Deep learning has transformed computer vision, but its effectiveness drops significantly with domain shifts—when training and testing data distributions differ. It is particularly prominent in multi-center medical studies, where multiple imaging modalities, varying acquisition protocols and sites can lead to substantial data discrepancies. Additionally, the need for large annotated datasets hinders clinical deployment, as obtaining expert annotations is time-consuming and expensive.

This project aims to address these challenges by: a) exploring parameter-efficient unsupervised domain adaptation techniques for multi-target medical domains b) developing novel methods for training-free adaptation of foundation models for medical organ segmentation c) exploring few-shot techniques to adapt foundation models for medical disease diagnosis.