Harmonized connectome resampling for variance in voxel sizes

Elyssa M. McMaster, Nancy R. Newlin, Gaurav Rudravaram, Adam M Saunders, Aravind R. Krishnan, Lucas W. Remedios, Michael E. Kim, Hanliang Xu, Jai Hoon Yoon, Derek B. Archer, Kurt G. Schilling, François Rheault, Laurie E. Cutting, Bennett A. LandmanMagnetic 2025 · DOI

TL;DR

This paper investigates how different voxel resolutions in diffusion MRI (dMRI) affect brain connectome analysis. They found that down-sampling dMRI data significantly alters connectome measures, but up-sampling to a common 1mm isotropic resolution can harmonize these results, leading to more reproducible and reliable brain connectivity maps.

What Problem Does This Paper Address?

Understanding the intricate wiring of the human brain, known as the connectome, is crucial for studying neurological disorders and cognitive functions. Diffusion MRI (dMRI) is a key technology for mapping these connections through fiber tractography. However, a major challenge is the variability introduced by different dMRI acquisition resolutions. This inconsistency makes it difficult to compare findings across studies, hindering the development of standardized diagnostic tools and research protocols. This paper addresses the lack of a comprehensive study that systematically quantifies the impact of voxel resolution on connectome analysis and proposes a harmonization strategy.

Methods

The researchers employed a systematic approach to evaluate the effect of voxel resolution. They started with high-resolution dMRI data and then generated multiple datasets by down-sampling and up-sampling the voxel sizes in 0.5 mm increments, creating both isotropic and anisotropic resolutions. For each resolution, they generated tractograms (representing fiber pathways) using the MRTrix3 software and then constructed connectomes (graph representations of brain connectivity) using the Desikan-Killany atlas. Finally, they analyzed these connectomes using 12 complex graph measures. Statistical significance of differences between resolutions was assessed using the Wilcoxon Sign-Rank test, and effect sizes were quantified using Cohen's d. They also performed qualitative visualizations of tractography bundles to assess anatomical plausibility.

Datasets

The primary dataset used in this study was from the Human Connectome Project – Young Adult (HCP-YA) dataset, which includes scan/rescan data from 44 healthy subjects (32 female). This dataset provided high-resolution dMRI and T1-weighted images. For translational impact analysis, data from a subset of HCP-YA subjects was resampled to match the resolutions of three major national studies: the Alzheimer's Disease Neuroimaging Initiative (ADNI), the Baltimore Longitudinal Study of Aging (BLSA), and a common clinical isotropic resolution of 2 mm.

Key Insights and Ideas

  • Voxel resolution is a critical factor: The study unequivocally demonstrates that even small changes in voxel resolution (as little as 0.5 mm) can lead to statistically significant differences in complex graph measures of brain connectomes.
  • Down-sampling introduces variability: Reducing voxel resolution leads to less reproducible and less reliable connectome measures, likely due to increased partial volume effects and the inclusion of multiple tissue types within a single voxel.
  • Up-sampling as a harmonization strategy: The most significant finding is that up-sampling lower-resolution data to a common 1 mm isotropic resolution can effectively mitigate the spatial sampling effect and harmonize connectome measures across different acquisition protocols.
  • Algorithmic vs. Information Loss: The observed differences between resolutions are more indicative of algorithmic challenges in tractography at lower resolutions rather than simple information loss from down-sampling.

Key Results

  • Nearly all pairwise comparisons between different resolutions yielded statistically significant differences in graph measures (p ≤ 0.05).
  • When up-sampling down-sampled resolutions in 0.5 mm increments, a mitigation of the spatial sampling effect was observed on both tractography and connectome graph measures.
  • For translational impact analysis, Cohen's d effect size was consistently large (≥ 1) when comparing study resolutions to 1 mm isotropic resolution data, highlighting the impact of resolution differences.
  • Qualitative visualizations showed that lower resolutions resulted in more spurious streamlines and less anatomically plausible tractography bundles compared to higher resolutions.

Limitations and Future Work

The authors acknowledge several limitations. The study primarily used HCP-YA data, which may not fully represent individuals with white matter diseases or disorders, or across different age groups. The dataset also had a bias towards female scans and did not account for site-specific differences or acquisition parameter variations beyond resolution. The choice of the Desikan-Killany atlas for parcellation might also introduce biases.

Future work could explore the effect of spatial sampling on diverse populations and investigate finer parcellation schemes. The authors suggest that resampling diffusion data to 1 mm isotropic resolution before performing DTI and tractography is a recommended strategy to minimize spatial sampling effects and improve reproducibility. They also note that while up-sampling can harmonize results, the degree of improvement varies across different graph measures, with "clustering" presenting a unique challenge.

Summary updated

Abstract

Diffusion MRI (dMRI) fiber tractography presents exciting opportunities to deepen our knowledge of human brain connectivity and discover novel alterations in white matter. To date, there has been no comprehensive study characterizing the effect of dMRI voxel resolution on the resulting connectome for subject data. We assessed the statistical significance of graph measures derived from dMRI data by comparing connectomes from the same scans across different resolutions with 44 subjects (32 female) from the Human Connectome Project - Young Adult dataset (HCP-YA) with scan/rescan data (88 scans). We explored 15 isotropic and anisotropic resolutions, generated tractography and connectomes, and compared graph measures between each resolution and its nearest larger and smaller resolutions. Nearly all pairwise comparisons yielded statistically significant differences in graph measures (p ≤ 0.05, Wilcoxon Sign-Rank Test). Upon up sampling the 14 down sampled resolutions in 0.5 mm increments, we observed mitigation of the spatial sampling effect on both the tractography and the connectome's complex graph measures. To investigate translational impact, we resampled 22 subjects from HCP-YA to the resolutions of two major national studies and up-sampled this data back to 1 mm isotropic with different interpolation methods. Similarity in results improved with higher resolution, even after initial down-sampling. To ensure robust tractography and connectomes, resample data to 1 mm isotropic resolution.
BibTeX
@inproceedings{mcmaster2025harmonized,
  author = {Elyssa M. McMaster and Nancy R. Newlin and Gaurav Rudravaram and Adam M Saunders and Aravind R. Krishnan and Lucas W. Remedios and Michael E. Kim and Hanliang Xu and Jai Hoon Yoon and Derek B. Archer and Kurt G. Schilling and François Rheault and Laurie E. Cutting and Bennett A. Landman},
  title = {{Harmonized connectome resampling for variance in voxel sizes}},
  booktitle = {Magnetic},
  year = {2025},
  doi = {10.1016/j.mri.2025.110424},
}
Figure