Abstract
Diffusion-weighted MRI (dMRI) and its derivative models provide critical insights into human brain organization and white matter variations. While previous research has demonstrated that complex graph measures of the connectome vary with spatial sampling, the specific biases introduced by anisotropic voxels—common in clinical settings—remain under-characterized. This study evaluates the impact of anisotropic sampling on microstructural measures (fractional anisotropy and mean diffusivity) and white matter bundle properties (volume, length, and surface area). Using scan-rescan data from 44 subjects in the Human Connectome Project Young Adult (HCP-YA) dataset, we compared three white matter bundles across various spatial resolutions. Our results indicate statistically significant differences in at least one microstructural and one bundle measure at every tested resolution. While upsampling to a 1mm isotropic grid improves the repeatability of tractography-derived bundle measures, it fails to recover original microstructural values, suggesting that the biases introduced by anisotropic sampling are irreversible for tissue physics metrics.
Introduction
In neuroimaging, the "voxel" serves as the fundamental unit of 3D spatial sampling. Ideally, these voxels are isotropic (perfect cubes), ensuring uniform spatial resolution in all directions. However, due to hardware limitations, signal-to-noise ratio (SNR) requirements, or the need for rapid clinical acquisition, dMRI data is frequently acquired with anisotropic voxels—rectangular prisms often elongated in one direction.
Elongated voxels introduce Partial Volume Effects (PVE), where a single voxel captures a mixture of different tissue types or fiber orientations, leading to blurred or biased data. While the impact of sampling on global connectomics is documented, there is a gap in understanding how these choices affect specific white matter bundles and the quantitative microstructural metrics used to diagnose disease. This study investigates whether the biases introduced by anisotropic sampling can be mitigated by mathematically upsampling data back into a high-resolution isotropic grid.
Methods
Data and Simulation Strategy
The researchers utilized the Human Connectome Project – Young Adult (HCP-YA) dataset, focusing on "scan-rescan" data from 44 healthy subjects. This allowed the team to distinguish between the effects of spatial sampling and the natural baseline variability of repeated scans.
To evaluate the impact of sampling, a "downsample-then-upsample" simulation was employed:
- Baseline: High-quality 1.25mm isotropic data served as the gold standard.
- Degradation: Lower-quality data was simulated by incrementally stretching voxels in the z-direction (superior-inferior) in 0.25mm steps, reaching a maximum of 2.75mm.
- Restoration: These anisotropic images were upsampled back to 1.25mm and 1mm isotropic grids using cubic interpolation.
Analysis Metrics
The study focused on three major white matter pathways: the Arcuate Fasciculus, the Corticospinal Tract (CST), and the Corpus Callosum (CC). Two categories of metrics were analyzed:
- Microstructure: Fractional Anisotropy (FA) and Mean Diffusivity (MD), which quantify the magnitude and directionality of water diffusion.
- Macrostructure (Tractography): Bundle volume, length, and surface area.
Tensor modeling was performed using MRTrix3, and automated bundle segmentation was conducted via TractSeg.
Results and Discussion
The Fragility of Microstructure
Quantitative maps (FA and MD) proved extremely sensitive to the initial sampling grid. The study found that once data is captured in an elongated voxel, the underlying physics is fundamentally altered. Every resolution change, even a minor 0.25mm increment, resulted in statistically significant differences ($p \le 0.05$) compared to the baseline. Cohen’s $d$ coefficients revealed "large" effect sizes ($> 0.8$) for microstructural changes at every stage of downsampling. Crucially, upsampling to a cubic grid did not restore the original contrast; it merely smoothed the errors introduced by the initial anisotropic sampling.
The Resilience of Tractography
In contrast to microstructural metrics, the macro-scale properties of white matter bundles demonstrated a degree of resilience. When data was resampled to a 1mm isotropic grid, bundle measures became more repeatable. This suggests that fiber-tracking algorithms, which rely on Fiber Orientation Distributions (FODs), benefit from the higher-resolution grid's smoothness, even if the underlying data quality is degraded.
However, this resilience was not uniform across the brain. The orientation of the anatomy relative to the voxel "stretch" played a critical role. The Corpus Callosum, which primarily runs left-to-right, was more negatively impacted by vertical (z-direction) stretching than the Corticospinal Tract, which runs superior-inferiorly.
Conclusion
This study demonstrates that while upsampling low-resolution dMRI data to a 1mm isotropic grid can improve the consistency of tractography and bundle identification, it cannot recover accurate microstructural measurements. For researchers and clinicians, the takeaway is twofold: first, the orientation of anisotropic voxels relative to the anatomy of interest must be carefully considered; second, while post-processing can stabilize fiber tracking, it cannot compensate for the loss of physical tissue information inherent in low-resolution sampling. Future work should explore whether advanced super-resolution or deep-learning reconstruction methods can better bridge this gap where standard interpolation fails.
Summary updated