3D Curve Inference for Diffusion MRI Regularization

Peter Savadjiev, Jennifer S. W. Campbell, G. Bruce Pike, Kaleem SiddiqiMICCAI 2005 · DOI

Introduction

Diffusion Magnetic Resonance Imaging (dMRI) is a vital non-invasive technique for mapping the white matter architecture of the living brain. By measuring the diffusion of water molecules, researchers can infer the orientation of axonal bundles. However, dMRI data is inherently noisy, and traditional regularization methods often fail in regions where fibers cross, branch, or merge. These methods typically assume a single fiber orientation per voxel, leading to blurred or incorrect reconstructions in complex areas. This paper presents a novel framework that treats white matter fibers as 3D space curves, utilizing geometric constraints to distinguish true biological structures from noise.

Geometric Framework: The Co-helicity Constraint

The core innovation of this approach is the transition from 2D curve inference to a 3D differential geometry model. While 2D methods often use "co-circularity" to check if vectors are tangent to the same circle, the authors argue that the osculating helix—a curve with constant curvature and torsion—is the optimal local approximation for 3D brain fibers.

Geometric representation of co-helicity: three vectors are considered compatible if they are tangent to the same 3D helix, defined by a specific curvature and torsion.

This model introduces a "triple constraint": whereas two vectors can define a circle in 2D, three vectors are required to determine if they are tangent to the same 3D helix. The algorithm implements this through relaxation labeling, an iterative process where neighboring voxels "vote" for compatible orientations. Over multiple iterations, physically plausible fiber paths are reinforced while random noise is suppressed.

Experimental Validation and Results

The researchers validated the 3D curve inference (CI) method using a combination of physical models and simulations:

  1. Biological Phantom: A physical model of two crossing rat spinal cords embedded in agar provided a known ground truth. The CI method reduced orientation error to ~10.5°, compared to ~27.1° in raw data.
  2. Synthetic Data: Simulations of noisy planar and helical curves showed the method could cut orientation error by more than half (from 24.9° to 11.2°).
  3. In Vivo Human Data: Scans of the human brain demonstrated that the method effectively recovers multiple fiber directions in complex regions like the centrum semiovale, providing cleaner maps for fiber tracking.
Validation overview: The top row shows the biological phantom (left) and synthetic helical data (middle). The bottom rows illustrate the algorithm's ability to resolve orientations in crossing regions.

Discussion and Conclusion

By modeling fibers as helices, the framework captures both curvature (bending) and torsion (twisting out of a plane), reflecting the true 3D nature of neural pathways. This unified approach works across different MRI protocols (DTI and HARDI) and naturally handles multi-fiber voxels. While computationally more intensive than simple smoothing, the significant gain in orientation accuracy suggests that 3D curve inference is a powerful tool for regularizing complex brain connectivity data. Future work will focus on integrating these geometric priors directly into full tractography pipelines.

Summary updated

BibTeX
@inproceedings{savadjiev2005d,
  author = {Peter Savadjiev and Jennifer S. W. Campbell and G. Bruce Pike and Kaleem Siddiqi},
  title = {{3D Curve Inference for Diffusion MRI Regularization}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2005},
  doi = {10.1007/11566465\_16},
}
Figure