Mirai AI reveals mammographic phenotypes tied to 5-year breast cancer risk
Mammogram AI model identifies tissue patterns and hidden shortcut artifacts like surgical clips.
Researchers at NYU (Jia et al.) developed a new interpretability framework for mammogram-based deep learning models by clustering patch embeddings from the pre-trained Mirai model. Rather than relying on single-image saliency maps, they extracted recurring phenotypes across a large patient cohort and linked them to 5-year breast cancer risk scores.
Their analysis revealed that risk-increasing phenotypes capture complex structures such as dense tissue and microcalcifications, but also shortcut artifacts like surgical clips—highlighting latent confounders. The phenotypes strongly correlated with older age and higher BI-RADS breast density. This work bridges tissue-level patterns and AI risk predictions, offering a path to uncover both genuine clinical signatures and hidden biases in breast cancer risk models.
- Clustered patch embeddings from Mirai model to identify recurring mammographic phenotypes linked to 5-year risk
- Risk-increasing phenotypes include dense tissue, microcalcifications, and shortcut artifacts (e.g., surgical clips)
- Phenotypes correlate with older age and higher BI-RADS density, revealing clinical signatures and potential confounders
Why It Matters
Exposes both biological risk markers and hidden biases in AI breast cancer screening models.