Synopsis

Keywords: Data Analysis, Aging

In this explorative study, the effect of imaging parameters, image quality, circadian rhythm, and aging on FreeSurfer’s estimates is investigated. To that end, the human phantom, an openly available data of a single subject scanned at 7T over 7 years with various MR protocols, is used.

Introduction

Freesurfer^1,2 is one of the most frequently applied tools in neuroimaging to quantify brain morphology such as variability in cortical thickness and regional brain volume due to aging³. Further, imaging parameters such as voxel volume can be a bias in the estimates⁴. In contrast to most studies which utilize harmonized protocols and large cohorts, here, an openly available dataset of a single subject scanned over approximately 7 years with various MR protocols (i.e. voxel volume: 0.1 mm³-1 mm³) is used. In this preliminary study, a broad, explorative analysis is conducted to investigate the effect of imaging parameters, image quality, circadian rhythm, and aging on FreeSurfer’s estimates.

Methods

Form the human phantom^5,6, 75 MPRAGE volumes of the single healthy male (29-year-old at study start) meeting the following criteria were selected: 32-channel head coil; maximal GRAPPA factor 2 (47x no GRAPPA, 28x GRAPPA 2, 1x excluded), maximal 7/8 partial Fourier (53x no partial Fourier; 22x 7/8; 6x excluded). Along with the image data, the FreeSurfer (v6 with conf2hires; recon-all script including resampling to 1mm isotropic resolution) and MRIQC⁷ results are provided readily in the data repository. For all 75 volumes gray to white matter contrast-to-noise ratio (CNR) was > 1.5 and the EFC⁸ < 0.75. For correlation and modeling the following input variables were selected empirically: voxel volume, partial Fourier, GRAPPA factor, CNR, year (starting from 0 at the first scan), and time of day. Significance of Pearson’s correlations were Bonferroni corrected. Linear models were used to understand how the combination of input variables explained the variance in FreeSurfer’s global and regional estimates. As a baseline, only the voxel volume is considered. Successively, more variables are added to the model: partial Fourier, GRAPPA factor, CNR, year, and time of day. Hence a total of six models are compared using adjusted R² values (corrects R² for number of variables used). The variable order was motivated by first considering imaging parameters, then the image quality, and lastly, aging and the circadian rhythm. This comparison provides a glimpse into the relevance of each variable in explaining the variance in the data while keeping the number of models investigated reasonably small.

Results

Voxel volume is distributed bimodally (see Fig.1). CNR increases with voxel volume, with a significant correlation. While data was acquired over approximately 7 years and between 8:00 to 22:00, the majority was acquired mid-day and in year 5 and 6. One-by-one Pearson correlation (Bonferroni corrected; see Fig.2) yields significant correlation for estimated total intracranial volume (eTIV) with voxel volume and CNR; brain volume with partial Fourier; total and cortical gray matter (GM) volume with years, respectively. No significant correlation with aging was found for subcortical GM volume. White matter (WM) volume correlates significantly with all imaging parameters and CNR. Surprisingly, ventricle volume decreased with aging, but the data was not normalized for changes in eTIV or global brain volume. To understand how the model variables jointly explain the variance in Freesurfer estimates, the adjusted R² is reported (see Fig.3 for global and Fig.4 for regional estimates, respectively). Further, average increase in adjusted R² across ROIs for all models are shown in Tab.1. The absolute and increase in adjusted R² was region-specific and different for volume, area, and thickness estimates. Further, comparing ROIs across hemispheres showed stark differences. Modeling variables reflecting the effective resolution, i.e. voxel volume and partial Fourier, explains largely variances in eTVI, WM volume, and on average for regional volume and thickness, but not surface estimates. Including the GRAPPA factor did not increase R² considerably. On average, adding CNR showed only for ROI volume a noteworthy increase. Aging, modeled by years, had a considerable contribution for all estimates on average, particularly cortical volumes. On average, time of day, modeling the circadian rhythm, increased considerably R² for all regional metrics but the average thickness.

Discussion

Using data acquired at 7T over 7 years, an explorative assessment of imaging parameters, quality, aging, and circadian rhythm on Freesurfer estimates for a single subject was conducted. Modeling the effective resolution, i.e. voxel volume and partial Fourier, had a considerable contribution to the data variance potentially due to image blurring. Time-related variables, i.e. modeling aging and circadian rhythm, had considerable, but ROI-specific effects. In the future, this initial, explorative analysis needs to be refined. First, further FreeSurfer options, i.e. longitudinal assessment, have to be considered. Secondly, more advanced models need to be employed. The here used linear assumption is suboptimal for i.e. the periodic circadian rhythm. Further, the increase in R² when adding a variable to the modeling is no direct measure of the contribution of the respective variable in isolation, but the overall model performance. Third, acquiring new data to better capture the circadian rhythm in within a day instead of pooling it from approximately 7 years would be ideal. Finally, adding further variables (i.e. hydration level) could improve the overall low to moderate R² values.

Acknowledgements

This work was funded by the DFG-MA 9235/1-1 and MA 9235/3-1