DIMOND++: Improving diffusion model optimization using diffusion priors

Motivation: Solving microstructure model parameters from noisy diffusion MRI signals requires sufficient data and is slow. Supervised deep learning-based parameter mapping methods provide satisfying results but often cannot be generalized to different diffusion-encoding schemes and microstructure models. Goal(s): Achieve high-generalizability and high-quality diffusion model parameter estimations from sparse sampled q-space data. Approach: DIMOND++ employs a latent diffusion model to learn the diffusion model parameter distribution prior and samples parameter map from the conditional probability of parameter map given acquired data using posterior sampling. Results: DIMOND++ outperforms conventional methods and learning-based method for fitting tensor model and kurtosis model in both in-distribution and out-of-distribution test. Impact: DIMOND++ has a high potential to transform diffusion model fitting. Its superior generalization capability and the ability to be deployed directly on any dataset will greatly enhance the clinical and neuroscientific applicability of diffusion MRI based microstructure and connectivity mapping.