Abstract: The unprecedented advances of modern machine learning have unlocked the potential for faster and more accurate data-driven analysis. However, ideal algorithmic setups often fall short in practice, particularly in diverse healthcare environments. As a result, the successful deployment of any approach depends on both the model and the data: theoretical foundations ensure methodological rigor and enhance the model's interpretability and scalability; meanwhile, data-driven models must rely on large, inclusive datasets to achieve robust and generalizable representations.

In this talk, I will present two key areas of my research that address the challenges posed by real-world data and enhance model interpretability and robustness. First, I will introduce a series of physics-driven learning approaches I developed to model spatiotemporal dynamics in brain perfusion. These approaches enable continuous reconstruction of perfusion imaging time series, providing interpretable insights for stroke diagnosis and significantly improving lesion detection. Second, I will discuss my work on modality-agnostic representation learning for medical imaging, which leverages domain randomization to create robust and generalizable foundation models that are resilient to variations in imaging modalities, resolutions, and external artifacts. These methods hold the potential to increase access to affordable, low-field MRI diagnostics. 

Looking ahead, I am excited to explore advanced physics-driven formulations for dynamic modeling in real-world scenarios, particularly, developing interactive models for predicting patient outcomes following interventional treatment. I will also continue focusing on creating robust and generalizable algorithms to improve diagnostic performance and promote accessible healthcare worldwide. By bridging theory, algorithms, and applications, my long-term goal is to enhance the resilience of machine learning, address the imperfections of real-world data, and ultimately contribute to a safer, more reliable, and accessible healthcare environment.

Bio: Peirong Liu is a postdoctoral researcher at Harvard Medical School and Massachusetts General Hospital, working with Dr. Juan Eugenio Iglesias. She received her PhD in Computer Science from UNC-Chapel Hill in 2023, where she was advised by Dr. Marc Niethammer. Peirong also completed two research internships with the Meta AI applied research team in New York City. Her research focuses on AI for healthcare, with an emphasis on developing robust and generalizable algorithms to advance reliable and accessible healthcare. Her work has been published in top-tier ML/CV conferences, including CVPR, ICCV, ECCV, and NeurIPS, as well as in leading medical image computing venues such as IEEE TMI, MICCAI, and IPMI, with multiple papers selected for oral presentations. She has been invited to speak at MIT, Harvard, Cornell, UCSD, and HKUST. Peirong was recognized as a Rising Star in EECS at MIT, and a Rising Star in Data Science at UCSD, UChicago and Stanford.