Deep learning for realistic synthesis of 3D contrast-enhanced MRI from non-contrast MRI volumes

Gadolinium-based contrast agents (GBCAs) markedly improve lesion conspicuity in MRI, but add cost, workflow burden, and small yet non-negligible safety risks. This motivates virtual contrast enhancement: synthesizing contrast-enhanced (CE) images from non-CE acquisitions so radiologists can detect enhancement without injection. We study this problem for brain MRI and adapt Posterior-Mean Rectified Flows (PMRF) to fully volumetric (3D) data. We explicitly target the perception-distortion trade-off: the tension between making images look more realistic to humans (perception) and keeping them numerically faithful to the ground-truth signal (distortion).
Our approach decouples fidelity from realism in two stages. First, a residual 3D U-Net predicts the conditional posterior mean of CE T1-weighted MRI from non-contrast inputs, yielding a low-distortion anatomical template. Second, a time-conditioned 3D rectified flow nudges this template toward the empirical CE distribution, with the number of Euler integration steps K in the second stage serving as a single, interpretable knob that trades perceptual realism for distortion.
We evaluate on multi-institutional cohorts against several posterior-mean and rectified-flow baselines. Key findings: (i) With a single integration step, PMRF achieves the lowest distortion among all tested methods while matching the posterior-mean baseline on perceptual quality. (ii) With larger step budgets, PMRF attains state-of-the-art perceptual realism comparable to the strongest flow baseline yet retains markedly lower distortion. (iii) Across step budgets, PMRF consistently lies on the empirical perception-distortion Pareto frontier. Qualitatively, PMRF sharpens enhancing rims, preserves non-enhancing cores, adds realistic micro-texture in normal brain, and avoids spurious hallucinations. These findings suggest that virtual CE with PMRF could reduce reliance on GBCAs in select scenarios.