Disentangled Anatomy-Disease Diffusion Model for UC Endoscopy
A novel framework named Disentangled Anatomy-Disease Diffusion (DADD) has been created by researchers to generate longitudinal endoscopic images of ulcerative colitis (UC) at various disease stages while maintaining individual anatomical details. The primary issue tackled is the intertwining of pathological textures with structural elements in medical imaging. DADD utilizes a latent diffusion model conditioned on two distinct embeddings: an image encoder trained on patient anatomy and an ordinal embedder focused on cumulative disease severity, adhering to the Mayo Endoscopic Score (MES) progression. To avoid contamination of anatomical data by disease information, a Feature Purifier—a mechanism based on cross-attention—detects and eliminates disease-related channels, resulting in refined anatomical tokens. These purified tokens, along with disease tokens, are then incorporated into the denoising network. The study appears on arXiv under ID 2605.01848.
Key facts
- DADD stands for Disentangled Anatomy-Disease Diffusion.
- It synthesizes longitudinal UC endoscopic images at controllable disease stages.
- The model uses a pretrained image encoder for anatomy and an ordinal embedder for disease severity.
- Disease severity follows the Mayo Endoscopic Score (MES) ordinal progression.
- A Feature Purifier erases disease-correlated channels from image embeddings.
- The framework is designed for controllable disease progression synthesis.
- The paper is on arXiv with ID 2605.01848.
- The approach preserves patient-specific anatomy while varying disease severity.
Entities
Institutions
- arXiv