increase in metabolic demand Note however that this explanation is not universally accepted.
2 m arTRAVASCUIAR )
6.2 EXPERIMENTAL METHOD
6.2.2 Functional MRI data collection
Functional MRI data were acquired on the 3T MRI scanner at the SPMMRC using an insert head gradient coil and whole head TEM RF volume coil (see chapter three). The sequence used was optimised for optimum BOLD contrast. Single shot MBEST gradient
echo (GE) EPI images were taken from a region of interest at the back of the head comprising 10 contiguous coronal slices. This region of interest covered the whole of the primary visual cortex and associated areas.
The use of rapid MR imaging to sample the haemodynamic response in both space and time requires careful selection of the imaging parameters in order to ensure an adequate temporal resolution, spatial resolution and signal to noise ratio. Optimisation of these three parameters introduces certain trade offs that must be considered prior to carrying out an fMRI experiment. In order to obtain good spatial resolution (i.e. a large matrix size and small voxels) long image acquisition times are required, which reduces the temporal resolution available. A reduction in voxel size will also lower the signal to noise ratio. Conversely, if a high temporal resolution is required, a small matrix and large voxels must be used in order to reduce the image acquisition time. However, whilst large voxels increase signal to noise ratio, a short volumar repetition time means that the longitudinal magnetisation does not have time to fully recover between images. This means that the initial longitudinal magnetisation available prior to each imaging pulse is reduced; hence the imaging signal is also reduced. Finally, if the voxel size is greater than the active region, then partial volume effects [3] will reduce the (BOLD) contrast to noise ratio. Given that the cerebral cortex is — 2 — 4 mm thick, this effect is highly likely.
For the visual BOLD experiment, the repetition time (TR) was set to 100 ms per slice,
giving an overall volumar repetition time, and hence temporal resolution, of 1 s. Using this TR, an imaging matrix size of 64 x 64 pixels per slice, with a slice thickness of 6 mm
and in plane resolution of 4 x 4 mm was achievable without placing too much stress on the gradient coils. The echo time was set to 40 ms in order to achieve sufficient BOLD contrast, but also maintain sufficient signal to noise ratio in the final images. The short repetition time also meant that the intensity recorded in the first volumar image was greater than that recorded in subsequent images. This Ti saturation effect is well
documented in fMRI literature and after a small number of volumes have been acquired, the system will reach a steady state. Those initial images were simply ignored. Following
steady state onset, the flip angle was optimised to give maximum intensity in the MR images.
6.2.3 Functional MRI data analysis
The recorded FMRI data were analysed using standard techniques. Analysis comprised pre-processing, motion correction, spatial smoothing, and statistical estimation. Full descriptions of each of these steps are given elsewhere [18], however for completeness, a brief description of each step is given below.
Data pre-processing comprised three simple steps. Initially the image data were
corrected for Nyquist ghost artefacts by manually changing the phase in alternate lines of k-space. The extent of ghost correction was gauged by eye and the phase angle was adjusted until the ghost was minimised. Following this, the
experimental pre-scans (saturation scans) were extracted to ensure that the effects of T1 recovery had reached a steady state. Finally, images were reduced in size
(segmented) eliminating the extra-cranial space. This step simply serves as a data reduction technique in order to speed up computational time.
One of the major detrimental effects to final functional image quality in fMRI is
subject motion. This can cause two types of artefact in fMRI time series: i) Suppose following the nth image, the subject's head either moves or rotates within the magnet. All subsequent images will be spatially displaced such that the time course of some voxel will result from two different brain regions. This effect can be corrected by motion correction algorithms (see for example [19]). ii) Perhaps less obvious is the fact that subject motion will destroy the steady state magnetisation that has built up over the duration prior to the nth image. This secondary effect is much harder to correct for using post-processing techniques. In order to minimise motion artefacts, motion plots from all five subjects were analysed. These plots, derived from the SPM99 motion correction algorithm, show 3 dimensional translation and rotation about the subjects' original position. Any dataset in which motion was larger than the smallest voxel dimension (i.e. 4 mm) was discarded in order to minimise steady state effects. The SPM99 motion
correction algorithm was then applied to the images (http://www.fil.ion.ucl.ac.uk/spm/).
A key conflict in the analysis of fMRI time series is that of efficiency against bias
[20]. Ideally one would like to maximise the efficiency in detecting genuine areas of activation, whilst preventing any bias in the final results that may arise due to data processing. This raises an important point since in general, the functional response is on a similar amplitude scale to the noise. In most cases, both spatial and temporal smoothing are applied to fMRI data in order to reduce the white noise level. However, whilst this will improve detection efficiency, if inappropriate filters are used they may bias the final results. (For example the use of a very large spatial filter may lead to an artificially large spatial extent of the final detected activation.) For the visual experiment, data were spatially smoothed using a 4 mm FWHM Gaussian smoothing kernel. This was thought to be sufficient to improve detection efficiency without biasing source extent. A low pass (Gaussian) temporal filter (FWHM = 4 s) was also applied to the time course data in order to remove high frequency (thermal) noise. In order to prevent low
frequency drift, a high pass temporal filter was applied with a 1501/z cut off
frequency.
Finally, statistical parametric maps of significant BOLD contrast were constructed using the General Linear Model. This is applied in a directly analogous way to that described in chapter four in relation to MEG data. The effect of interest in the GLM design matrix comprised a simple boxcar waveform defining the visual stimulus delivery. In order to account for the delayed haemodynamic response, this waveform was convolved with a standard haemodynamic response function in SPM99. In order to take into account any time latency in the BOLD response, the simple HRF model was supplemented with its temporal derivative. To test the significance of the final functional images, all BOLD statistical parametric maps were thresholded using a corrected p-value of 0.05. Co-registration of activation maps onto anatomical images was achieved using the AIR (Automated Image Registration) algorithm [19] in MEDx® (Sensor Systems) software.