Collaborative Research: CISE MSI: RCBP: SCH: Advancing Breast-Cancer Detection in Ultrasound Imaging through Active- and Weakly-Supervised Learning

Early detection of breast cancer is critical to decreasing mortality, and breast-ultrasound imaging is commonly employed in early diagnosis due to its widespread availability, portability, and affordability. Yet, breast ultrasound is inherently noisy and of low contrast, characteristics that challenge its effectiveness at breast-cancer diagnosis and hinder application of automated deep-learning methods. Furthermore, the training such state-of-the-art deep learning requires extensive training samples equipped with costly manual labeling by radiologists. Therefore, this project aims to develop breast-cancer detection using deep learning that can function effectively with minimal annotations as well as within the substantial noise inherent to breast ultrasound. Beyond boosting public-health outcomes, broader-impact activities include the participation of women and minority students in the research. Specifically, this project aims to develop a weakly-supervised breast-cancer detection using active and weakly-supervised learning to help doctors diagnose breast cancer in ultrasound images. Weakly-supervised object localization will be applied to avoid reliance on bounding-box-level annotations, instead employing a transformer-based classification network to detect breast cancer. Furthermore, active learning will select the most informative images for training an improved breast-cancer detection model by iteratively choosing the most relevant breast-ultrasound images based on detection outcomes to successively refine the model's capabilities. The success of this project will bring both social and technological benefits, its outcomes enhancing women's health by assisting early detection of breast cancer, particularly benefiting underrepresented groups. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.