Empirical estimation of intravoxel structure with persistent angular structure and Q-ball models of diffusion weighted MRI

Vishwesh Nath, Kurt G. Schilling, Prasanna Parvathaneni, Justin A. Blaber, Allison E. Hainline, Zhaohua Ding, Adam W. Anderson, Bennett A. LandmanJournal of Medical Imaging 2018 · DOI

TL;DR

This paper empirically compares two advanced MRI techniques, Persistent Angular Structure (PAS-MRI) and Q-ball imaging, for mapping complex white matter structures in the brain. They found that PAS-MRI is more reproducible, especially at lower b-values, but the two methods suggest fundamentally different brain microarchitectures, cautioning against combining results from studies using different approaches.

What Problem Does This Paper Address?

Diffusion Tensor Imaging (DTI), a standard MRI technique, struggles to accurately represent complex white matter structures in the brain where nerve fibers cross, bend, or branch within a single voxel. This limitation hinders our understanding of brain connectivity. To overcome this, High Angular Resolution Diffusion Imaging (HARDI) methods like PAS-MRI and Q-ball imaging were developed. However, their empirical performance and reproducibility, especially in real-world clinical scenarios, were not well-characterized. This paper aims to bridge that gap by empirically comparing PAS-MRI and Q-ball imaging using a carefully acquired dataset to assess their reproducibility and agreement.

Methods

The study employed a rigorous data acquisition and processing pipeline. A single healthy volunteer was scanned multiple times (11 sessions) with varying diffusion weighting parameters (b-values of 1000 and 3000 s/mm²) and a high number of diffusion directions (96). This allowed for the creation of "gold standard" models by combining data from multiple scans.

The core of the methodology involved:

  • Data Pre-processing: Standard steps like eddy current correction, motion correction, and susceptibility distortion correction were applied using established tools (FSL's topup and eddy).
  • Registration and Normalization: Data from different sessions were registered to a common MNI (Montreal Neurological Institute) space for consistent analysis.
  • HARDI Model Reconstruction:
    • Q-ball: Reconstructed using Spherical Harmonic Basis Functions with an order of 8, following established methods.
  • PAS-MRI: Reconstructed using maximum entropy spherical deconvolution with a reduced encoding model.
  • Peak Reconstruction: Both methods' outputs were processed to identify the dominant fiber orientations (peaks) using a peak-finding algorithm.
  • Metrics for Comparison:
    • Success Fraction (SF): Measures agreement between estimated fiber populations and their orientations, with a tolerance of 20 degrees.
  • Peak Fraction: Quantifies the strength of detected fiber populations.
  • Symmetric Angular Error (SAE): Assesses the orientation error between estimated and gold standard fiber peaks.

Datasets

The study utilized a newly acquired, high-quality dataset from a single healthy volunteer. This dataset consisted of:

  • 11 scanning sessions over successive days.
  • 96 diffusion-weighted gradient directions per b-value.
  • Two b-values: 1000 s/mm² and 3000 s/mm².
  • 10 minimally weighted reference images (b0's) per b-value shell.

This resulted in a total of 2144 diffusion-weighted volumes. The dataset was designed to specifically assess reproducibility and inter-model agreement, serving as a benchmark for comparing PAS-MRI and Q-ball. No pre-existing public datasets were used for the primary analysis.

Key Insights and Ideas

  • PAS-MRI is More Reproducible: The study consistently found PAS-MRI to be more reproducible than Q-ball imaging, particularly at lower b-values (1000 s/mm²). This suggests PAS-MRI might be more reliable in clinical settings where higher b-values are not always feasible.
  • Different Brain Microarchitectures: A striking finding was the significant disagreement between the intravoxel structures estimated by PAS-MRI and Q-ball. This suggests that these two methods, while both aiming to capture complex fiber arrangements, might be sensitive to different aspects of the underlying tissue microstructure, leading to fundamentally different interpretations of brain connectivity.
  • B-value Matters: The choice of b-value (1000 vs. 3000 s/mm²) significantly impacted the results for both methods, highlighting the importance of acquisition parameters in HARDI.
  • False Positives in PAS-MRI: For single-fiber scenarios, PAS-MRI showed higher agreement but also a higher standard deviation at lower FA values, suggesting a potential for detecting false positives. Q-ball, while less reproducible in these cases, was more consistent.
  • Three-Fiber Scenarios Remain Challenging: Both methods struggled with three-fiber populations, showing low reproducibility and agreement, indicating that accurately resolving such complex arrangements remains a significant challenge.

Key Results

  • Success Fraction (SF):
    • For single-fiber cases, PAS-MRI showed higher SF than Q-ball, especially when PAS was the gold standard.
  • For two-fiber cases, Q-ball generally agreed with PAS at high b-values, but a substantial portion of PAS two-fiber cases were not detected by Q-ball.
  • Three-fiber cases showed poor agreement for both methods.
  • Symmetric Angular Error (SAE):
    • Single-fiber cases had lower SAE than multi-fiber cases for both methods.
  • SAE generally decreased with increasing Fractional Anisotropy (FA).
  • PAS-MRI showed lower SAE than Q-ball in most scenarios.
  • Interestingly, SAE was lower for three crossing fibers compared to two crossing fibers.
  • Peak Fraction:
    • Q-ball identified more voxels with single fiber populations than PAS at both b-values.
  • PAS showed a wider spread of peak fractions compared to Q-ball's narrower distribution.
  • Inter-Model Comparison: Generally showed low agreement between PAS and Q-ball, suggesting they capture different aspects of intravoxel structure.

Limitations and Future Work

  • Single Subject Study: The study was conducted on a single healthy volunteer. While this allowed for a controlled and reproducible dataset, it limits the generalizability of the findings to a broader population. Future work should involve larger cohorts with diverse demographics and pathologies.
  • Low SNR: The study acknowledged that the signal-to-noise ratio (SNR) of the acquired data was relatively low, which can impact the accuracy and reproducibility of HARDI methods. Future studies could investigate performance with higher SNR data.
  • Three-Fiber Complexity: The challenges in accurately resolving three-fiber populations persist. The authors suggest that robust fitting of HARDI methods, potentially with outlier volume removal, could lead to improvements.
  • Histological Validation: The paper mentions that comparison with histological validation could indicate which methods are more accurate in regions of disagreement. This is a crucial next step for ground-truthing these imaging techniques.
  • Deeper Understanding of ODF vs. FOD: The authors highlight the need for a deeper understanding of the differences between Orientation Distribution Function (ODF) and Fiber Orientation Distribution (FOD) methods, particularly regarding peak fractions, to better interpret the results from different approaches.
  • Clinical Feasibility of B-values: While higher b-values generally improve Q-ball performance, they are not always clinically feasible. The study's focus on 1000 and 3000 s/mm² is relevant for clinical applications.

Summary updated

Abstract

The diffusion tensor model is nonspecific in regions where micrometer structural patterns are inconsistent at the millimeter scale (i.e., brain regions with pathways that cross, bend, branch, fan, etc.). Numerous models have been proposed to represent crossing fibers and complex intravoxel structure from <i>in vivo</i> diffusion weighted magnetic resonance imaging (e.g., high angular resolution diffusion imaging-HARDI). Here, we present an empirical comparison of two HARDI approaches-persistent angular structure MRI (PAS-MRI) and Q-ball-using a newly acquired reproducibility dataset. Briefly, a single subject was scanned 11 times with 96 diffusion weighted directions and 10 reference volumes for each of two [Formula: see text] values (1000 and [Formula: see text] for a total of 2144 volumes). Empirical reproducibility of intravoxel fiber fractions (number/strength of peaks), angular orientation, and fractional anisotropy was compared with metrics from a traditional tensor analysis approach, focusing on [Formula: see text] values of 1000 and [Formula: see text]. PAS-MRI is shown to be more reproducible than Q-ball and offers advantages at low [Formula: see text] values. However, there are substantial and biologically meaningful differences between the intravoxel structures estimated both in terms of analysis method as well as by [Formula: see text] value. The two methods suggest a fundamentally different microarchitecture of the human brain; therefore, it is premature to perform meta-analysis or combine results across HARDI studies using a different analysis model or acquisition sequences.
BibTeX
@article{nath2018empirical,
  author = {Vishwesh Nath and Kurt G. Schilling and Prasanna Parvathaneni and Justin A. Blaber and Allison E. Hainline and Zhaohua Ding and Adam W. Anderson and Bennett A. Landman},
  title = {{Empirical estimation of intravoxel structure with persistent angular structure and Q-ball models of diffusion weighted MRI}},
  journal = {Journal of Medical Imaging},
  year = {2018},
  doi = {10.1117/1.jmi.5.1.014005},
}
Figure