Predictive Impact of Ultrasound for Breast Cancer Screening

Breast cancer remains a major cause of cancer-related mortality among women. While mammography has long been the primary screening modality, personalizing supplemental imaging recommendations remains limited in practice due to cost and resource constraints. Ultrasound has gained traction as a supplemental modality for women with dense breasts, but its broader diagnostic utility to improve cancer diagnosis remains less well understood. In this work, we study the predictive impact of ultrasound, defined as the additional predictive value it provides over mammography for breast cancer screening. We first characterize the predictability of breast cancer from mammography and ultrasound individually, as well as from their interaction. We show that, among patients for whom both modalities were collected in the NYU Langone Health Breast Cancer Dataset between 2012 and 2020 (a subpopulation predominantly consisting of women with dense breasts), ultrasound contributes substantially more to prediction than mammography or their interaction. At the patient level, we model ultrasound acquisition as an individual decision problem using a latent-variable framework. Given only a mammogram, we generate plausible latent representations and their corresponding predictions. The variation in these predictions provides a patient-level estimate of the predictive impact of ultrasound. We find that this measure captures model-defined ultrasound utility (expected patient-level predictive gain) better than conventional breast-density-based selection, providing a step toward more personalized ultrasound acquisition policies for breast cancer screening.