BlochShift: Physics-Guided Diffusion with Bloch-Based Drift for MRI Image Translation

Multi-contrast MRI improves diagnostic workflows, but comprehensive protocols are frequently unavailable due to scan-time limits and site-dependent acquisition variability. We study paired T1 ↔ T2 MRI translation under multi-site scanner/protocol shift with an emphasis on few-step inference suitable for clinical pipelines. We propose BlochShift, a physicsguided diffusion formulation that introduces trajectory-level cross-domain control by (i) constructing a spatially varying domain-mixture field from simplified Bloch-derived signal models and coarse tissue priors, and (ii) injecting its closed-form time derivative as an explicit restoration drift in the reverse-time dynamics. This stepwise drift pulls each denoising update toward target-consistent contrast and reduces translation drift under large integration steps. Experiments on the multi-center IXI dataset show that BlochShift achieves SSIM/PSNR =0.75 / 23.13 for T 1 T 2 and 0.83 / 24.18 for T 2 T 1, outperforming GAN and diffusion-bridge baselines in both directions.