Deep generative priors for 3D brain analysis

We introduce diffusion models as general-purpose priors for solving a wide range of medical imaging inverse problems in brain MRI analysis. A score-based diffusion prior is trained extensively on diverse brain MRI data and paired with flexible forward models to tackle tasks including super-resolution, bias field correction, and inpainting. The framework also refines outputs from existing deep learning methods. Testing on clinical and research MRI datasets demonstrated state-of-the-art performance producing consistent, high-quality solutions without requiring paired training datasets, highlighting the potential of diffusion models as versatile tools for 3D brain MRI analysis.