Summary: Retinal image registration is a fundamental step in the analysis of ophthalmic images, but remains challenging because much of the image can appear homogeneous, while the most informative anatomical structures, such as blood vessels and bifurcations, are sparse and highly localised. This talk presents a structure-aware framework for deformable retinal image registration based on Gaussian Primitive Optimization (GPO). The method models salient vascular keypoints as trainable Gaussian control primitives, allowing local anatomical correspondences to guide the estimation of a globally coherent deformation field. The talk will discuss how explicit structural modelling can make retinal registration more robust in challenging settings, and will also briefly connect this work to broader efforts in multimodal retinal image analysis, including topology-aware image fusion. More generally, the seminar will highlight the value of embedding anatomical structure directly into learning and optimisation frameworks for medical imaging.
Bio: Xin Tian is a Postdoctoral Researcher in the Knight Group at the University of Oxford, U.K. She received the B.Eng. degree in Biomedical Engineering from Tianjin University, China, the M.Sc. degree in Electrical and Electronic Engineering from the University of Bristol, U.K., and the Ph.D. degree (2024) from the Visual Information Laboratory, University of Bristol, U.K., supervised by Prof. Alin Achim, Dr Lindsay Nicholson and Dr Nantheera Anantrasirichai. Since 2024, she has been with the University of Oxford, leading a multimodal AI research stream on sepsis subtyping and biomarker discovery in collaboration with Danaher. Her research interests span multimodal biomedical image analysis — registration, fusion, and translation — with emphasis on structure-aware and graph-based learning, optimal transport, and weakly supervised models for complex multimodal data.