Generative AI-Enhanced Microcalcification Detection in Full-Field Digital Mammography: Reducing False Positives with High Sensitivity

While generative AI shows promise across industries, its advantages over state-of-the-art segmentation AI in radiology remain unclear. This study addresses this gap by developing a generative AI model to refine calcification detection in full-field digital mammograms (FFDM), improving specificity while maintaining high sensitivity. A segmentation AI initially extracted calcification pixels, achieving 98.0% sensitivity but a low positive predictive value (PPV) of 3.2%. To enhance detection, our generative AI used the segmentation AI output as a structural prior, transforming calcification-positive pixels into calcification-free pixels and generating a corrected result by subtraction. Trained on true calcification-free regions, it categorized densities within and around each calcification as interior or exterior. Our approach improved PPV by 2.28-fold (from 3.2% to 7.3%), surpassing prior generative AI models by 146-fold (from 0.05% to 7.3%), while maintaining sensitivity above 95%. It also reduced patient-level detection errors for small calcifications (5.17-fold, from 27.43% to 5.31%), high-exterior density (4.20-fold, from 53.85% to 12.82%), and low-interior density (2.89-fold, from 63.95% to 22.09%). This study serves as a seminal reference, demonstrating generative AI’s radiological significance beyond the latest segmentation models, with potential to redefine screening accuracy and generate high-fidelity virtual normal FFDM references.