Neural fields for adaptive white matter microstructural modeling

Abstract

Diffusion magnetic resonance imaging (MRI) is sensitive to the micron-level diffusion motion of water molecules, allowing in vivo characterization of white matter architecture in the brain. Microstructural models like diffusion tensors or the Standard Model are typically fit independently on each voxel, limiting their ability to incorporate local spatial context and adapt to regional tissue complexity. Here, we introduce FiberField, an approach using neural fields that selects from multiple candidate models with different numbers of fibers to estimate fiber configurations and properties. This method incorporates spatial context through coordinate representations and model-weighted priors inspired by Bayesian model averaging to adapt to local microstructure. To compare FiberField with ground truth fiber properties, we validate on simulated diffusion MRI before applying to in vivo data. In simulated diffusion MRI, FiberField accurately estimates fiber directions and diffusivities using a multi-stick model (angular error: $2.2 \pm 2.8°$, diffusivity root mean square error (RMSE): $4.4 \times 10^{-4} \pm 5.2 \times 10^{-4}$ mm$^{2}$/s, fiber sensitivity: $0.87 \pm 0.21$). FiberField resolves fibers at 25% smaller angular separations than fod2fixel, and the diffusivity RMSE is 52% lower within fiber crossings than the Standard Model. On acquired diffusion MRI, FiberField's estimated fiber directions agree with those from MRtrix3's fod2fixel while also mapping microstructural properties such as diffusivity (angular difference: $10.5 \pm 7.1°$). Using neural fields allows for adaptive white matter modeling for selecting from a set of alternative models while maintaining accurate fiber property estimation.