Webinar #27: Infinite from finite: Medical Vision Compositionality & Foundation Models-Dr. Yuting He

Title: Infinite from finite: Medical Vision Compositionality and Foundation Models Abstract: The increasing specialization of medicine has created a rapidly expanding landscape of imaging modalities, diseases, and clinical tasks. However, the development of medical AI remains constrained by limited annotations, long-tailed distributions, and heterogeneous imaging protocols. These limitations make the conventional one-model-per-task paradigm challenging to scale, particularly for rare diseases, emerging modalities, and specialized clinical scenarios. This talk will introduce compositionality as a principle for building generalizable medical vision systems from finite resources. Medical images are generated through combinations of different components, like topological and geometric relationships, pathological variations, and modality-specific appearances. By identifying these reusable primitives and learning how they can be recombined, medical AI models will achieve systematic generalization beyond the tasks and distributions observed during training. This talk will discuss this perspective through three interconnected research directions. 1) At the data level, it will present generative data-engine models that deconstruct images into topology and imaging appearance, and then recombine these components to generate diverse training distributions for few-shot medical image segmentation. 2) At the representation level, it will describe topology- and geometry-informed self-supervised learning methods that use the structural consistency of human anatomy as an inductive bias, enabling more reliable correspondence discovery and dense semantic representation learning. 3) At the model level, it will introduce self-composing neural architectures that learn emergent modular representations, allowing specialized and shared knowledge to be dynamically coordinated across heterogeneous medical modalities and clinical tasks. These studies demonstrate how finite data, finite structural rules, and finite computational components can be composed into scalable learning systems. The broader goal is to move medical AI beyond isolated task-specific models toward resource-efficient, adaptable, and generalizable foundation models for biomedical imaging. Bio: Yuting He is currently a Research Associate in the Department of Biomedical Engineering at Case Western Reserve University, USA. He received his Ph.D. in Computer Science from Southeast University in 2023. His research lies at the forefront of medical representation learning, interactive medical AI, and foundation models for healthcare, with a focus on developing structure-aware and human-centered computational frameworks for multimodal medical data. Dr. He has published over 40 peer-reviewed articles in top-tier journals and conferences, including Nature Communications, IEEE T-PAMI, IEEE T-MI, Medical Image Analysis, CVPR, ICCV, MICCAI, etc. His scholarly contributions are widely recognized, evidenced by prestigious honors such as the Bao Gang Outstanding Student Award. In addition to his research, Dr. He is actively involved in the academic community, serving as an Area Chair for MICCAI 2026, Senior Program Committee member for IJCAI-ECAI 2026, and a regular reviewer for leading venues including IEEE T-MI, IEE T-NNLS, NeurIPS, CVPR, AAAI, etc.