread adoption of risk-based assessment,” Spear added.
Visit our RADCast for full coverage of RSNA 2025.
