SCH: Robust Multimodal Longitudinal AI for Enhanced Breast Cancer Screening - The overarching objective of this project is to optimize and personalize breast cancer screening by advancing
artificial intelligence (AI) algorithms. Breast cancer is a major global health concern. Early diagnosis is key
to reducing disease burden and mortality, making breast cancer one of the few cancers regularly screened
for asymptomatic women. Accordingly, the American Cancer Society recommends annual screening for
women age 40 and older, leading to a substantial number of scans conducted yearly. Mammography, with
an estimated 39 million annual screenings in the US, is the primary breast cancer screening method. Digital
Breast Tomosynthesis (DBT) or 3D mammography has become a routine screening method for its 3D tissue
view, improving cancer detection. However, the complexity of mammograms, the high radiologist workload,
and the low prevalence of breast cancer in screening examinations challenges radiologists’ performance and
increases the risk of false positives and missed diagnoses. Computer-aided detection (CAD) systems have
been introduced to assist radiologists, but suffer from high false positives, straining healthcare systems and
causing needless worry for patients. Recent advancements in deep learning (DL), a subfield of AI, have
enabled the development of new CAD (AI-CAD) systems that can transform clinical decision-making.
However, AI-CAD faces limitations when dealing with DBT images. Current DL models often rely on just
current scans, lack real-world generalizability, and struggle with the high-resolution DICOM format and
complexity of DBT images. To enhance breast cancer screening, we propose a more realistic AI-CAD
system, which mimics a radiologist's approach to breast cancer screening by comparing current and prior
mammograms, incorporating clinical risk factors, and accounting for image configuration variations.
Furthermore, we consider practical implementation of such a system for clinical use.
This project aims to achieve its main objective through three specific goals: i) to achieve a robust,
explainable, and efficient representation of DBT images for cancer identification by employing self attention
graph learning models; ii) to develop a multimodal longitudinal learning model for integrating history images
and clinical information, thereby creating a more human-like intelligence CAD system to improve breast
cancer detection and localization; iii) to practically implement AI-CAD for clinical use by utilizing expert-in-
the-loop continual learning to ensure model adaptivity to changes in practice, and an algorithm-hardware
codesign model compression to ensure computational feasibility. Furthermore, the project will provide a
comprehensive DBT database. In this project both publicly available and in-house DBT images will be
used to train, evaluate, and validate the proposed models. Due to the widespread use of mammography
and critical importance of early detection in patient survival, successful completion of this project will
provide a means to transform breast cancer diagnosis through greater accuracy, thereby improving patient
outcomes and reducing healthcare cost. The innovative developed algorithms will contribute to advancing
DL especially in biomedical image analyses and can be used for analyzing other 3D biomedical images
such as CT scans and brain MRI.
