Metadata-Aware Adaptation of a Generative Foundation Model for Conditional CMR Synthesis

Generating synthetic medical images is a vital strategy for addressing data scarcity and class imbalance in clinical datasets. However, ensuring that these synthetic images accurately reflect specific patient characteristics—such as pathology, age, or BMI—remains a significant challenge. This study investigates whether a latent diffusion model can be effectively conditioned on structured patient metadata to synthesize clinically meaningful cardiac magnetic resonance (CMR) images without requiring complex geometric inputs.

The researchers fine-tuned a pretrained Stable Diffusion model using 59,058 short-axis CMR images from the UK Biobank. To improve metadata adherence and mitigate the effects of class imbalance, they introduced three specific strategies:

• Metadata-Free Classifier-Free Guidance (MF-CFG): A technique that strengthens the conditioning signal during inference by contrasting metadata-aware prompts against metadata-free versions.
• Contrastive Batching (CB): A sampling method that increases metadata diversity within training batches to ensure the model is exposed to a wider range of patient attributes.
• Inverse-Frequency Sampling (IFS): A weighting strategy that prioritizes underrepresented metadata values during training to counteract class imbalance.

The combined framework achieved a Fréchet Inception Distance (FID) of 37.47, representing a 57.04% improvement over a standard fine-tuned baseline and a 28.68% improvement over a previous CMR-specific diffusion model that required additional cardiac geometry inputs. The authors found that MF-CFG was the primary driver of distributional fidelity. While the model successfully generated realistic images that aligned with demographic metadata, it struggled most with disease-specific conditioning, highlighting that while these techniques improve synthesis, they do not fully resolve the challenges posed by highly imbalanced clinical data.