Codomain scale space and regularization for high angular resolution diffusion imaging

Luc FlorackWorkshops @ CVPR 2008 · DOI

Introduction

In the field of medical imaging, High Angular Resolution Diffusion Imaging (HARDI) has emerged as a critical tool for mapping the complex architecture of fibrous tissues, such as the white matter in the human brain. Unlike its predecessor, Diffusion Tensor Imaging (DTI), HARDI can resolve multiple crossing fibers within a single voxel. However, the raw data acquired in HARDI is frequently plagued by noise and lacks inherent smoothness. To address this, researchers employ regularization—a mathematical process designed to suppress noise while preserving structural integrity.

Traditionally, regularization schemes have relied on first-order (Tikhonov) or second-order methods. This paper introduces a novel "infinite order" regularization framework based on scale-space theory, establishing a rigorous mathematical connection between scale-space representations and the widely used Tikhonov regularization.

The Codomain Scale Space Paradigm

The core of the proposed method is the concept of a codomain scale space. Rather than applying smoothing to the spatial domain (the physical location of the voxels), the author applies it to the codomain—the unit sphere where the HARDI signal resides.

This approach is operationalized through an infinite-order partial differential equation (PDE), specifically the heat equation applied to the unit sphere. To solve this PDE efficiently, the author utilizes spherical harmonics. This decomposition "diagonalizes" the Laplace-Beltrami operator (the spherical equivalent of the Laplacian), transforming complex differential equations into manageable algebraic operations. This allows the "infinite order" smoothing to be implemented as a simple, one-parameter filter.

Linking Scale Space to Tikhonov Regularization

The most significant theoretical contribution of this work is the establishment of a closed-form link between scale-space theory and Tikhonov regularization. The author proves that the Tikhonov-regularized signal is a "scale superposition" of the scale-space signal, related via the Laplace transform.

The Operator Identity

The paper establishes the following identity: $$\tilde{S}_s = (I - s\Delta_g)^{-1} S = \frac{1}{s} \int_0^\infty \exp(-t/s) S_t dt$$

In this formula, $S_t$ represents the scale-space signal at "blurring time" $t$, and $\tilde{S}_s$ represents the Tikhonov-regularized signal with parameter $s$. This identity demonstrates that scale-space is the more fundamental, or "parent," framework. While a Tikhonov signal can always be derived from a scale-space representation, the reverse process is mathematically ill-posed.

Physical Interpretation

Beyond the abstract math, the author provides a physical interpretation of the relationship: the Tikhonov regularization parameter ($s$) is effectively the expectation value (the average) of the scale-space blurring time ($t$) under a negative-exponential distribution. This provides a bridge between the temporal evolution of the heat equation and the static smoothing parameters used in clinical practice.

Spectral Analysis and Attenuation

By examining the attenuation coefficients of spherical harmonics, the paper reveals why these methods behave differently in practice:

  • Scale Space: Coefficients decay exponentially: $e^{-t\ell(\ell+1)}$.
  • Tikhonov Regularization: Coefficients decay as a rational function: $1/(1 + s\ell(\ell+1))$.

This comparison shows that scale-space regularization suppresses high-frequency noise much more aggressively than Tikhonov regularization. Both methods, however, maintain asymptotic consistency: at a scale of zero, they return the raw noisy signal; at an infinite scale, they both converge to a perfect average of the signal over the entire sphere.

Conclusion and Future Directions

This paper provides a robust theoretical foundation for HARDI regularization by unifying scale-space theory and Tikhonov methods. While the work is primarily mathematical, it suggests that "infinite order" smoothing is not only computationally feasible via spherical harmonics but also offers a more fundamental approach to noise reduction.

As this study is conceptual, the primary limitation is the lack of empirical validation on clinical datasets. Future research is required to determine how these different smoothing orders affect practical outcomes, such as the accuracy of fiber tracking (tractography) and the resolution of crossing fibers in real-world medical imaging pipelines.

Summary updated

BibTeX
@inproceedings{florack2008codomain,
  author = {Luc Florack},
  title = {{Codomain scale space and regularization for high angular resolution diffusion imaging}},
  booktitle = {Workshops @ CVPR},
  year = {2008},
  doi = {10.1109/cvprw.2008.4562967},
}
Figure