Abstract
This paper introduces a novel framework for registering diffusion-weighted images (DWI) that simultaneously optimizes for structural alignment and fiber reorientation. By integrating these components into a single "geodesic shooting" cost function within a Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework, the method achieves significantly higher accuracy than traditional sequential approaches. Experimental results on in vivo human brain data demonstrate that explicit orientation optimization leads to superior alignment of complex white matter structures and lower registration errors.
Introduction
Diffusion-Weighted Imaging (DWI) is a critical MRI technique for mapping the brain's white matter architecture. Comparing DWI data across subjects requires image registration—warping individual brains into a common coordinate system. A unique challenge in DWI is that local diffusion directions must be mathematically reoriented during warping to maintain physical consistency with the new anatomy. Most existing methods perform this in two decoupled steps: finding an anatomical warp and then reorienting fibers. This paper proposes that fiber orientations contain valuable information that should actively guide the warping process, leading to more accurate and biologically plausible registrations.
Methods
The authors propose an LDDMM framework using geodesic shooting. Key technical innovations include:
- Direct Q-Space Registration: The method operates on raw diffusion signals rather than simplified models like Diffusion Tensors, allowing for model-agnostic analysis post-registration.
- Fiber Basis Functions (FBFs): Complex signals are represented as combinations of Watson distributions, enabling the modeling of crossing fibers.
- Explicit Orientation Optimization: The registration cost function is reformulated to explicitly account for how deformations rotate fibers, allowing the algorithm to use orientation errors to refine the anatomical warp.
- Multi-Kernel Strategy: Multiple Gaussian kernels are used to concurrently register structures at various spatial scales.
Results and Evaluation
The framework was evaluated using in vivo brain data from 11 healthy adults. One subject was designated as the target, and the remaining 10 were registered to it.
The proposed method achieved a mean Root Mean Square (RMS) error of 9.8, significantly lower than scalar registration (11.5) and the authors' previous iterative method (10.6) ($p < 10^{-3}$). As shown in the error maps below, the proposed method produces the lowest error across the brain volume compared to standard affine, scalar, and vector registration techniques.
Qualitative alignment with the target image is also significantly improved, as demonstrated by the sharpness of the registered brain slices and the alignment of fine-grained pathways.
Discussion
By treating fiber reorientation as a fundamental driver of the registration process, the proposed framework achieves superior alignment of complex anatomical pathways. The use of geodesic shooting ensures that transformations are diffeomorphic—smooth, topology-preserving, and invertible. While the method is computationally demanding (requiring ~4.2 hours per registration), its high accuracy and model-agnostic nature make it a powerful tool for high-end neuroimaging research. Future work will focus on improving computational efficiency and validating the approach on larger clinical datasets.
Summary updated