Susceptibility-induced distortion that varies due to motion: Correction in diffusion MR without acquiring additional data

Jesper L. R. Andersson, Mark S. Graham, Ivana Drobnjak, Hui Zhang, Jon CampbellNeuroImage 2018 · DOI

TL;DR

This paper introduces a novel method to correct for distortions in diffusion MRI caused by magnetic field variations that change with subject movement. By cleverly using the diffusion data itself, the technique estimates and corrects these dynamic distortions without requiring extra scans, making it more efficient and applicable to existing datasets.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI), particularly Echo Planar Imaging (EPI) used in diffusion MRI, is susceptible to distortions caused by variations in the main magnetic field. These variations, known as susceptibility-induced off-resonance fields, arise because different tissues (like air in sinuses) disrupt the magnetic field unevenly. Traditionally, these distortions are corrected by measuring a "field map" once. However, this assumes the field map is static. The problem is that when a subject moves, especially with rotations not aligned with the main magnetic field, the way their head disrupts the field changes, leading to different distortions in each acquired volume. This means a single, static field map correction is insufficient, leaving residual distortions that can affect the accuracy of diffusion MRI analyses. Existing methods to address this either require extra scans or are not well-suited for diffusion MRI.

Methods

The authors propose a post-processing method that estimates the susceptibility-induced field as it changes with subject motion during the scan. The core idea is to use a generative model that predicts how each diffusion-weighted image (DWI) volume should look based on estimated motion and field distortions. By comparing these predictions to the actual acquired images, the method iteratively refines estimates of the changing susceptibility field.

Key aspects of the method include:

  • Taylor Expansion of the Field: The changing susceptibility field is modeled as a Taylor expansion around a reference orientation. This allows the field to be represented by a base field and its derivatives with respect to rotational parameters (pitch and roll).
  • Generative Model: A model predicts the observed DWI signal, accounting for rigid-body motion, eddy currents, and the dynamic susceptibility field.
  • Inverse Problem: The method frames the problem as an inverse one: given the observed data and the generative model, estimate the parameters of the dynamic susceptibility field. This is solved using a Gauss-Newton iterative approach.
  • No Additional Acquisitions: Crucially, the method does not require acquiring extra field maps or specialized scans. It leverages the inherent information within the diffusion data itself.
  • Integration with Existing Tools: The method is implemented as an extension to the existing eddy tool, which already corrects for eddy currents and subject motion.

Datasets

The authors validated their method using:

  • Simulated Data: Realistic diffusion MRI data was generated using the POSSUM simulator. This allowed for precise control over movement types (volumetric vs. continuous), movement magnitudes (normal vs. large), eddy currents (present/absent), phase-encode direction, and signal-to-noise ratio (SNR). This provided a "ground truth" for comparison.
  • Human Data: Data from a healthy volunteer scanned multiple times under different motion conditions (lying still, controlled movement, mimicking movement) was used to test the method on real-world scenarios.

Key Insights and Ideas

  • Motion-Induced Field Variability is Significant: The study highlights that subject motion, particularly out-of-plane rotations, causes substantial and non-trivial changes in susceptibility-induced distortions, rendering static field map corrections inadequate.
  • Diffusion Data is Rich with Information: The core insight is that the diffusion-weighted images themselves contain enough information to estimate these dynamic field changes, eliminating the need for extra acquisition sequences.
  • Efficiency Through Iteration: By framing the problem as an inverse problem solved iteratively, the method efficiently refines estimates of the dynamic field without requiring explicit field measurements.
  • Synergy with Existing Corrections: The method seamlessly integrates with existing motion and eddy current correction pipelines, offering a comprehensive solution.

Key Results

  • Accurate Derivative Field Estimation: The method accurately estimates the rate of change of the susceptibility field with respect to subject orientation (pitch and roll), achieving high correlations with the "true" fields in simulations. For example, correlations between estimated and true derivative fields were consistently high across various simulation parameters.
  • Improved Image Correction: Correcting for the dynamic susceptibility field significantly improved the correlation between corrected images and the "ground truth" compared to methods that only correct for static susceptibility, eddy currents, and motion. For "large" movements, the improvement was particularly pronounced.
  • Robustness: The method demonstrated robustness across different SNRs (20-40) and phase-encode directions. The presence of eddy currents also did not significantly degrade performance when jointly estimated.
  • Test-Retest Reproducibility: Human data showed good test-retest reproducibility for the estimated derivative fields, indicating the reliability of the method.

Limitations and Future Work

  • Movement Magnitude: The accuracy of estimating the "true" derivative fields was somewhat affected by the magnitude of movement, with larger movements showing slightly better results, potentially due to a clearer "estimate of truth" or more information about susceptibility-by-movement effects.
  • Spin-Echo EPI: The method is not directly applicable to spin-echo EPI data due to its different phase behavior.
  • Future Work:
    • Breathing: The authors suggest extending the framework to account for breathing-induced field variations, which also change over time.
  • Prospective Motion Correction: Combining the method with prospective motion correction could simplify the algorithm by using external motion estimates.
  • Higher b-values: The method is expected to work well for higher b-values (up to 5000) based on previous work.
  • Other Movement Parameters: The framework can be extended to model changes in the field with respect to all rigid-body movement parameters, not just pitch and roll.

Summary updated

BibTeX
@article{andersson2018susceptibility,
  author = {Jesper L. R. Andersson and Mark S. Graham and Ivana Drobnjak and Hui Zhang and Jon Campbell},
  title = {{Susceptibility-induced distortion that varies due to motion: Correction in diffusion MR without acquiring additional data}},
  journal = {NeuroImage},
  year = {2018},
  doi = {10.1016/j.neuroimage.2017.12.040},
}

Loading related content...

Figure