Synthesizing longitudinal ulcerative colitis progression via disentangled latent diffusion

What To Know

This paper presents a machine-learning method (DADD) that generates synthetic endoscopic images of ulcerative colitis with controllable disease severity using a latent diffusion model. The approach separates patient-specific anatomy from disease-related texture and uses an ordinal embedding tied to Mayo endoscopic scores to steer severity changes.

On a labeled UC image dataset the method produced synthetic images that closely matched real-image metrics and improved classifier performance when used as augmentation. The article is a technical imaging/AI methods paper rather than clinical research.

Findings show promise for producing realistic, severity-controlled synthetic endoscopy frames that could help image-analysis research or training data augmentation, but it does not report clinical outcomes or patient-facing interventions.

If you care about how imaging AI might support future IBD research, this is relevant; it does not change care now and is not a clinical tool.