Non-rigid Registration of High Angular Resolution Diffusion Images Represented by Gaussian Mixture Fields

Guang Cheng, Baba C. Vemuri, Paul R. Carney, Thomas H. MareciMICCAI 2009 · DOI

Introduction

In medical imaging, registration is the critical process of aligning multiple images into a common coordinate system. For brain imaging, Diffusion-Weighted MRI (DWMRI) is frequently used to map white matter pathways. The traditional standard, Diffusion Tensor Imaging (DTI), models water diffusion as a simple ellipsoid. However, DTI fails in regions of "fiber crossings" where multiple neural pathways intersect, as a single tensor cannot represent multiple directions. High Angular Resolution Diffusion Imaging (HARDI) was developed to overcome this, but registering such complex data remains a significant challenge due to the mathematical complexity of the objects and the need for accurate re-orientation during image warping.

Methodology

This paper introduces a novel framework for the non-rigid registration of HARDI data by representing it as Gaussian Mixture Fields (GMF). The proposed pipeline consists of four key components:

  1. Gaussian Mixture Fields (GMF): Diffusion at each voxel is represented as a mixture of multiple Gaussians, naturally accounting for multiple fiber populations.
  2. Thin-Plate Splines (TPS): Non-rigid deformations are modeled using TPS to ensure flexible and smooth spatial warping.
  3. L2 Distance Objective: The authors derived a closed-form expression for the derivatives of the $L_2$ distance between Gaussian mixtures. This allows for efficient optimization using the L-BFGS algorithm.
  4. Simplified Re-orientation: The "Preservation of Principle Directions" (PPD) strategy was extended to GMFs, providing a mathematically simpler and faster re-orientation method than previous DTI-based approaches.

By assuming cylindrical symmetry for individual fibers, the authors simplified complex matrix algebra into manageable scalar equations, making the gradient calculation computationally feasible.

Experimental Evaluation

The framework was evaluated using both synthetic and real-world datasets:

  • Synthetic Data: 2D images of crossing fiber bundles were subjected to random deformations and Rician noise at four levels (SNR 50, 20, 10, and 5).
  • Real Data: Rat brain DWMRI scans were used, including a longitudinal study of a single rat post-injury and inter-subject registration between two different rats.
Figure 1: Performance comparison on synthetic fiber crossings across varying Signal-to-Noise Ratios (SNR). The GMF-based method (blue) consistently shows lower error than DTI (red) and GA (green) methods, especially in crossing regions.

Key Results

Superiority in Complex Regions

In synthetic tests, the GMF-based registration significantly outperformed DTI and Generalized Anisotropy (GA) methods. This advantage was most pronounced in regions with fiber crossings, where DTI's single-tensor model is inherently limited. The algorithm demonstrated high resilience to noise, maintaining accuracy down to SNR 10.

Biological Validity and Alignment

In experiments involving rat brain scans, the registration successfully aligned disrupted fiber tracts in the corpus callosum. Checkerboard visualizations confirmed seamless anatomical alignment between source and target images. Quantitatively, the dissimilarity between consecutive scans in the longitudinal study decreased significantly following registration.

Figure 3: Registration results on rat brain DWMRI. Panels (a-f) show anatomical alignment through checkerboard views and fiber orientation maps. Panels (g-h) quantify the reduction in dissimilarity before and after registration.

Conclusion

The GMF-based registration framework provides a robust solution for aligning complex brain tissues that DTI cannot handle. While more computationally demanding than tensor-based methods, the use of closed-form gradients and sparsity-aware computation makes it feasible for high-resolution research. Future work may extend this framework to higher-order representations or the creation of high-fidelity brain atlases that incorporate crossing fiber information.

Summary updated

BibTeX
@inproceedings{cheng2009non,
  author = {Guang Cheng and Baba C. Vemuri and Paul R. Carney and Thomas H. Mareci},
  title = {{Non-rigid Registration of High Angular Resolution Diffusion Images Represented by Gaussian Mixture Fields}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2009},
  doi = {10.1007/978-3-642-04268-3\_24},
}
Figure