PURD: A Prior-Guided and Uncertainty-Aware Residual Diffusion Model for Multi-Modal MRI Synthesis

Multi-parametric magnetic resonance imaging (mpMRI) provides complementary diagnostic information; however, its clinical utility is often hindered by the issue of missing modalities resulting from prolonged scanning times, high costs, and the potential risks associated with contrast agents. While diffusion models (DMs) have demonstrated superior performance in synthesizing missing modalities, the stochastic nature of the reverse sampling process frequently compromises anatomical consistency. Furthermore, the lack of effective uncertainty assessment for synthesized results limits their practical value in clinical diagnosis. To address these challenges, we propose a Prior-guided Uncertainty-aware Residual Diffusion model (PURD), designed to generate accurate and anatomically consistent multi-modal MR images. First, PURD introduces a CLIP-based prior image generator that constrains the solution space via an implicit multi-modal feature injection mechanism, thereby ensuring the integrity of critical anatomical structures during the synthesis process. To enhance the reliability of the synthesized results, the model employs an uncertainty-aware guidance strategy: by quantifying generative stochasticity to construct pixel-level uncertainty maps, we inject these maps as Spatial Conditioning Masks into the reverse sampling process, achieving adaptive optimization for the reconstruction of complex regions. Furthermore, the model adopts a residual diffusion framework-replacing traditional Gaussian diffusion-to more effectively capture subtle lesion details. Extensive experiments conducted on two public brain MRI datasets and one in-house dataset demonstrate that images generated by PURD exhibit superior realism and fidelity. Specifically, compared to the second-best performing methods, our approach achieved an average increase of 0.61dB in PSNR and 0.008 in SSIM.