Diffusion-based generative fiber orientation restoration from severe signal loss in diffusion-weighted MRI.

Fiber Orientation Distributions (FODs) is a widely used model to represent diffusion MRI (dMRI) data. However, susceptibility-induced distortion can cause signal loss and corrupted reconstruction of FODs in multiple brain regions even after the well-established distortion correction, which can severely affect fiber tracking and connectivity analysis. Generative models, such as diffusion models, have been successfully applied in various image restoration tasks. Their application to FODs, however, poses unique challenges since FODs are four-dimensional images represented by spherical harmonics (SPHARM), with the fourth dimension exhibiting order-related dependency. In this paper, we propose a novel diffusion model for FOD restoration. We use volume-order encoding to enhance the ability of the diffusion model to generate individual FOD volumes at all SPHARM orders. Moreover, we add cross-attention features extracted across all SPHARM orders to capture the order-related dependency of FOD volumes. Building upon our previous distortion severity modeling framework, we also condition the diffusion model with FODs of surrounding areas to maintain geometric coherence. We trained and tested our model using data from the UK Biobank (n = 1315) and the Lifespan Human Connectome Project Aging (HCP-Aging) (n = 679). We demonstrate high accuracy of the generated FODs in both the brainstem and orbitofrontal lobes. The tractography results on the corticospinal tract (CST) using the generated FODs show improved performance on both UK Biobank and HCP-Aging datasets. On the HCP-Aging data, we show that CST reconstructed from generatively restored FODs using data from a single phase-encoding (PE) direction achieved better performance than tractography results based on merged data from two PEs.