Effects of Tissue-Specific Smoothing Approaches on Statistical Analysis in Quantitative MRI

Quantitative MRI (qMRI) provides interpretable maps of brain microstructural properties, such as myelination and iron content. However, standard Gaussian smoothing can compromise interpretability due to partial volume effects, particularly at grey matter (GM) and white matter (WM) boundaries. To address this, we compared three tissue-specific smoothing (TSmoo) approaches: TSmoo-LC (linear and compensated), TSmoo-NC (non-linear and compensated) and TSmoo-NI (Non-linear, Intensity-weighted).
Using open-access qMRI data from 138 healthy participants (19–75 years), we reproduced Callaghan et al. (2014) to evaluate age-related changes. TSmoo-LC and TSmoo-NC yielded similar spatial patterns of age-related iron content increases and myelination decreases, but TSmoo-LC detected more significant voxels and clusters. In contrast, TSmoo-NI showed fewer significant voxels and clusters, reflecting lower effective smoothing and higher resel counts. Voxelwise log-likelihood (LL) maps revealed that TSmoo-NI provided a better model fit at GM-WM boundaries, while TSmoo-LC excelled in cortical GM and TSmoo-NC in deep WM.
These findings highlight the impact of smoothing strategies on statistical sensitivity and anatomical specificity. TSmoo-LC and TSmoo-NC are suitable for maximizing detection power, with TSmoo-LC better capturing small variations within cortical gray matter and TSmoo-NC performing better in the homogeneous core of deep white matter. TSmoo-NI preserves anatomical detail, particularly at tissue boundaries.
This study underscores the importance of selecting a TSmoo method based on the scientific objective: maximizing sensitivity or preserving anatomical specificity. This work provides insights into optimizing qMRI analyses for aging and neurodegenerative research.