Physics-informed Latent Diffusion Model multi-echo chemical shift-encoded liver MRI generation

Motivation: There is a lack of large liver CSE-MRI datasets, paired with ground-truth results, for the training of Deep Learning (DL)-based methods to quantify PDFF maps. Goal(s): To design a state-of-the-art generative model to synthesize realistic multi-echo liver MR images given a set of arbitrary MR scan parameters. Approach: To use a physics-driven latent diffusion model to synthesize realistic liver CSE-MR images with different compositions and geometries. Results: The synthesized data enabled the training of convolutional neural networks for PDFF estimation with biases at liver ROIs comparable to models trained on real datasets. Impact: We successfully generated realistic multi-echo liver MR images and relevant quantitative maps to train deep learning models for PDFF estimation. The combination of limited real samples and numerous synthetic images for training enabled an improved performance compared to real-only datasets.