Introduction
Diffusion Magnetic Resonance Imaging (dMRI) is a critical tool for non-invasively mapping the brain's white matter by tracking the diffusion of water molecules. Traditionally, this diffusion is modeled using rank-2 tensors, visualized as 3D ellipsoids. However, these simple models fail in regions with complex microstructures, such as crossing or branching nerve fibers.
To capture these complexities, researchers utilize Higher Order Tensors (HOTs)—specifically rank-4 and rank-6 tensors. While HOTs provide the necessary mathematical flexibility, clinical MRI scanners often produce data at a spatial resolution (1–2 $mm^3$) that is too coarse to resolve fine tissue structures, leading to blocky models. Enhancing this resolution requires interpolation, but standard methods for rank-2 tensors do not generalize to HOTs. Direct linear interpolation, the common fallback, often violates physical constraints like positive definiteness, resulting in anatomically impossible models.
Methodology: The Tucker Decomposition Process (TDP)
This paper introduces the Tucker Decomposition Process (TDP), a non-parametric Bayesian framework designed to interpolate HOT fields of any even rank while strictly maintaining physical validity.
The TDP combines two powerful mathematical concepts:
- Tucker Decomposition: A multi-linear algebra technique that decomposes a tensor into a "core" tensor and a set of factor matrices.
- Gaussian Processes (GPs): The elements of the factor matrices are modeled as independent Gaussian Processes indexed by spatial coordinates. This allows the model to learn global spatial correlations across the brain tissue.
By structuring the synthesis of the tensor field through this decomposition ($\mathcal{T} = \mathcal{D} \times_1 \mathbf{A} \dots$), the resulting interpolated tensors are mathematically guaranteed to be symmetric and positive definite. Inference is performed using Markov chain Monte Carlo (MCMC) methods, specifically Metropolis-Hastings and elliptical slice sampling, to estimate the posterior distribution of the tensor field at unmeasured locations.
Experimental Validation
The authors evaluated TDP against synthetic and real dMRI data (3.0T scanner, 90 gradient directions). High-resolution fields served as the ground truth and were downsampled by a factor of two to create training sets for upsampling.
Performance was quantified using the tensorial Frobenius distance. In rank-2 scenarios, TDP (error: 0.287) significantly outperformed linear interpolation (0.756) and Log-Euclidean methods (0.432). For HOTs, where advanced competitors are scarce, TDP consistently outperformed linear interpolation:
- Rank-4 Real Data: TDP error of 2.030 vs. 2.727 (Linear).
- Rank-6 Real Data: TDP error of 1.947 vs. 2.664 (Linear).
Discussion and Key Insights
The primary advantage of TDP is its true generalization. Unlike previous probabilistic models like Generalized Wishart Processes, which are restricted to rank-2 tensors, TDP’s formulation naturally extends to any even-ranked tensor. Furthermore, the use of GPs ensures a high degree of spatial coherence, producing anatomically plausible fiber reconstructions that avoid the jagged artifacts typical of local interpolation methods.
However, the model's reliance on stationary GP kernels presents a limitation: it tends to over-smooth regions with abrupt structural changes, such as sharp tissue boundaries.
Conclusion
The Tucker Decomposition Process provides a mathematically rigorous and physically constrained framework for the super-resolution of complex dMRI data. By successfully interpolating higher-order tensors, TDP enables more accurate modeling of the brain's intricate connectivity. Future work aims to incorporate non-stationary covariance functions to better handle sharp anatomical transitions, further refining the model's precision in heterogeneous tissue environments.
Summary updated