Research

Project 1. Predicting up to 10 year breast cancer risk using longitudinal mammographic screening history

Recent deep learning (DL) risk models based on mammography have shown superiority in short-term risk prediction compared to traditional risk factor-based models. However, those models primarily rely on single-time exams and thus ignore the temporal changes in breast tissues that can be extracted from sequence exams. Here, we present the Multi-Time Point Breast Cancer Risk Model (MTP-BCR) a novel temporospatial DL risk model that integrates traditional BC risk factors and longitudinal mammography data to identify subtle changes in breast tissue indicative of future malignancy.

Project 2. Ordinal Learning: Longitudinal Attention Alignment Model for Predicting Time to Future Breast Cancer Events from Mammograms

Rethinking the mammogram based risk prediction model: precision and explainability are vital in breast cancer risk assessment for developing personalized screening and prevention strategies. We introduce OA-BreaCR, a method that precisely evaluates both the likelihood and timing of future breast cancer occurrence using sequential mammograms. To enhance Precision: By employing ordinal learning, OA-BreaCR models temporal information based on the 'time-to-future-event' ordering among patients, improving the precision of time predictions. To improve Explainability: The method utilizes an attention alignment mechanism to effectively track high-risk breast tissue changes over time, enhancing the model's interpretability.

Project 3. Global-local learning for explainable breast cancer risk prediction from screening mammograms

To mimic the process of mammogram reading by radiologists, which firstly figure out globally high-risk characteristics (e.g. dense tissue, architectural distortion et al.), and then further analyze the local suspicious areas (such as micro-calcifications), we propose a global-local transformer framework-based (GL) risk prediction model.