Balancing Fidelity, Utility, and Privacy in Synthetic Cardiac MRI Generation: A Comparative Study

TL;DR

This study benchmarks three generative models for synthetic cardiac MRI, finding diffusion models like DDPM best balance image quality, utility for segmentation, and privacy preservation, aiding safe data augmentation.

Contribution

It systematically compares DDPM, LDM, and FM architectures for synthetic CMR generation, highlighting diffusion models' superior balance of fidelity, utility, and privacy.

Findings

DDPM achieves the best balance between fidelity, utility, and privacy.

Diffusion models outperform others in limited-data scenarios.

FM offers promising privacy with slightly reduced task performance.

Abstract

Deep learning in cardiac MRI (CMR) is fundamentally constrained by both data scarcity and privacy regulations. This study systematically benchmarks three generative architectures: Denoising Diffusion Probabilistic Models (DDPM), Latent Diffusion Models (LDM), and Flow Matching (FM) for synthetic CMR generation. Utilizing a two-stage pipeline where anatomical masks condition image synthesis, we evaluate generated data across three critical axes: fidelity, utility, and privacy. Our results show that diffusion-based models, particularly DDPM, provide the most effective balance between downstream segmentation utility, image fidelity, and privacy preservation under limited-data conditions, while FM demonstrates promising privacy characteristics with slightly lower task-level performance. These findings quantify the trade-offs between cross-domain generalization and patient confidentiality, establishing a framework for safe and effective synthetic data augmentation in medical imaging.