Non-negative Spherical Deconvolution (NNSD) for Fiber Orientation Distribution Function Estimation

Jian Cheng, Rachid Deriche, Tianzi Jiang, Dinggang Shen, Pew-Thian YapCDMRI/MMBC @ MICCAI 2013 · DOI

TL;DR

This paper introduces Non-Negative Spherical Deconvolution (NNSD), a novel method for estimating the fiber orientation distribution function (fODF) in diffusion MRI. NNSD guarantees non-negativity of the fODF across the entire unit sphere, unlike previous methods, and offers improved accuracy and efficiency, particularly for high-resolution data.

What Problem Does This Paper Address?

Diffusion Magnetic Resonance Imaging (dMRI) is a powerful tool for visualizing white matter by tracking the diffusion of water molecules. However, traditional methods like Diffusion Tensor Imaging (DTI) struggle with complex fiber structures (e.g., crossings, kissing fibers). High Angular Resolution Diffusion Imaging (HARDI) techniques aim to overcome this by estimating the fiber Orientation Distribution Function (fODF), which describes the probability of fiber orientations within a voxel.

Spherical Deconvolution (SD) is a common approach to estimate fODFs. It models the diffusion signal as a convolution of the fODF with a fiber response function. However, existing SD methods face several limitations:

  • Incomplete Non-negativity: Many methods only enforce non-negativity on discretized points of the unit sphere, leading to potential negative fODF values elsewhere.
  • False Positive Peaks: They can produce spurious peaks, especially in regions with low anisotropy (e.g., gray matter, CSF), leading to an overestimation of fiber bundles.
  • Computational Inefficiency: Some methods that do enforce non-negativity globally require computationally expensive numerical integration.
  • Limited Angular Resolution: Discrete representation-based methods can restrict peak detection to predefined orientations.

This paper addresses these issues by proposing NNSD, a method that guarantees global non-negativity and offers improved accuracy and efficiency.

Methods

The core innovation of NNSD lies in its representation of the fODF. Instead of directly modeling the fODF, NNSD represents its square root as a linear combination of Spherical Harmonics (SH) basis functions. This clever approach ensures that the reconstructed fODF is inherently non-negative.

Key aspects of the NNSD method include:

  • Square Root Representation: $\Phi(u) = (\sum c_{\ell m} Y_{\ell m}(u))^2$, where $\Phi(u)$ is the fODF, $c_{\ell m}$ are SH coefficients, and $Y_{\ell m}(u)$ are SH basis functions. This formulation guarantees $\Phi(u) \ge 0$.
  • SH Basis Functions: Utilizing SH basis functions allows for a closed-form solution for spherical integration, which is crucial for efficient computation.
  • Riemannian Gradient Descent: The optimization problem to estimate the SH coefficients is solved using a Riemannian gradient descent algorithm, which operates on the manifold of unit-norm vectors.
  • Adaptive Stopping Condition (NNSD-ASC): An adaptive stopping condition is introduced to balance convergence speed and accuracy, particularly for noisy data.

NNSD is presented as a Continuous Representation-based SD (CR-SD) method, leveraging the advantages of SH representations for efficient computation and accurate peak detection on the continuous unit sphere.

Datasets

The paper evaluates NNSD using:

  • Synthetic Data: Generated signals simulating both isotropic and anisotropic fiber profiles, with varying Signal-to-Noise Ratios (SNRs) and crossing angles. This synthetic data allows for controlled evaluation of anisotropy, non-negativity, and crossing fiber resolution.
  • Real Human Data:
    • Data from the Human Connectome Project (HCP) with high resolution (1.25mm³), characterized by low SNR.
  • Additional real human data with b=2000s/mm², 120 gradient directions, and 2mm isotropic voxel dimensions.

The paper does not introduce new benchmark datasets but rather uses existing data to compare NNSD against established methods.

Key Insights and Ideas

  • The Power of the Square Root: Representing the square root of the fODF is a simple yet elegant solution to enforce global non-negativity, a persistent challenge in fODF estimation.
  • SH Efficiency: The use of SH basis functions is key to NNSD's computational efficiency, enabling closed-form solutions for integration and faster processing compared to methods relying on numerical integration.
  • Robustness to Noise and Anisotropy: NNSD, especially its adaptive variant (NNSD-ASC), demonstrates remarkable robustness in low-anisotropy regions and handles noisy, high-resolution data effectively, reducing spurious peaks.
  • Improved Contrast: NNSD-ASC shows a better contrast between isotropic and anisotropic regions in GFA maps, indicating more accurate fiber tract characterization.

Key Results

  • Guaranteed Non-negativity: NNSD is the first SH-based method to guarantee non-negativity of the fODF across the entire unit sphere.
  • Reduced False Positives: NNSD significantly reduces false positive peaks, particularly in low-anisotropy regions, compared to CSD, MESD, and L1NNLS.
  • Improved Accuracy: NNSD-ASC achieves higher success ratios and lower Mean Difference of Angles (MDA) for crossing fibers, especially at larger crossing angles.
  • Efficiency: NNSD is computationally efficient, with an implementation capable of estimating fODFs for 1000 voxels in about 8 seconds on a laptop (for L=6).
  • High-Resolution Performance: NNSD-ASC shows particular strength on high-resolution, low-SNR data (HCP), maintaining isotropic fODFs in isotropic regions and sharp fODFs in anisotropic regions, outperforming other methods.

Limitations and Future Work

  • Parameter Sensitivity: While NNSD is generally robust, the paper notes that the outcome can be sensitive to the choice of the stopping condition parameter $\delta$ (or $T$ for NNSD-ASC). For instance, a very small $\delta$ can lead to fuzzier fODFs in low-anisotropy regions.
  • Single Shell Data: The current implementation primarily focuses on single-shell data. The authors suggest that NNSD can be extended to multi-shell data by incorporating a signal response function defined in 3D space.
  • Denoising and Spatial Regularization: The experiments were conducted without pre-processing steps like denoising or spatial regularization. The authors acknowledge that incorporating these techniques could further improve results in practical applications.
  • Future Applications: The authors believe the clean fODF fields generated by NNSD will be highly beneficial for downstream applications such as fiber tractography.

Summary updated

BibTeX
@inproceedings{cheng2013non,
  author = {Jian Cheng and Rachid Deriche and Tianzi Jiang and Dinggang Shen and Pew-Thian Yap},
  title = {{Non-negative Spherical Deconvolution (NNSD) for Fiber Orientation Distribution Function Estimation}},
  booktitle = {CDMRI/MMBC @ MICCAI},
  year = {2013},
  doi = {10.1007/978-3-319-02475-2\_8},
}
Figure