From Physics to Clinical Concepts: Pathways to Generalisable and Trustworthy AI in Medical Imaging

Abstract: 
Recent breakthroughs in artificial intelligence have dramatically expanded its role in medical imaging, unlocking new possibilities in diagnosis, treatment planning, and prognostic assessment. Yet the biomedical domain brings unique challenges: annotated datasets remain scarce and costly, and clinical adoption hinges on transparent models that physicians can trust and interrogate. By embedding physical principles, such as those governing tracer kinetics and photon attenuation, into the learning process, models can respect real-world constraints and avoid generating non-physical information. At the same time, representing decisions through clinically meaningful concepts, rather than opaque features, ensures interpretability and diagnostic relevance. In this talk, I will share recent efforts to overcome these hurdles, learning effectively from limited data and designing AI architectures that are interpretable by design. Through illustrative case studies in Positron Emission Tomography (PET), I will highlight both the strides we have made and the open questions that remain on the path toward truly reliable, trustworthy AI for healthcare.

Speaker bio:
Dr. Ruwan Tennakoon is a Senior Lecturer in Artificial Intelligence at RMIT University, Australia, and an Honorary Senior Lecturer at the University of Melbourne. With over a decade of experience developing AI algorithms for healthcare, his research focuses on enhancing PET and CT imaging for improved diagnosis and patient outcomes. His PhD addressed robust image registration for low-dose lung CT, followed by industry experience at IBM Research, where he contributed to three U.S. patents in ophthalmology AI. Dr. Tennakoon has led projects exceeding $4 million in funding and published over 50 papers in top venues, including CVPR, IEEE PAMI, TIP, and TMI.