Breast Imaging in Focus: Can AI Models for Breast Cancer Detection and Risk Prediction Be Used Interchangeably?

Artificial intelligence (AI) is increasingly being used in breast cancer screening for two purposes: detecting breast cancers on mammography and predicting future risk of breast cancer. Although closely related, these applications address different clinical questions. A new study, published in Radiology Imaging Cancer, evaluated whether AI models developed for breast cancer detection and short-term risk prediction can be used interchangeably.

This retrospective study used a Swedish mammography dataset containing 20,187 mammograms from 7430 women who underwent screening between 2008 and 2016. Among these exams, 741 were associated with a breast cancer diagnosis within two years, including 524 screen-detected cancers and 217 interval cancers.¹

Four AI models were evaluated: two detection models, RSNA Detection and Transpara Detection (ScreenPoint Medical), and two risk prediction models, Mirai Risk and Transpara Risk (ScreenPoint Medical). Each model was assessed in two settings. The detection setting included screen-detected and interval cancers whereas the short-term risk setting focused on predicting interval cancers following negative screening exams. Performance was evaluated using the area under the receiver operating characteristic curve (AUC), along with sensitivity and specificity at clinically relevant thresholds.

What the Research Revealed About Mammography-Based AI Tools for Breast Cancer Detection and Risk Prediction

There were three important findings. First, in the detection setting, Transpara Detection and Transpara Risk performed similarly with an AUC of 0.92 for both models in comparison with 0.90 for RSNA detection and 0.88 for Mirai risk. However, at the sensitivity of radiologist double reading, which was 70.7 percent, Transpara Detection achieved a specificity of 97.1 percent in comparison with 94 percent for Transpara Risk. Thus, despite similar AUCs, the detection model demonstrated higher specificity at this operating threshold.¹

Second, in the short-term risk setting, Transpara Risk performed best with an AUC of 81 percent compared with 76 percent for both Transpara Detection and Mirai Risk, and 72 percent for RSNA Detection. At 90 percent specificity, Transpara Risk identified 49.8 percent of interval cancers compared with 41 percent for Transpara Detection. These findings suggest that models optimized for risk prediction may be better suited to identifying women who subsequently develop interval cancer.¹

Third, the investigators examined agreement between Transpara Detection and Transpara Risk. At 90 percent specificity, only 11.1 prcent of interval cancers were identified by the risk model without also being flagged by the detection model. According to the study, 47.9 percent of interval cancers were missed by both models. These results demonstrate overlap between detection and risk prediction while highlighting the limitations of both approaches.¹

Several study limitations should be considered. This study was a retrospective case-control study and multiple exams from the same woman were included. The models were evaluated independently rather than in combination with radiologists, and some patient information required by the transfer models had to be imputed. In addition, all exams were obtained from a single mammography vendor in a Swedish screening population, limiting generalizability. Finally, the study evaluated model performance rather than clinical outcomes.

Key Considerations with AI Detection and Risk Prediction Models in Breast Imaging

These findings have important implications for breast cancer screening. First, AI detection and risk prediction models should not be considered interchangeable even when their performance overlaps. Second, AI-based risk prediction models could help identify women who would benefit from supplemental screening, but whether this approach improves cancer detection or reduces interval cancers remains uncertain. Third, the agreement analysis raises questions about the incremental benefit of combining detection and risk prediction models as relatively few interval cancers were identified exclusively by the risk prediction model.

The key takeaway is that AI-based breast cancer detection and risk prediction models are related but distinct applications. Models should be selected and validated based on their intended clinical purpose. Future prospective studies are needed to determine whether integrating these models into screening workflows improves patient outcomes.

Dr. Bahl is an associate professor at Harvard Medical School and a breast radiologist at Massachusetts General Hospital in Boston.

Reference

1. Wang YK, Cockmartin L, Wagner T, et al. Comparison of artificial intelligence detection and short-term breast cancer risk models within and beyond their intended use in mammography screening. Radiol Imaging Cancer. 2026;8(5):e250609. doi: 10.1148/rycan.250609.