Gaussian process regression can turn non-uniform and undersampled diffusion MRI data into diffusion spectrum imaging

Jens Sjölund, Anders Eklund, Evren Özarslan, Hans Knutsson

Abstract

We propose to use Gaussian process regression to accurately estimate the diffusion MRI signal at arbitrary locations in q-space. By estimating the signal on a grid, we can do synthetic diffusion spectrum imaging: reconstructing the ensemble averaged propagator (EAP) by an inverse Fourier transform. We also propose an alternative reconstruction method guaranteeing a nonnegative EAP that integrates to unity. The reconstruction is validated on data simulated from two Gaussians at various crossing angles. Moreover, we demonstrate on non-uniformly sampled in vivo data that the method is far superior to linear interpolation, and allows a drastic undersampling of the data with only a minor loss of accuracy. We envision the method as a potential replacement for standard diffusion spectrum imaging, in particular when acquistion time is limited.

Introduction

Diffusion MRI (dMRI) is a powerful tool for non-invasively probing the microscopic structure of biological tissues, such as brain white matter, by measuring the displacement of water molecules. The gold standard for characterizing this displacement is the Ensemble Averaged Propagator (EAP), typically calculated via Diffusion Spectrum Imaging (DSI). However, DSI requires measuring the MRI signal on a dense, regular Cartesian grid in "q-space," leading to prohibitively long scan times. While researchers often use multi-shell sampling to save time, translating that data back into a high-resolution EAP usually relies on simple linear interpolation, which fails to capture the underlying physics and is prone to significant errors.

Methods

This paper introduces a Gaussian Process Regression (GPR) framework to accurately estimate diffusion signals from sparse, non-uniformly sampled data. The core of the approach is a custom-designed covariance function (kernel) that factorizes into two parts:

  • Angular Part: Uses Legendre polynomials to ensure the signal is symmetric ($E(q) = E(-q)$) and isotropic in its prior.
  • Radial Part: Uses a radial basis function on a logarithmic scale to capture the decaying nature of the diffusion signal across different shells.

To ensure physical validity, the authors formulated the EAP reconstruction as a constrained weighted least-squares problem, ensuring the resulting EAP is non-negative and integrates to unity. The model is further improved by data augmentation, incorporating synthetic points at the origin (signal = 1) and at large radii where the signal has completely decayed.

Results

The method was validated using simulated crossing-fiber data and in vivo data from the Human Connectome Project (HCP).

Comparison of exact and estimated EAPs for crossing fibers at 30°, 60°, and 90° angles.

In simulated crossing-fiber scenarios, the GPR method significantly outperformed linear interpolation in estimating the EAP. As shown in Figure 1, the estimated contours closely match the exact shapes across various crossing angles. Quantitatively, the method reduced the relative error in estimating key scalar indices, such as the return-to-origin probability, by approximately 40-50% compared to standard interpolation.

Normalized mean absolute difference as a function of the fraction of included gradient orientations for various b-values.

A critical finding is the method's robustness to undersampling. Figure 2 demonstrates that the GPR method maintains low error even when a large fraction of measurements are randomly removed. Notably, using only 20% of the available gradient orientations with GPR yielded performance comparable to using 95% of the data with standard methods. This suggests a massive potential for reducing clinical MRI scan times.

Discussion

By integrating physical symmetries and signal behavior directly into the GPR kernel, the proposed method turns undersampled dMRI data into high-quality DSI maps. The GPR approach naturally handles the convex nature of signal decay, avoiding the overestimation errors typical of linear interpolation. While GPR is computationally more intensive, its ability to provide uncertainty estimates and handle non-uniform sampling makes it a superior alternative for clinical applications. Future work may explore encoding spatial correlations between neighboring voxels and optimizing sampling patterns to further enhance reconstruction accuracy.

Summary updated

Abstract

We propose to use Gaussian process regression to accurately estimate the diffusion MRI signal at arbitrary locations in q-space. By estimating the signal on a grid, we can do synthetic diffusion spectrum imaging: reconstructing the ensemble averaged propagator (EAP) by an inverse Fourier transform. We also propose an alternative reconstruction method guaranteeing a nonnegative EAP that integrates to unity. The reconstruction is validated on data simulated from two Gaussians at various crossing angles. Moreover, we demonstrate on non-uniformly sampled in vivo data that the method is far superior to linear interpolation, and allows a drastic undersampling of the data with only a minor loss of accuracy. We envision the method as a potential replacement for standard diffusion spectrum imaging, in particular when acquistion time is limited.
BibTeX
@inproceedings{sjolund2016gaussian,
  author = {Jens Sjölund and Anders Eklund and Evren Özarslan and Hans Knutsson},
  title = {{Gaussian process regression can turn non-uniform and undersampled diffusion MRI data into diffusion spectrum imaging}},
  booktitle = {ISBI},
  year = {2016},
  doi = {10.1109/isbi.2017.7950634},
}
Figure