M.S. Candidate: Meryem Mine Kurt
Program: Multimedia Informatics
Date: 29.08.2025 / 13:30
Place: B-223

Abstract: Disease progression modeling presents significant challenges in medical imaging due to the scarcity of longitudinal data and the inherent class imbalance in medical datasets. This thesis introduces a novel conditional diffusion framework for synthesizing realistic disease progression sequences from cross-sectional data, with a focus on ulcerative colitis endoscopic imaging. The proposed approach employs specialized ordinal class embeddings that capture the progressive nature of disease severity, enabling the generation of smooth transitions between discrete Mayo Endoscopic Score levels. Two embedding strategies are developed: a Basic Ordinal Embedder using linear interpolation between severity classes, and an Additive Ordinal Embedder that explicitly models the cumulative nature of pathological features. The framework is built upon Stable Diffusion v1.4 with custom modifications for medical imaging applications, incorporating advanced training techniques including Exponential Moving Average, Min-SNR-γ weighting, and class-balanced sampling. The methodology is evaluated using the LIMUC dataset through comprehensive quantitative metrics including. The framework aims to transform classification-based datasets into continuous progression models, enabling fine-grained disease severity control and realistic intermediate stage synthesis. This work addresses the critical limitation of longitudinal data scarcity in medical research and provides a foundation for enhanced clinical training, treatment planning, and disease understanding across various progressive medical conditions.