Marina Rakic^{1}, Luis Miguel Lacerda^{1}, Ahmad Beyh^{2}, Pedro Luque-Laguna^{1}, Rachel Barrett^{2}, Francisco De Santiago Requejo^{1}, Steven Williams^{1}, Gareth Barker^{1}, Fernando Zelaya^{1}, and Flavio Dell'Acqua^{1,2}

A well-known dilemma in DW-MRI acquisitions is to determine the extent to which signal-to-noise ratio (SNR) can be can be sacrificed in favour of higher spatial resolution on one hand, and in favour of shorter acquisition time on the other. In this study we quantify the reproducibility of spherical deconvolution results at 3 spatial resolutions with and without denoising, as it is still unclear how denoising methods6 affect the uncertainty in subsequent diffusion model fitting and whether it introduces or improves bias in modelled fibre direction.

METHODS

Two datasets were acquired with multi-band DW-MRI (MB=3) on a 3T MR750 GE scanner: dataset A contains 360(4x90) diffusion-weighted volumes and dataset B contains 180(2x90), acquired at an isotropic resolution of 2mm, TE=72.0ms and b=2000s/mm2. Dataset B was also acquired at 1.7mm and 1.5mm isotropic resolutions, with TE=73.9ms and TE=88.8ms, respectively. Data was pre-processed using eddy and topup (FSL software
JSD in Figure 1 is
an index of uncertainty of the overall FOD profile and it shows that higher
spatial resolution noticeably affects uncertainty of FODs in the 90 directions
dataset where higher JSD indicates lower reproducibility. JSD maps exhibit
spatial patterns similar to well known SNR maps^{9} of accelerated acquisition
sequences, since the variation in each FOD component is highly sensitive to
noise, which is amplified towards the centre of the brain, i.e. further away
from the receive coil. JSD maps show the reproducibility is efficiently
recovered by denoising the 90 DW directions data (row 2) and is comparable to
180 DW volumes (row 3). This is confirmed by JSD histogram distribution plots
(Figure 1), where the proportion of voxels in lower JSD ranges is higher in
denoised data, with most evident effect of denoising in 1.7mm spatial
resolution.

In Figure 2, CI maps show increased angular uncertainty with higher spatial resolution. This effect is more prominent for 90 directions compared to 180. Denoising 90 directions dataset results moderately improves angular confidence interval of the 1st and 2nd peak. Unlike JSD maps, CI maps appear to follow the pattern of white matter anisotropy suggesting the peak orientation angular error is affected by noise as well as anatomical complexity. This could explain why the effects of denoising are more prominent in reducing the shape uncertainty, and less obvious in peak direction angular uncertainty.

Both, JSD and CI, show the uncertainty of 90 DW-volumes dataset is brought closer to the 180-directions dataset by MPPCA denoising method. Finally Figure 3 shows the angular bias of 90 directions compared to reference dataset of 360 directions (i.e. more reliable) remains unchanged before and after denoising. Tying these results together suggests the uncertainty in diffusion model fitting is efficiently reduced with results comparable to doubling the number of data, without systematically affecting the angular bias.

(A)
Jensen-Shannon Divergence maps of
600 bootstrap fibre orientation distributions per voxel. (B) The histogram
distribution plots of JSD values with bin size of 0.001 for three different
isotropic spatial resolutions. (C) An example of FOD variability with the
transparent blue cloud around the zoomed in FOD represents the standard
deviation of FOD components.

Angular
90% confidence interval of the 1st (B) and 2nd (C)
highest local peak in fibre orientation distribution function for the slice
location shown in (A), based on local maxima extraction from 600 bootstrap
FODs.

The
angular difference maps, for the slice location shown on the far left, of 90
DW volumes (directions) dataset compared to 360 DW volumes for non-denoised
data (left) and denoised data (right).