next up previous
Next: Null Data Study Up: Accuracy Assessment: Motion Correction Previous: Interpolation Scheme

Choice of Template Image

In an attempt to increase the accuracy of the scheme and as a final parameter investigation, a method using a mean image template was implemented. This scheme generates a mean image for the series by averaging all the volumes over time after the first three stages of trilinear interpolation-based motion correction have been carried out. In doing so we hope to be registering all volumes to a more generalised target which exhibits less overall variation from each volume in the series than the original target (middle) volume previously used. This new mean image is a robust target to which the original time series is then registered, again using three trilinear interpolation stages and an optional final sinc interpolation stage. Because we are registering to a mean image, we no longer have ``gold standard'' values for the transformations found by the correction scheme. Therefore, to quantify the accuracy of the correction, a median absolute residual variation (MARV) score was created by initially de-meaning each voxel time-series and then measuring the median value of the residual absolute values in this time-series. That is,

$\displaystyle MARV(x,y,z) = \sum_{t=1}^N \vert I_t(x,y,z) - I_{mean}(x,y,z) \vert / N. $

This produce a volume of MARV scores for each voxel and the median of these values (over the volume) is then taken as a summary measure. This is effectively a measure characterising the level of inter-volume intensity variation (presumed to be due to subject motion) after retrospective motion correction has been applied. While this can only work for activation-free data (so that in perfect alignment the variance should be at minimum) it can give us a clear impression of the accuracy of the motion correction scheme. Because SPM rejects information outside a mask obtained from the data (end slice effects), the corrected median images were masked according to the corrected SPM data so that the measure reflected a consistent comparison across the schemes. The results shown in Figure 15 correspond to the MARV values generated after running MCFLIRT and SPM on the null-activation data set for both the low and severe motion designs. Results using the RMS measure (Table 3) revealed that, although all three schemes provide sub-voxel accuracy, AIR 3.08 using Least Squares (which we found to give better results than the standard AIR measure) and Windowed Sinc interpolation was almost an order of magnitude worse than basic 3 stage tri-linear MCFLIRT. Accordingly we decided not to compare it further. However, we note that AIR was primarily designed to solve several different registration problems that arise in tomographic data sets [18] rather than optimised for FMRI motion correction.

Table 3: RMS deviation values for synthetic null data
Uncorrected AIR SPM MCFLIRT
Sum of squared intensity errors 936.5866 406.8876 1.6405 1.5171
RMS error (mm) 2.3360 1.7570 0.1064 0.1102


Figure 15: Median Absolute Residual Variation (MARV) values for corrected data processed by different motion correction schemes: Uncorrected, MCFLIRT w. CR, MCFLIRT w. NC, MCFLIRT w. CR & mean, MCFLIRT w. NC & mean, SPM99 run at full quality, with sinc interpolation and interpolation error adjustment
\begin{figure*}
\begin{center}
\psfig{figure=mare_mean.eps,width=2\figwidth}
\end{center}
\end{figure*}

>From these results we conclude that for some cases (generally the low motion data), MCFLIRT with the Correlation Ratio cost function produces significantly smaller errors than SPM99 whilst in other cases (some of the high motion data) both methods give similar results. This can be seen be comparing the heights (MARV values) of the SPM bars with the CR bars (typically the best MCFLIRT cost function) where a 30% to 40% reduction can be seen in the first, second and fourth cases. Slightly surprisingly we found that the use of a mean image template gave no discernable improvement in accuracy. We conclude that for artificial data where the motion is purely rigid, there is no advantage to using a (possible blurred) average image over an image from the original data. We would expect that the mean template scheme could yield greater accuracy where the data includes some physical motion-induced artefacts and the choice of a reference image from the original dataset is not so obvious.
next up previous
Next: Null Data Study Up: Accuracy Assessment: Motion Correction Previous: Interpolation Scheme
Peter Bannister 2002-05-03