Progressive Disease Image Generation with Ordinal-Aware Diffusion Models

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Ulcerative Colitis (UC) lacks longitudinal visual data, hindering disease progression modeling.
• Current computer-aided diagnosis systems are limited by sparse intermediate disease stages and discrete scoring systems like the Mayo Endoscopic Score (MES).

Purpose of the Study:

• To develop novel ordinal embedding architectures for conditional diffusion models to generate realistic UC progression sequences from endoscopic images.
• To address the limitations of sparse intermediate disease stages and discrete scoring in UC.

Main Methods:

• Adapted Stable Diffusion v1.4 with specialized ordinal embeddings (Basic Ordinal Embedder and Additive Ordinal Embedder).
• Converted discrete MES categories into continuous progression representations using ordinal embeddings.
• Modeled cumulative pathological features for enhanced disease representation.

Main Results:

• The Additive Ordinal Embedder demonstrated superior distributional alignment and disease consistency compared to alternatives.
• Generated sequences showed smooth transitions between severity levels with maintained anatomical fidelity.
• Achieved strong performance metrics including CMMD 0.4137, recall 0.6331, Quadratic Weighted Kappa 0.8425, and UMAP Silhouette Score 0.0571.

Conclusions:

• Established a foundation for transforming static medical datasets into dynamic progression models for UC.
• Demonstrated the effectiveness of ordinal-aware embeddings in capturing disease severity relationships.
• Enabled the synthesis of underrepresented intermediate disease stages, supporting applications in medical education, diagnosis, and synthetic data generation.