Revealing Mammographic Phenotypes in Deep Learning Breast Cancer Risk Models

2026-06-24

AI Summary

Current deep learning models for breast cancer risk prediction lack systematic identification of reproducible mammographic phenotypes across populations, limiting interpretability of their decision-making and susceptibility to confounding factors. This study leverages image patch embeddings extracted by the pretrained model Mirai and integrates deep embedded clustering, saliency map analysis, and statistical associations with clinical variables to systematically discover and characterize reproducible phenotypes significantly associated with five-year breast cancer risk in a large-scale cohort. These phenotypes—including dense tissue, microcalcifications, and surgical clip artifacts—exhibit strong correlations with increasing age and higher BI-RADS density categories. Moreover, they reveal the complex tissue structures relied upon by AI models and highlight potential shortcut artifacts exploited during risk prediction, thereby mapping opaque risk scores onto clinically interpretable imaging features.

Abstract

Mammogram-based deep learning models have improved breast cancer risk prediction, but the learned imaging patterns remain underexplored. Existing interpretability methods rely on single-image saliency maps, failing to identify recurring mammographic phenotypes across large patient cohorts. By clustering patch embeddings from a pre-trained model, Mirai, we isolate recurring phenotypes linked to 5-year cancer risk. Analyses show risk-increasing phenotypes capture complex structures (e.g., dense tissue, microcalcifications) and shortcut artifacts (e.g., clips). These phenotypes correlate strongly with older age and higher BI-RADS density. Our framework connects tissue patterns to AI risk scores, revealing clinical signatures and potential latent model confounders.

Problem

• mammographic phenotypes
• breast cancer risk
• deep learning
• interpretability
• patch embeddings

Innovation

• mammographic phenotypes
• deep learning interpretability
• patch embedding clustering
• breast cancer risk prediction
• AI confounders