Mammographic density is a dynamic risk factor for breast cancer and affects the sensitivity of mammography-based screening. While automated machine and deep learning-based methods provide more consistent and precise measurements compared to subjective BI-RADS assessments, they often fail to account for the longitudinal evolution of density. Many of these methods assess mammographic density in a cross-sectional manner, overlooking correlations in repeated measures, irregular visit intervals, missing data, and informative dropouts. Joint models, however, are well-suited for capturing the longitudinal relationship between biomarkers and survival outcomes. We present the DeepJoint algorithm, an open-source solution that integrates deep learning for quantitative mammographic density estimation with joint modeling to assess the longitudinal relationship between mammographic density and breast cancer risk. Our method efficiently analyzes processed mammograms from various manufacturers, estimating both dense area and percent density--established risk factors for breast cancer. We utilize a joint model to explore their association with breast cancer risk and provide individualized risk predictions. Bayesian inference and the Monte Carlo consensus algorithm make the approach reliable for large screening datasets. Our method allows for accurate analysis of processed mammograms from multiple manufacturers, offering a comprehensive view of breast cancer risk based on individual longitudinal density profiles. The complete pipeline is publicly available, promoting broader application and comparison with other methods.
翻译:暂无翻译