Data Science and Analytics Thrust Seminar | Towards Reliable AI in Medical Imaging: Cross-Domain Generalization and Multi-Modal Robustness

Medical imaging has been an indispensable pillar of modern healthcare for accurate disease diagnosis and improved treatment. Cutting-edge AI techniques, such as deep learning, are opening new doors to intelligent and precise medical image analysis for next-generation healthcare. While deep neural networks (DNNs) have achieved remarkable success in various medical image computing tasks, the generalization and robustness of DNNs are often challenged by the data heterogeneity and multi-modal nature of medical images in complex real-world scenarios. Such limitations hinder the wide application of deep models in clinical practice. In this talk, I will present my interdisciplinary research on exploring reliable AI models in medical imaging from three perspectives, i.e., unsupervised domain adaptation, privacy-preserving domain generalization, and robust multi-modal learning. To tackle these challenging topics, a variety of deep learning techniques will be covered, including meta learning, federated learning, feature disentanglement, adversarial learning, knowledge distillation, etc. The up-to-date progress and promising future directions of building reliable AI models will also be discussed.