Image-Conditioned Diffusion Preserves Latent PhysiologicalSignals and Enables Cross-Site Transfer of Synthetic MedicalImages

Image-to-image diffusion fine-tuned on Human Phenotype Project retinal fundus photographs preserves more source-specific physiological signal than text-conditioned generation, including for hemoglobin and waist circumference.

• Why it matters: Privacy limits sharing large medical datasets, and text-conditioned diffusion can omit subtle distributed biomarkers needed by downstream physiological models.
• What they did: The study fine-tuned Stable Diffusion for image-conditioned (I2I) generation on retinal fundus images, compared it to text-conditioned generation, and trained estimators on I2I-synthetic data to transfer from Human Phenotype Project to UK Biobank.
• The result: I2I-synthetic pretraining transferred across cohorts with performance comparable to real-image pretraining, especially when target-site labels were scarce, and increasing I2I noise reduced re-identification risk for retinal images but not chest radiographs.

AI summary of the abstract · may contain errors · not medical advice · read the paper.