194. Learning Beyond the Training Distribution: Geometry, Robust Augmentation, and Conditional Learning in Medical Imaging

Abstract:

Deep learning models often rely on dataset-specific patterns, limiting their ability to generalize to new domains and acquisition settings. This seminar presents a line of research conducted during my PhD at the Sano Centre for Computational Medicine that investigates how incorporating additional structure into the learning process can improve robustness and generalization. I will discuss three complementary perspectives: geometry-aware regularization, which constrains models using anatomical and geometric priors; robust data augmentation, which enriches information embedded in training data with prior knowledge; and conditional learning, where auxiliary shape representations are used to guide model predictions. Through examples from recent work, including geometry-based learning objectives, augmentation-driven robustness studies, and decoder conditioning with shape-derived embeddings, I will show how these approaches encourage models to learn representations that are less dependent on dataset-specific characteristics and more transferable to unseen data. Together, these studies illustrate how explicit inductive biases can improve the reliability of deep learning systems in medical imaging.