Latent bridge matching synthesizes breast MRI contrast, cuts gadolinium need

DCE-MRI is central to breast cancer imaging, but it relies on gadolinium-based contrast agents that increase scan burden and carry safety concerns. A team led by Sina Amirrajab from the arXiv preprint 2608.10000 proposes a latent bridge matching (LBM) framework to synthesize peak-enhanced breast DCE-MRI directly from pre-contrast images. Unlike conventional latent diffusion models (LDMs) that start from Gaussian noise, LBM learns a conditional bridge between paired pre-contrast and peak-enhanced VAE latents. A latent UNet predicts the remaining correction from intermediate bridge states, keeping the generation anchored to patient-specific anatomy and enabling iterative refinement.

On 91 DUKE validation cases, they tested two conditioning variants: pre-contrast conditioning and tumor-conditioning using tumor masks. The tumor-conditioned variant reduced MSE from 1.023 to 0.940 and FRD from 7.523 to 4.716, while boosting tumor SSIM from 0.355 to 0.429—a 21% improvement. It also beat the evaluated LDM baseline. However, the method still depends on ground-truth tumor masks at inference, limiting practical use. The authors call for further work on generalization and removing this dependency. Still, these results position LBM as a promising pre-contrast-anchored formulation for virtual contrast enhancement in breast MRI.