Retraining-free multi-task MR image restoration using a pretrained diffusion model with Bayesian optimization

Magnetic resonance imaging (MRI) often requires long scan times, and efforts to shorten scans can lead to image degradation, such as undersampling artifacts, increased noise, and reduced spatial resolution. A restoration framework that can address multiple types of degradation without task-specific retraining would therefore be valuable. In this study, we evaluated whether a pretrained score-based diffusion model combined with Bayesian optimization, termed task guidance design with Bayesian optimization (TGD-BO), could provide retraining-free, multi-task MR image restoration for compressed sensing (CS) reconstruction, super-resolution (SR), and denoising (DN). TGD-BO incorporates data-consistency guidance and gradient guidance into the reverse diffusion process and parameterizes their time-varying strengths using compact low-dimensional B-spline-based guidance schedules. For each task, the schedule was optimized by Bayesian optimization using four paired degraded/reference slices from a single adaptation case, and the resulting task-specific schedule was then applied unchanged to an independent test set. In single-task experiments, TGD-BO improved upon the fixed-guidance diffusion baseline in denoising, compressed-sensing reconstruction, and super-resolution, although the improvement in super-resolution was modest. Compared with task-specific supervised baselines, TGD-BO did not uniformly achieve higher peak signal-to-noise ratio (PSNR) or structural similarity (SSIM); in super-resolution, it showed lower PSNR and SSIM but better perceptual similarity as assessed by learned perceptual image patch similarity (LPIPS). In joint-task experiments, schedules optimized directly for DN + CS and DN + SR yielded numerically better image-quality metrics than applying the corresponding single-task schedules. Sensitivity analyses showed that the selected schedule and test performance were unchanged when 1, 4, or 8 slices from the same adaptation case were used, whereas changing the adaptation case substantially affected test performance. These results suggest the feasibility of adapting a single pretrained diffusion prior across multiple MR image restoration tasks through guidance calibration without task-specific weight retraining, while indicating that robustness to adaptation-case selection and computational cost require further investigation.