Nonlinear Registration of Diffusion MR Images Based on Fiber Bundles

Ulas Ziyan, Mert R. Sabuncu, Lauren O'Donnell, Carl-Fredrik WestinMICCAI 2007 · DOI

Introduction

Registration is a fundamental task in medical imaging, enabling the comparison of brain scans across different subjects by aligning them into a single coordinate system. In Diffusion Tensor Imaging (DTI), this process is particularly complex because each voxel contains a "tensor"—a 3D representation of water diffusion—rather than a simple scalar value. Traditional registration methods often treat DTI data as simple scalar maps or attempt to match voxels independently, frequently ignoring the long-range fiber tracts that define the brain's structural connectivity. Furthermore, nonlinear warping requires rotating these tensors to maintain their alignment with the transformed tissue geometry, a process that is often mathematically difficult and prone to noise.

Methodology

Researchers have developed a novel registration framework that utilizes major white matter fiber bundles to drive a nonlinear transformation. The proposed approach employs a three-stage pipeline:

  1. Bundle Extraction & Probability Mapping: Major fiber bundles (such as the corpus callosum) are automatically labeled in each subject and converted into 3D spatial probability maps.
  2. Local Affine Alignment: For every corresponding bundle pair, the system calculates an optimal 9-parameter affine transformation (rotation, scale, and shear) to align these specific structures.
  3. Polyaffine Fusion: Using a log-Euclidean polyaffine framework, these local transforms are fused into a single, smooth, and invertible global warp field.

A significant technical highlight of this work is the derivation of an analytical solution for the Jacobian matrix (the local gradient of the warp). By avoiding the noisy approximations of standard finite-difference methods, this analytical approach allows for much more precise reorientation of diffusion tensors after the brain is warped.

Experimental Results

The study utilized 15 full-brain DTI scans, benchmarking the proposed bundle-based method against standard global affine registration and the popular "demons" algorithm, a high-dimensional nonlinear registration method.

Visual comparison of registration results and Jacobian maps for Affine (Aff), Demons (Dem), and the proposed Polyaffine (PA) methods.

As shown in Figure 1, the proposed polyaffine (PA) method produces "Jacobian maps" (bottom row) that reflect actual anatomical structures, indicating that the deformation field respects the brain's underlying geometry. In contrast, the "demons" algorithm produces maps that appear more like random noise. Visual inspection of the corpus callosum (indicated by yellow arrows) confirms that the bundle-based approach achieves superior alignment of fiber tracts compared to benchmark methods.

Quantitative performance comparison across different brain structures using the Fiber-Tensor Fit (FiT) metric.

Quantitative validation was performed using a new measure called Fiber-Tensor Fit (FiT), which quantifies how well a warped fiber tract aligns with the actual diffusion tensors in the target brain. As illustrated in Figure 2, the proposed method performed significantly better than standard affine registration and achieved results comparable to the more computationally intensive "demons" algorithm across various anatomical regions, including the corpus callosum (CC), cingulum (Ci), and fornix (Fr).

Discussion and Conclusion

By anchoring registration in the brain's structural "superhighways," this method ensures that the resulting warps are anatomically meaningful and connectivity-preserving. The use of the polyaffine framework also guarantees that the resulting warp is invertible, a crucial feature for medical research that requires moving data reliably between subject and atlas spaces.

While the current iteration uses a single affine component per structure, future work may involve increasing the degrees of freedom for complex bundles or integrating the FiT metric directly into the optimization process. Overall, this bundle-based approach provides a mathematically sound and anatomically consistent alternative to traditional voxel-wise nonlinear registration methods.

Summary updated

BibTeX
@inproceedings{ziyan2007nonlinear,
  author = {Ulas Ziyan and Mert R. Sabuncu and Lauren O'Donnell and Carl-Fredrik Westin},
  title = {{Nonlinear Registration of Diffusion MR Images Based on Fiber Bundles}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2007},
  doi = {10.1007/978-3-540-75757-3\_43},
}
Figure