[4O1-IS-2a] Machine learning

10:00 AM - 10:15 AM JST(1:00 AM - 1:15 AM UTC)

[4O1-IS-2a-05] Semantic Latent Space Construction for Ophthalmic ImagingToward Disease-Consistent Representation Learning Across Heterogeneous Modalities

〇Rongkai Sun¹, Kazumasa KISHIMOTO^1,2, Osamu SUGIYAMA³, Hiroshi TAMURA^1,4 (1. Graduate School of Informatics, Kyoto University, 2. Kyoto University Hospital, 3. Department of Information Science, KINDAI University, 4. Institute of Liberal Arts and Sciences, Kyoto University)

work-in-progress,[[online]]

Keywords:

Multimodal Consistency,Disease-Consistent Representation,Semantic Latent Space,Oculomics,Representation Learning

Oculomics explores ophthalmic images as non-invasive biomarkers for ocular and systemic diseases and has attracted growing attention in recent years. In this work, we review and organize representative oculomics studies to summarize common modeling paradigms and the types of semantic information captured by different imaging modalities. Motivated by this overview, we investigate disease-consistent semantic latent space learning for mapping multiple modalities into a shared, medically meaningful representation. Color fundus photography, fundus fluorescein angiography, and optical coherence tomography capture complementary yet heterogeneous manifestations of the same disease state. As an initial feasibility study, we examine semantic consistency between CFP and FFA using a conditional latent diffusion framework. Preliminary results indicate that vascular-related information can be inferred from non-invasive CFP.