Abstract
Diffusion MRI affords valuable insights into white matter microstructures, but suffers from low signal-to-noise ratio (SNR), especially at high diffusion weighting (i.e., b-value). To avoid time-intensive repeated acquisition, post-processing algorithms are often used to reduce noise. Among existing methods, non-local means (NLM) has been shown to be particularly effective. However, most NLM algorithms for diffusion MRI focus on patch matching in the spatial domain (i.e., x-space) and disregard the fact that the data live in a combined 6D space covering both spatial domain and diffusion wavevector domain (i.e., q-space). This drawback leads to inaccurate patch matching in curved white matter structures and hence the inability to effectively use recurrent information for noise reduction. The goal of this paper is to overcome this limitation by extending NLM to the joint x-q space. Specifically, we define for each point in the x-q space a spherical patch from which we extract rotation-invariant features for patch matching. The ability to perform patch matching across q-samples allows patches from differentially orientated structures to be used for effective noise removal. Extensive experiments on synthetic, repeated acquisition, and real data demonstrate that our method outperforms state-of-the-art methods, both qualitatively and quantitatively.
Introduction
Diffusion Magnetic Resonance Imaging (dMRI) is a non-invasive technique essential for mapping the complex microstructures of the human brain. By measuring the diffusion of water molecules, researchers can infer the orientation and integrity of white matter tracts. However, dMRI is inherently limited by a low signal-to-noise ratio (SNR), a problem that is exacerbated when using high b-values to resolve fine structural details. While increasing scan time through repeated acquisitions can improve SNR, it is often impractical in clinical settings.
Post-processing denoising techniques, particularly Non-Local Means (NLM), have emerged as a solution. NLM works by identifying similar image patches and averaging them to suppress noise while preserving structural edges. Traditionally, NLM operates in the 3D spatial domain ($x$-space). However, this approach struggles with the brain's curved geometry. When a fiber bundle changes direction, its local signal profile changes, making it difficult for spatial-only algorithms to find matching patches—a phenomenon known as the "rare patch effect." This study introduces XQ-NLM, a framework that extends denoising into a joint 6D space ($x-q$ space) to leverage the full self-similarity of diffusion data.
Methodology: The XQ-NLM Framework
The core innovation of XQ-NLM is the transition from searching for similarity in 3D spatial coordinates to searching across the combined spatial and diffusion wavevector ($q$) domains. By treating the dMRI data as a 6D volume, the algorithm can identify self-similar patterns that are hidden when looking at spatial slices alone.
Spherical Patch Matching and Rotation Invariance
To effectively match patches in $q$-space, the researchers defined a spherical patch for every point in the 6D space. Because white matter structures can be oriented in any direction, a direct comparison of signal intensities is insufficient. To achieve rotation invariance, the method employs two key transforms:
- Azimuthal Equidistant Projection (AEP): This projects the spherical patch onto a 2D disc.
- Polar Complex Exponential Transform (PCET): This extracts features from the 2D disc that remain constant regardless of the structure's orientation.
This allows the algorithm to recognize that a horizontal fiber bundle and a vertical fiber bundle of the same tissue type are fundamentally similar, significantly increasing the pool of available patches for averaging.
Noise Adaptation
MRI noise typically follows Rician or non-central Chi distributions rather than a standard Gaussian distribution. To ensure mathematical accuracy, XQ-NLM includes a preprocessing step that transforms the signal to follow a Gaussian distribution, allowing for more precise noise estimation and removal.
Experimental Validation
The performance of XQ-NLM was evaluated against several state-of-the-art methods, including Adaptive NLM (ANLM), NLSAM, and MPPCA, using three distinct datasets:
- Synthetic Data: Created with the Phantomαs toolbox, these simulations provided a ground truth for complex geometries like branching and crossing fibers under various noise conditions.
- Repeated Acquisition Data: A high-SNR reference was established by averaging 25 separate scans of a single subject, providing a "gold standard" for real-world performance testing.
- Human Connectome Project (HCP) Data: Real-world multi-shell data were used to test the algorithm's efficacy on high-quality, complex clinical samples.
Results and Discussion
Quantitative and Qualitative Superiority
XQ-NLM demonstrated significant improvements over existing benchmarks. In synthetic tests featuring 10% Rician noise, XQ-NLM achieved a Peak Signal-to-Noise Ratio (PSNR) improvement of up to 6.42 dB over the next best method (NLSAM).
Qualitatively, the method excelled at preserving the physical properties of the tissue. It produced more accurate Fractional Anisotropy (FA) maps and Orientation Distribution Functions (ODFs). Notably, XQ-NLM successfully eliminated "spurious peaks"—false fiber directions that often appear as artifacts at tissue boundaries in other denoising methods. The algorithm also demonstrated superior edge preservation, removing the blurring and "halo" effects common in traditional NLM applications.
The Advantage of 6D Information
The study highlights that moving to a 6D space dramatically increases the "effective sample size" for denoising. Because $q$-space signal profiles are generally smoother than spatial intensities, the algorithm can find more reliable matches. This increased redundancy allows for a tighter matching criterion, which reduces estimation bias and leads to a cleaner final image.
Conclusion
XQ-NLM represents a significant advancement in dMRI post-processing by recognizing that diffusion data is inherently multidimensional. By utilizing rotation-invariant features in a joint $x-q$ space, the method overcomes the limitations of spatial-only denoising, particularly in the brain's complex, curved white matter tracts.
While the 6D search space increases computational complexity, the gains in SNR and structural accuracy are substantial. Future work may extend this $q$-space matching strategy to other critical dMRI tasks, such as image registration, interpolation, and the construction of high-resolution brain atlases.
Summary updated