Leveraging Multi-Modal AI for Selective Use of Ultrasound in Breast Cancer Screening

Project Summary Supplemental breast screening ultrasound is recommended for women with dense breasts, accounting for nearly half of U.S. women over age 40. While breast ultrasound improves cancer detection rates, its cost-effectiveness remains low for women at average risk of breast cancer, with a cost-effectiveness ratio of $728,000 per quality- adjusted life year (QALY) gained. This limited efficiency is driven by high false-positive rates in breast ultrasound interpretation, leading to unnecessary diagnostic workups and biopsies, and by one-size-fits-all guidelines that recommend supplemental ultrasound for all women with dense breasts. This proposal aims to improve the diagnostic accuracy and cost-effectiveness of breast cancer screening by building an AI system to determine the necessity of supplemental screening ultrasound. Our core hypothesis is that in a substantial portion of the screening population, it is possible to forgo currently guideline-recommended supplemental ultrasounds without impacting diagnostic accuracy. To verify this hypothesis, we propose the following aims. First, we will build a multi-modal deep neural network, Multi-Modal Diagnoser (MMD), that integrates current and prior mammographic and ultrasound imaging data to improve detection of breast cancer. To train MMD, we will assemble a substantial dataset of over 2.12 million exams (420,000 patients) from our institution, as well as two external datasets from other institutions for independent evaluation. MMD will serve as the foundation for our second model, the Ultrasound Benefit Predictor (UBP). UBP will utilize screening mammography images to determine the necessity of supplemental ultrasound for women with dense breasts through an interpretability mechanism that highlights specific mammographic regions where ultrasound could add diagnostic value. Finally, we will critically assess these two elements of our AI system through a retrospective study, examining their impact on radiologists' performance and cost-effectiveness through the lens of patients’ QALY. This project seeks to pioneer the integration of AI in breast cancer screening, leading to more personalized patient care and optimized healthcare resource allocation.