Revealing Mammographic Phenotypes in Deep Learning Breast Cancer Risk Models

Deep learning models have demonstrated significant success in predicting 5-year breast cancer risk from mammograms, yet they remain largely opaque "black boxes." While traditional interpretability methods like saliency maps highlight important regions in individual images, they fail to identify recurring, population-level imaging patterns. This study aims to bridge this gap by systematically mining and characterizing the specific mammographic phenotypes that influence AI-driven risk scores.

The researchers utilized a large dataset of over 72,000 mammograms to analyze the internal representations of the Mirai risk model. Instead of looking at individual pixels, they cropped mammograms into patches and encoded them using the model's pre-trained image encoder. By applying soft k-means clustering to these patch embeddings, the team isolated distinct, recurring phenotypes. They then used logistic regression to link the prevalence of these phenotypes to 5-year cancer risk and employed radiomic feature analysis to quantify the biological and structural characteristics of the identified patterns.

The analysis revealed that the model's risk predictions are driven by a mix of clinically meaningful and artifactual signals. High-risk phenotypes were associated with increased local heterogeneity, such as dense fibroglandular tissue, architectural distortions, and scattered microcalcifications. However, the model also captured "shortcut" signals, such as biopsy clips, which likely serve as proxies for a patient's prior clinical history rather than direct indicators of future malignancy. Furthermore, the study found that these AI-identified phenotypes correlate with established clinical risk factors like age and BI-RADS density, yet provide additional, localized information that whole-breast density metrics alone do not capture.