Volunteer Motion; Realignment

For BOLD data, the necessity of motion correction has been well established (Hajnal et al., 1994; Friston et al., 1996). Image realignment has become a standard step in the BOLD data analysis and there are many software packages around that can do this, among which the well-known SPM package (SPM, 1999).

It is not hard to imagine the possible benefits of realignment for FAIR: if the head shifts > 1 voxel in a particular direction, the FAIR sequence could be measuring signal from white matter at one TI point and of grey matter for the next. A shift in and out of CSF areas will have even greater effects.

Standard realignment packages such as SPM(SPM, 1999) cannot deal with FAIR data for two reasons:

FAIR only covers 1-10 slices and not a whole brain volume.

- The FAIR data has different contrasts for different inversion times. Realignment

programs optimise for a minimal sum of squares difference between images which is not possible if the contrasts are different.

The first problem will become less and less important as the brain coverage of FAIR increases by using faster readout sequences. For the single-slice FAIR sequence

implemented in this project, however, this is an issue, because it is based on a single slice acquisition. It means all movement correction is limited to in-plane motions as there is no data to establish and correct out-of-plane movement. The second problem is more general and applies to all multi-TI acquisitions. Two-dimensional realignment for repeat

acquisitions o f one inversion time is not a problem, realignment over multiple inversion times is.

Three different analysis approaches were developed for a standard FAIR dataset: 1) no realignment

2) realignment of repeat acquisitions of every inversion point TI (‘intra-TI’ realignment); this means that averaged images for different inversion times TI might still be misaligned with respect to each other.

3) realignment of repeat acquisitions within and between inversion points (‘inter-TI’ realignment).

The inter-TI realignment was implemented as follows:

- For each inversion time TI, before the acquisition o f its nav repeats, a single shot, single

slice Mo image was acquired. These Mqsnapshots were realigned to the first acquired

Mo snapshot, thus giving the required in-plane image transformations at the beginning o f each TI point. These image rotations were applied to the first image of the FAIR data at each inversion point. The other repeats (nav-1) were then realigned to the first image o f each TI point series. This realignment scheme is illustrated below. By using the Mo snapshots as reference scans, realignment across different contrasts is thus possible.

(1)

realignm ent Mo snapshots to Mo sn a p sh o t-1

(2)

c o p y required im age rotations from Mo snapshot-x to ti* (l) im age.

1) Mo snapshot-1; 2) Mo snapshot-2; 3) Mo snapshot-3;

til : repeats |0....nav]

ti2 : repeats |@....nav]

tig : repeats 10....nav]

nti) Mo snapshot-nti; tinti : repeats |0....nav] (3)

realignm ent tix repeats t o t ix ( l)

nti is the number of TI points, nav is the number of repeats for each TI point.

The ‘intra-TF realignment is simply the realignment of the nav averages for every inversion point TI. This method thus does not realign between inversion points TI and the Mo snapshot data are therefore not used.

These realignment procedures were applied to a FAIR data set acquired with a slice thickness of 5 mm, a short tr of 2.65s (with global saturation in the pulse sequence) and a b-

factor of 5 s/mm^. This experimental setup will be described more extensively in the next section. Mo snapshots were also acquired immediately before the start o f every Tl-point acquisition. The standard deviation averaged over the slice and all TI points for FAIR data are given in Table 5-16. The standard deviation is given as a percentage o f that of

unrealigned data. Unrealigned intra-TI realignment inter-TI realignment sd FAIR over slice 100% 91% 91%

Table 5-16 E feet of realignment on summed sd of FAIR data: mean sd over inversion points and over all voxels (sdWZi sdj as percentage of unrealigned sd.

The realignment procedure is indeed reducing the variability of the FAIR data. The reduction in the total standard deviation is 9%, and there is no difference for intra-TI and inter-TI realignment.

To see if the data has been realigned correctly, the different FAIR data sets (unrealigned, intra-TI realigned, inter-realigned) were fitted using the CBFD model, with inputs from a TMF fit of SL data. The percentage of voxels passing the Q> 0.001 threshold for the goodness-of-fit test on the CBFD results was 88%, 90% and 89% for the

unrealigned, intra-realigned and inter-realigned data, respectively. So the realignment only has a small effect on the goodness-of-fit results. However, realigning repeats of one TI

point will reduce the mean value’s standard deviation and thus might decrease the chance

of a voxel passing the goodness-of-fit test. More information comes from looking at the sum o f squared differences between the data and the model. For the intra-realigned data, the sum o f squared differences is reduced by 13%, and by 10% for the inter-realigned data. So both intra- and inter- realignment improve the fit. The intra-realignment seems to be a little better.

As the realignment procedures are weighted towards edges, a risk with inter­

realignment is that the movements perpendicular to the slice can lead to a shift in the edges, which the algorithm tries to correct, thus potentially leading to an increase in the

realignment can also suffer from this problem, but as the time scale for the intra­ realignment is much shorter (18-36 acquisitions for one inversion point for intra­ realignment vs 6*18 or 6*36 acquisitions for inter-TI realignment) the out-of-plane movements will generally be smaller for intra-realignment.

Because the intra- and inter-realignment both result in comparable improvements and the intra-realignment is less sensitive to out-of-plane movement, the intra-TI

realignment will be used in this thesis. Inter-realignment will probably be the method of choice once whole brain FAIR data sets become available, as this method can reduce variability between Tl-points, i.e. between subsequent CBF measurements.

slice after out-of-plane m ovem ent slice o f interest