Estimation of Non-negative ODFs Using the Eigenvalue Distribution of Spherical Functions

Evan Schwab, Bijan Afsari, René VidalMICCAI 2012 · OpenReview

Abstract

Current methods in high angular resolution diffusion imaging (HARDI) estimate the probability density function of water diffusion as a continuous-valued orientation distribution function (ODF) on the sphere. However, such methods could produce an ODF with negative values, because they enforce non-negativity only at finitely many directions. In this paper, we propose to enforce non-negativity on the continuous domain by enforcing the positive semi-definiteness of Toeplitz-like matrices constructed from the spherical harmonic representation of the ODF. We study the distribution of the eigenvalues of these matrices and use it to derive an iterative semi-definite program that enforces non-negativity on the continuous domain. We illustrate the performance of our method and compare it to the state-of-the-art with experiments on synthetic and real data.

Introduction

In High Angular Resolution Diffusion Imaging (HARDI), Orientation Distribution Functions (ODFs) are essential for representing the probability of water molecule diffusion in specific directions. As probability density functions, ODFs must inherently be non-negative. However, standard estimation techniques using Spherical Harmonics (SH) often fail this requirement due to noise or low-order approximations, resulting in "negative probabilities."

Traditional state-of-the-art methods attempt to address this by enforcing non-negativity at a discrete set of points on a grid. Unfortunately, this does not guarantee non-negativity across the entire continuous surface of the sphere. Such mathematical artifacts can significantly skew statistical comparisons in clinical population studies. This paper introduces a novel approach to ensure strict non-negativity across the continuous domain by leveraging the spectral properties of Toeplitz-like matrices.

Methodology

The proposed method shifts from local point-checking to a global structural constraint on the SH coefficients. The approach is built upon three mathematical pillars:

  1. Toeplitz Analogue for Spheres: Extending 1D signal processing concepts, the authors construct a matrix ($T_\ell$) from SH coefficients using Gaunt Coefficients. This matrix serves as a spherical analogue to the Toeplitz matrix used in Fourier analysis.
  2. Eigenvalue Distribution Theorem: A key theoretical insight is that as the size of the $T_\ell$ matrix increases, its smallest eigenvalue converges to the absolute minimum value of the ODF on the sphere.
  3. Iterative Semi-Definite Programming (SDP): To handle the infinite-dimensional nature of the problem, an iterative algorithm is employed. The method calculates eigenvalues for several finite matrix sizes and uses a curve-fitting technique ($q \frac{1}{(\ell+d)^r} + b$) to predict the "true" minimum ($b$) of the continuous ODF. An SDP is then solved to adjust the SH coefficients until this predicted minimum is non-negative.

Experimental Evaluation

The method was rigorously tested using both synthetic and real-world datasets:

  • Synthetic Data: A field of 100 ODFs simulating 1, 2, and 3 crossing fibers with complex Gaussian noise added at SNR levels of 5, 10, and 20 dB.
  • Real Data: A HARDI human brain dataset (128 x 128) with 94 gradient directions.

To verify non-negativity, evaluations were performed over an extremely fine mesh of 1.002 million points on the sphere.

Results

The "Continuous Non-negativity" (CN) method demonstrated significant improvements over existing techniques in both accuracy and physical validity.

Quantitative performance comparison: Average Riemannian error and percentage of negative area across SNR levels.

Elimination of Negative Values

On the high-resolution test mesh, the CN method achieved 0% negative values. In comparison, the previous "Discrete Non-negativity" (DN) method left negative values in over 99% of the estimated ODFs, with an average negative area of 0.04%.

Accuracy and Visual Fidelity

The CN method consistently yielded lower Riemannian error than both unconstrained least squares (LS) and discrete methods, particularly in high-noise environments (SNR 5 dB). Qualitatively, the reconstructions on real brain data appear sharper and more physically plausible, successfully removing the negative lobes that typically plague standard reconstructions.

Visual comparison of ODF field reconstructions demonstrating the elimination of negative lobes.

Discussion and Conclusion

While the CN method offers superior accuracy and physical validity, it comes with a higher computational cost. The SDP approach has a complexity of approximately $O(L^6)$, resulting in processing times of 70–159 seconds per ODF—roughly two orders of magnitude slower than discrete methods. Future work will focus on optimizing the implementation, potentially moving from general-purpose solvers to custom algorithms to improve runtime.

By ensuring that ODFs strictly adhere to the mathematical axioms of probability distributions, this method provides a more robust foundation for large-scale population studies and the detection of biomarkers for neurological diseases.

Summary updated

Abstract

Current methods in high angular resolution diffusion imaging (HARDI) estimate the probability density function of water diffusion as a continuous-valued orientation distribution function (ODF) on the sphere. However, such methods could produce an ODF with negative values, because they enforce non-negativity only at finitely many directions. In this paper, we propose to enforce non-negativity on the continuous domain by enforcing the positive semi-definiteness of Toeplitz-like matrices constructed from the spherical harmonic representation of the ODF. We study the distribution of the eigenvalues of these matrices and use it to derive an iterative semi-definite program that enforces non-negativity on the continuous domain. We illustrate the performance of our method and compare it to the state-of-the-art with experiments on synthetic and real data.
BibTeX
@inproceedings{schwab2012estimation,
  author = {Evan Schwab and Bijan Afsari and René Vidal},
  title = {{Estimation of Non-negative ODFs Using the Eigenvalue Distribution of Spherical Functions}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2012},
  url = {https://openreview.net/forum?id=wn4mi_Y2JQJ},
}
Figure