Toward explainable AI approaches for breast imaging: adapting foundation models to diverse populations

Foundation models can be adapted for medical imaging across diverse populations: The study successfully adapted BiomedCLIP, a foundation model, for breast density classification and demonstrated strong generalization on external datasets (RSNA and EMBED), with AUC values ranging from 0.80-0.93, suggesting these models can work reliably across different patient populations and imaging settings.

Multi-modality training improves AI system flexibility without sacrificing accuracy: The multi-modality model achieved comparable accuracy (0.74) to single-modality approaches (0.73) while maintaining the ability to work across different imaging types (synthesized 2D, digital mammography, and digital breast tomosynthesis), making the AI system more practical for real-world clinical deployment where imaging equipment varies.

Explainability tools help verify that AI systems focus on clinically relevant features: GradCAM visualizations confirmed the models consistently highlighted clinically appropriate areas in mammograms, providing transparency into how the AI makes decisions—a critical requirement for responsible deployment in healthcare where clinicians need to understand and trust AI recommendations.

Addressing class imbalance through weighted contrastive learning improves fairness: The study used weighted contrastive learning to handle imbalanced data, which helps prevent the AI from performing poorly on underrepresented groups—an important consideration for ensuring the technology works equitably across diverse patient populations.