Representation-guided volumetric diffusion for label-free and slice-consistent 3D brain MRI synthesis

Abstract

Existing methods for 3D brain MRI synthesis face a trade-off between efficiency, anatomical guidance, and volumetric coherence. Slice-wise or largely 2D diffusion pipelines are computationally efficient but can suffer from inter-slice inconsistency, whereas fully volumetric models better capture 3D structure at higher computational cost. Hybrid slice-aware designs alleviate discontinuities through sequential consistency mechanisms, but coherent 3D anatomical modelling during denoising remains only indirectly addressed. In addition, methods that improve anatomical realism often rely on explicit structural guidance, such as segmentation masks, limiting applicability when annotations are unavailable. To address these limitations, we propose a representation-guided volumetric diffusion framework for label-free and slice-consistent 3D brain MRI synthesis. Instead of using structural labels, we derive an implicit anatomical prior from a single reference slice with frozen visual encoders and inject it into denoising on volumetric latents. We further introduce depth-aware contextual modelling during reverse diffusion to strengthen inter-slice coherence. Experiments on CamCAN show improved synthesis quality over representative baselines, and transfer results on OASIS-3 demonstrate efficient cross-cohort adaptation by updating lightweight conditioning modules.