A unified and domain-robust framework enables controllable multimodal brain MRI synthesis

Multimodal magnetic resonance image (MRI) synthesis can shorten examinations, reduce cost, and furnish contrasts that are difficult or impossible to acquire, but current models remain hard to control and brittle to shifts in scanner, protocol, and pathology. We present a unified, domain-robust synthesis framework that modulates spatial carriers with physically informed, domain-stable statistical guidance. A balanced soft-plus-hard conditioning scheme enables a single model to perform diverse contrast mapping, while allowing smooth user adjustment of contrast weighting. Evaluated on 991 test subjects from five public and three private datasets spanning cross-dataset, cross-protocol, and related cross-pathology scenarios, our approach matches or outperforms advanced baselines for both in-domain synthesis and inference under unseen domains without finetuning. Clinician Turing test and downstream glioma grading experiments further indicate realistic appearance, preservation of pathology-relevant information, and potential practical value under clinically heterogeneous conditions. By jointly delivering controllability, robustness to real-world variability, and physically informed, domain-stable guidance within one architecture, this work advances MRI synthesis toward deployment in heterogeneous clinical environments.