Structured sparsity for spatially coherent fibre orientation estimation in diffusion MRI

Anna Auría, Alessandro Daducci, Jean-Philippe Thiran, Yves WiauxNeuroImage 2015 · DOI

TL;DR

This paper introduces a novel method for reconstructing the complex 3D structure of white matter fibers in the brain from diffusion MRI data. By leveraging the spatial coherence of these fibers across neighboring voxels, their approach achieves more accurate and robust fiber orientation estimates, even with significantly reduced data acquisition.

What Problem Does This Paper Address?

Diffusion MRI (dMRI) is a powerful, non-invasive technique for studying the brain's white matter by measuring water diffusion. The challenge lies in accurately inferring the orientation and distribution of nerve fiber bundles within each tiny voxel of the brain. Traditional methods often treat each voxel independently, which can lead to inaccuracies, especially in regions where multiple fiber bundles cross. State-of-the-art techniques like Diffusion Spectrum Imaging (DSI) require a large number of measurements, making them time-consuming for clinical use. Spherical deconvolution (SD) methods offer a compromise, but still struggle with accuracy and robustness, particularly when data is sparse or noisy. This paper aims to improve the accuracy and efficiency of fiber orientation estimation by incorporating spatial information directly into the reconstruction process.

Methods

The core of this work is a new formulation for reconstructing the Fiber Orientation Distribution Function (FOD) across multiple voxels simultaneously. Instead of processing each voxel in isolation, the method exploits the anatomical assumption that fiber orientations are spatially coherent between neighboring voxels. This is achieved by defining a "spatially structured sparsity prior" directly on the FOD field. The algorithm uses a reweighting scheme, inspired by compressed sensing techniques, to enforce this structured sparsity. Essentially, it iteratively refines the FOD estimates by considering the information from surrounding voxels, promoting smoother and more anatomically plausible fiber orientations. This approach is referred to as L2L0NW.

Datasets

The authors evaluated their method using two types of datasets:

  • Simulated Phantom Data: This dataset was generated to mimic real dMRI data and was used for the HARDI reconstruction Challenge 2012. It features a 16x16x5 volume with 5 different fiber bundles, including crossing and kissing tracts, with varying fractional anisotropy and constant diffusion properties. The data was artificially contaminated with Rician noise at different Signal-to-Noise Ratios (SNR = 10, 20, 30) and undersampled across various q-space acquisition schemes (30, 20, 15, 10, and 6 samples).
  • Real HARDI Data: A dataset acquired at b=3000 s/mm² with 256 directions was used. This dataset was retrospectively undersampled to create three additional datasets with 30, 20, and 10 diffusion directions. The SNR in the b=0 images was approximately 30.

Key Insights and Ideas

  • Spatial Coherence is Key: The most significant insight is that explicitly incorporating the spatial coherence of fiber orientations between neighboring voxels dramatically improves reconstruction accuracy and robustness, especially under challenging conditions like low SNR or undersampling.
  • Structured Sparsity: Moving beyond voxel-wise sparsity, the paper introduces a "spatially structured sparsity prior" that directly models the anatomical continuity of white matter tracts. This is a more powerful constraint than simply enforcing sparsity within each voxel.
  • Efficiency Gains: By leveraging spatial information, the method can achieve comparable or superior results with significantly fewer diffusion-weighted images (e.g., as few as 15 samples), leading to faster acquisition times and making it more clinically viable.
  • Robustness to Noise and Undersampling: The proposed L2L0NW method demonstrates remarkable resilience to both low SNR and reduced numbers of diffusion gradient directions, outperforming existing methods in these scenarios.

Key Results

  • Phantom Data: L2L0NW consistently outperformed the voxel-by-voxel L2L0 method and the denoising-based JAMMLSE+L2L0 approach across various undersampling rates and noise levels. For instance, at SNR=20 and with only 10 samples, L2L0NW achieved a success rate (SR∠) of 72%, while L2L0 and JAMMLSE+L2L0 dropped to 36% and 38%, respectively.
  • Real Data: Similar improvements were observed with real data. For the hardi10 dataset (10 samples), L2L0NW achieved a mean angular error (θ) of 13.6° ± 10.5°, significantly better than L2L0 (19.8° ± 11.25°) and JAMMLSE+L2L0 (19.8° ± 12.0°). Qualitatively, L2L0NW reconstructions showed neater, sharper fiber profiles with fewer spurious peaks, especially in regions with complex fiber crossings.
  • Comparison to Other Methods: When compared to a suite of state-of-the-art methods from the HARDI reconstruction Challenge 2012, L2L0NW achieved comparable or better success rates (SR∠) using significantly fewer samples (e.g., 15 samples) than many other methods.

Limitations and Future Work

  • Computational Cost: While the method is efficient in terms of data acquisition, the reweighting scheme still involves iterative computations. The authors note that their implementation is affordable for a single processor.
  • Fixed Bound: The current formulation uses a single bound for the total number of fibers across the entire volume. Preliminary investigations suggest that fixing individual bounds per voxel might be too constraining.
  • Future Work: The authors suggest exploring further improvements by enabling undersampling in Fourier space (k-space) for each acquired q-space image, which could lead to even greater acceleration. They also propose extending their framework to handle more complex microstructural features and configurations in regions with multiple fiber populations. The integration of spatial regularization of q-space images as an additional prior is also mentioned as a potential enhancement.

Summary updated

BibTeX
@article{auria2015structured,
  author = {Anna Auría and Alessandro Daducci and Jean-Philippe Thiran and Yves Wiaux},
  title = {{Structured sparsity for spatially coherent fibre orientation estimation in diffusion MRI}},
  journal = {NeuroImage},
  year = {2015},
  doi = {10.1016/j.neuroimage.2015.04.049},
}
Figure