Two threads: reducing shortcut learning & confounding, and creating physiologically grounded augmentation for segmentation. Research intern at the Stanford Center for Biomedical Informatics Research since 2022.
Methods to reduce spurious correlations in image-based diagnosis and segmentation, and evaluation setups that surface failure modes beyond average performance.
Simulation-driven perturbations that preserve anatomy / physics while expanding training support in clinically realistic ways.
Practical approaches to improve segmentation robustness under shift, including augmentation design and careful metric & visual validation.
Full list and citations on Google Scholar.