Abstract
Registration of diffusion-weighted magnetic resonance images (DW-MRIs) is a key step for population studies and the construction of brain atlases. Given the high dimensionality of the data, registration is usually performed by relying on scalar representative images, such as fractional anisotropy (FA) or non-diffusion-weighted (b0) images, thereby ignoring much of the directional information conveyed by the DW-MR dataset itself. Alternatively, model-based registration algorithms have been proposed to exploit information on preferred fiber orientations, but these can be limited by the accuracy of the underlying model (e.g., diffusion tensors). This work introduces a framework based on angular interpolation (AI) of diffusion-weighted volumes for both linear and non-linear registration. By avoiding reliance on specific local diffusion models, AI improves registration accuracy and provides registered raw DW-MRI data suitable for any subsequent analysis.
The Challenge of DW-MRI Registration
Diffusion-weighted MRI is a powerful tool for mapping the brain's white matter, but it presents a massive data challenge. A single scan contains hundreds of volumes, each measuring water diffusion along a specific 3D direction. To compare brains across a population, researchers must "register" these images—warping them into a common coordinate space.
The standard shortcut is to register a single representative scalar image. However, this ignores the rich directional information in the full dataset. Alternatively, model-based registration aligns estimated diffusion tensors. The catch is that if the model is wrong (e.g., it cannot handle crossing fibers) or the estimation is noisy, the registration suffers. Furthermore, when you warp a brain, the underlying fibers rotate, meaning the signal measured at a fixed gradient direction is no longer valid. You must "reorient" the data, a process that is often an afterthought rather than part of the optimization itself.
Methodology: Angular Interpolation (AI)
The authors propose Angular Interpolation (AI) as a solution. Instead of registering a model, they register the raw signal directly.
- The Core Idea: When a spatial transformation rotates a voxel, AI re-estimates the signal intensity for that voxel by interpolating between the intensities of the nearest available gradient directions in the original "q-space" (the sphere of diffusion directions).
- Linear and Non-Linear Integration: While AI had been proposed for simple linear (affine) registration, this paper generalizes it to non-linear registration. They integrated AI directly into the cost function of FNIRT (FSL’s non-linear registration tool).
- Optimization: To make this computationally feasible, the team implemented the algorithm in C++ using OpenMP for parallelization and kd-trees to rapidly find the nearest gradient neighbors.
- Signal Preservation: They utilized Blackman-windowed sinc interpolation for spatial moves and a small smoothing parameter for angular moves to prevent "tensor swelling"—a common artifact where interpolation artificially lowers the perceived anisotropy of fibers.
Key Insights and Results
Because AI works on the raw signal, it is model-agnostic. It does not care if a voxel contains one fiber, two crossing fibers, or complex "kissing" fibers, avoiding the bias inherent in assuming a Gaussian (DTI) distribution. By including AI inside the registration loop, the algorithm "sees" the intensity changes caused by fiber rotation during the optimization process, leading to a more accurate fit.
In synthetic tests, the AI-based linear registration achieved significantly lower Mean Fiber Orientation Error (FOE) compared to standard b0-based registration. In non-linear inter-subject registration, the AI framework was highly competitive with specialized tensor-based tools like DTI-TK, while significantly outperforming standard scalar-based FNIRT. Despite the high dimensionality, the optimized implementation allows non-linear registration to complete in 30–45 minutes on a standard workstation.
Conclusion
This framework provides a robust alternative to volume and tensor-based approaches. By outputting registered raw DW-MRI data, it ensures reproducibility and allows researchers to apply any future high-order model (like ODFs or multi-shell models) to the already-aligned raw data. Future work may extend this concept to multi-shell diffusion imaging (HARDI), where the radial component of q-space also varies.
Summary updated