Synthesizing longitudinal ulcerative colitis progression via disentangled latent diffusion

Synthesizing medical images at controllable disease stages while preserving patient-specific anatomy is hindered by the entanglement of pathological textures and structural features. We address this challenge for ulcerative colitis endoscopy, where severity follows a continuous ordinal progression along the mayo endoscopic score (MES).

We propose disentangled anatomy-disease diffusion (DADD), a latent diffusion framework conditioned on a pretrained image encoder for anatomy and an ordinal embedder for disease severity. A feature purifier suppresses disease-correlated channels to produce cleaner anatomical representations and reduce semantic conflict between source anatomy and target pathology. These representations are injected into the denoising network through a triple-pathway cross-attention module with resolution-dependent routing. We further introduce delta steering, a training-free signal derived from ordinal embeddings that enables explicit single-pass control of disease transitions at inference.

On the labeled images for ulcerative colitis (LIMUC) dataset, the hierarchical DADD variant achieved a regression-judge quadratic weighted kappa (QWK) of 81.53 and root-mean-square error (RMSE) of 0.46, approaching the performance obtained on real images (83.28 QWK, 0.47 RMSE). It also achieved a Fréchet inception distance (FID) of 27.20 and improved downstream classification from 74.84% accuracy/83.14 QWK with real-only training to 77.75% accuracy/85.77 QWK when synthetic images were added.

DADD improves controllable synthesis by disentangling patient-specific anatomy from pathological texture, enabling anatomically consistent visualization of severity progression and healing.