Anatomy-based quality metric of diffusion-weighted MRI data for accurate derivation of muscle fiber orientation

Nadya Shusharina, Xiaofeng Liu, Evangelia Kaza, Miranda Lam, Stephan E. Maier, Jonghye WooarXiv 2024 · DOI · arXiv

TL;DR

Researchers have developed a new "Mixing Index" to evaluate the quality of diffusion-weighted MRI (DW-MRI) data in skeletal muscles. By using the known anatomy of bipennate muscles—which have two distinct fiber directions—as a built-in benchmark, this metric provides a single scalar value to optimize the trade-off between image scan time and the accuracy of muscle fiber mapping.

What Problem Does This Paper Address?

In treating soft tissue sarcomas, clinicians must estimate how microscopic tumor cells spread beyond the visible mass. These cells tend to migrate along the paths of least resistance: the 3D interfaces of muscle fibers. DW-MRI is the primary tool for mapping these fiber orientations, but it is notoriously sensitive to noise.

To get high-quality data (high Signal-to-Noise Ratio or SNR), patients often need to stay still for long scan durations, which is difficult in a clinical setting. Currently, there isn't a robust way to determine if a "fast" scan is "good enough" for accurate fiber tracking. The authors address this gap by creating an anatomy-based quality metric that moves beyond simple SNR to measure how well the MRI can actually distinguish between different biological structures.

Methods

The researchers focused on the rectus femoris, a "bipennate" muscle in the thigh. Think of a bipennate muscle like a feather: fibers attach to a central tendon at two different angles. This known geometry provides a perfect "ground truth" for testing MRI accuracy.

Key technical steps included:

  1. Imaging: Scanning a healthy volunteer on a 3T MRI at two different resolutions (resulting in high and low SNR).
  2. Segmentation: Using both manual labeling and a 2D U-Net (convolutional neural network) to identify 12 individual thigh muscles.
  3. PCA & Clustering: The diffusion tensor at each voxel has a principal eigenvector (the fiber direction). The authors used Principal Component Analysis (PCA) to reduce the 9-dimensional eigenvector data into a 3D space, then applied k-means clustering to group voxels with similar orientations.
  4. The Mixing Index (MI): They adapted a "nearest-neighbor mixing index" from materials science. This formula calculates how much voxels from the two different fiber compartments (the two sides of the "feather") are spatially intermingled. A low MI indicates a clean, accurate separation of the two fiber groups.

Datasets

The study utilized a dataset acquired from a single healthy volunteer:

  • Morphological Data: T1-weighted MR images for anatomical reference.
  • Diffusion Data: DW-MRI scans with 12 gradient directions, performed at two resolutions ($3 \times 3 \text{ mm}^2$ for high SNR and $1.3 \times 1.3 \text{ mm}^2$ for low SNR).
  • Benchmarks: Manual segmentations were compared against automated segmentations produced by a U-Net trained on previously published thigh muscle data.

Key Insights and Ideas

  • Anatomy as a Calibration Tool: The most elegant idea here is using "biological ground truth." Because we know the rectus femoris has two distinct fiber directions separated by a tendon, any "mixing" of these directions in the data must be due to noise or poor resolution.
  • Bimodal vs. Unimodal Distributions: In high-SNR scans, the histogram of fiber angles shows two distinct peaks (bimodal), representing the two sides of the bipennate muscle. In low-SNR scans, these peaks merge into one blurry "blob," proving that lower quality data loses the ability to resolve complex anatomy.
  • The "Mixing Index" Advantage: Unlike SNR, which only tells you about signal strength vs. background noise, the Mixing Index jointly characterizes both spatial and directional uncertainty. It tells you if the noise is actually breaking the biological model you're trying to build.

Key Results

  • Mixing Index Sensitivity: The high-SNR scan yielded an MI of 0.16, while the low-SNR scan rose to 0.22. This confirms the metric can quantitatively distinguish between different acquisition protocols.
  • Segmentation Accuracy: The U-Net achieved a Dice Similarity Coefficient (DSC) of 0.80 for the rectus femoris.
  • The "Good Enough" Threshold: Interestingly, the authors found that a DSC of 0.80—while generally considered "good" in medical imaging—was not sufficient for high-precision fiber analysis. The automated segmentation included "edge" voxels from neighboring muscles that contaminated the fiber orientation clusters, suggesting that fiber-tracking applications require even higher segmentation precision than standard volume measurements.

Limitations and Future Work

The primary limitation is the small sample size (one volunteer), which served as a proof-of-concept for the metric. Additionally, the U-Net segmentation needs refinement to avoid "bleeding" into adjacent muscles, which skewed the Mixing Index.

Future work aims to establish the exact mathematical relationship between the Mixing Index, SNR, and specific acquisition parameters (like the number of gradient directions). Ultimately, the goal is to integrate this metric into MRI consoles to provide real-time feedback to technicians on whether a scan has reached the quality threshold required for tumor spread modeling.

Summary updated

Abstract

Diffusion-weighted MRI (DW-MRI) is used to quantitatively characterize the microscopic structure of soft tissue due to the anisotropic diffusion of water in muscle. Applications such as fiber tractography or modeling of tumor spread in soft tissue require precise detection of muscle fiber orientation, which is derived from the principal eigenvector of the diffusion tensor. For clinical applications, high image quality and high signal-to-noise ratio (SNR) of DW-MRI for fiber orientation must be balanced with an appropriate scan duration. Muscles with known structural heterogeneity, e.g. bipennate muscles such as the thigh rectus femoris, provide a natural quality benchmark to determine fiber orientation at different scan parameters. Here, we analyze DW-MR images of the thigh of a healthy volunteer at different SNRs and use PCA to identify subsets of voxels with different directions of diffusion tensor eigenvectors. We propose to use the mixing index of spatial co-localization of the clustered eigenvectors as a quality metric for fiber orientation detection. Comparing acquisitions at different SNRs, we find that high SNR results in a low mixing index, reflecting a clear separation of the two compartments of the bipennate muscle on either side of the central tendon. Because the mixing index allows joint estimation of spatial and directional noise in DW-MRI as a single parameter, it will allow future quantitative optimization of DW-MRI protocols for soft tissue.
BibTeX
@misc{shusharina2024anatomy,
  author = {Nadya Shusharina and Xiaofeng Liu and Evangelia Kaza and Miranda Lam and Stephan E. Maier and Jonghye Woo},
  title = {{Anatomy-based quality metric of diffusion-weighted MRI data for accurate derivation of muscle fiber orientation}},
  year = {2024},
  eprint = {2406.03560},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2406.03560},
}
Figure