Figure VI.1: An example of a PET scan
In this chapter the problem of matching PET images with MEG/EEG/MRI images is addressed. A Positron Emission Tomography (PET) scanner is capable of registering functional brain activity (cf. Lammertsma, 1992). Therefore it is interesting to compare the information from PET scans with the information derived from MEG/EEG/MRI combinations. A comparison of the localization results obtained with these various modalities may help to find those areas which are active with a higher degree of confidence. PET may, for instance, be able to indicate that the number of dipoles taken in MEG/EEG inverse procedures is adequate. However, it should be kept in mind, that each modality probes different properties of the brain which may bear a rather complex relationship.
The time resolution of PET scans is much lower than that of MEG and EEG, the minimum scan time being about 1 second. However, in order to perform useful measurements usually a scan time of at least 20 seconds is required. To obtain a reliable PET image may require several seconds to several minutes. The voxels are approximately 3.5 x 3.5 mm, and the slice thickness is about 3 mm. The spatial resolution is several mm. The high noise content of the images makes analysis of PET scans difficult for single individuals. Often averages are taken over several subjects. The PET studies discussed in this chapter have been carried out at the PET research unit of the University Hospital at Groningen.
PET is a technique that produces images of the distribution of a previously
administered radionucleide. The radioactive tracer is inhaled or injected
into the vascular system and thus transported to the brain. Consequently,
due to radioactive decay positrons are emitted in all directions within the
brain. A positron is annihilated within a very short distance from the
point of emission when meeting an electron. At the moment of annihilation,
two gamma particles are emitted, traveling in opposite directions. The two
gamma particles are detected as coincidence events by appropriately placed
scintillation detectors. The time difference needed for the gamma
particles to reach the detectors is sometimes used to calculate the
location where the annihilation took place, but usually only the direction
of the particles is taken into account. The locations found are not
exactly the same locations as where the radioactive material emitted the
positron, but the difference in location is less than 3 mm. Averaging the
data reduces this distance considerably.
In order to measure the brain activity of a subject, he or she is placed in the
PET scanner, which consists of a number of rings of detectors, in our case
16. Each ring scans one slice. Additional slices are interpolated
from the measurements. A typical PET apparatus can scan 31 slices of 3 mm
thickness.
In order to increase the accuracy of the reconstruction of the location where annihilation of the positrons takes place, the attenuation factors of the various tissues within the body can be calculated before the actual experiment is carried out by performing a transmission scan. The subject's head is exposed to radioactive sources, and the detectors measure the amount of radioactivity received. A transmission scan can be compared to a low resolution CT-scan. Since the source is known, the attenuation factors for radiation of the various tissue types can be calculated from these measurements .
A number of positron-emitting tracers can be used in PET-recordings. A [15O] labeled compound can be used to measure the regional cerebral blood flow (rCBF). Because the measurement time for rCBF with PET is short, a [15O]-labeled compound is an appropriate choice for activation studies. In Groningen [15O]-Water is used. In general, in activation studies paired images of rCBF are obtained, one image in a control state and one during the presentation of a stimulus. The neurophysiological activity associated with the activated state is determined by subtracting the image of the control state from that of the stimulated state. Regions with increased activity are the areas with increaced rCBF. It can be assumed that the increased blood flow is a consequence of brain activity. After injection of the tracer, the scanner counts the gamma quanta, indicating the emitted positrons over a period of time, called a frame. Different frames can be taken one after another. in this way changes in the activity over a time lapse can be recorded to obtain a reasonable signal-to-noise ratio. The time resolution of a PET scan taken to measure rCBF is a few seconds. This means that even using the shortest time resolution possible, the measured activity is the integral of the total brain activity over a few seconds.
Figure VI.1: An example of a PET scan
An example of a PET scan is shown in figure VI.1. As the PET scan has its own coordinate system and its voxels do not have the same dimensions as the voxels in the IVIRI, and because there are few obvious reference points common to the two images, combining the two is not a trivial procedure. A number of different approaches have been described in the literature. Two procedures are followed; the first one uses markers, which is in theory a straight-forward procedure giving unambiguous results, but it assumes that both the PET and the MRI scans have to be performed with markers on the subject's head (cf. Eisen et al., 1991). However, the region of the head that is scanned with the PET scanner is usually smaller than that in the MRI, so that it is not sufficient to place the markers on the usual anatomical landmarks, e.g. the auricular points and the nasion. To obtain the desired information additional PET scans have to be performed where the bed on which the subject lies is moved over a known distance.
The second method uses feature matching (Borgefors, 1988; Valentino et al., 1991; Besl and McKay, 1992; Eisen et al., 1992; Walter et al., 1992), where features of the head that are obtainable from both PET and MRI are matched. For example, such features are the outer surface of the head, the skull or even the contour of the ventricles in the brain. A frequently used method is to locate points on the outer surface of the head from the MRI scan and the PET transmission scan. Both sets of points are then matched together (Arun et al., 1987; Levin et al., 1988; Neiw, 1991; Pelizzari et al, 1989; Walter et al., 1992). This is the basis for our matching procedure.
In order to use the feature matching procedure to combine the MRI with the PET images, points on the outer surface of the head are taken from both types of scans. In order to retrieve the surface points from the MRI we use the same method which is employed to generate a realistically shaped model of the head, that was described in chapter IV of this thesis. In order to be able to extract surface points from the PET scan, the transmission scan has to be analysed. The transmission scan shows the outline of the head as well as the headrest in each slice. A basic threshold operation can separate the background from the head and the headrest. An erosion separates the head from the headrest, just in case pixels from the head connect with pixels from the headrest. Once again, to find all the pixels that are part of the head, a region growing operation is started. One point at the front of the head is used as seed, somewhere at the midline of the slice.
Once all the pixels that are part of the head have been determined, a dilation is used to make up for the pixels lost in the erosion. The mass midpoint of the head in this slice is calculated, and from this centre-point, points on the surface of the head are selected, just as it was done with the MRI scans. However, in the case of PET, a constant number of points in each slice can be used.
The coordinates of each surface point are expressed in mm, so that the different size of the PET voxels is automatically taken into account. A transformation is used to translate the PET axis to the axis used in the MRI scans. Points are also taken on the outer surface of the head from the already segmented MRI scans, but now in every slice.
A least-squares algorithm is applied to reduce the distances between the PET and the MRI points. Since there are many more MRI points than PET points, for each PET point the distance is calculated to the nearest MRI point. The algorithm searches for six parameters that characterize a three-dimensional translation vector and three angles through which to rotate the PET points. The rotation is carried out according to the following formula:
in which (X0, Y0, Z0) are the original PET
coordinates, and 
, 
and 
are respectively the rotation around
the x, y and z axes. The rotation is always carried out around the
centroid of the PET points.
The least-square search algorithm is divided in two steps. In the first step, the routine only searches for a suitable translation vector. The rotation is fixed at an angle of 30 degrees backwards, which is the approximate scan direction for PET with respect to the orientation of the MR1 slices. The second step optimizes the translation as well as the rotation. The algorithm uses the method of steepest descent to find the translation and rotation. Although the translation does not cause any problem, the proper rotational parameters are difficult to find, not in the least due to local minima, and to the fact that the sum of all the distances hardly changes with a change in rotation. Operator assistance is usually required for the most optimal fit. The approach of Neiw (1991) to solve this problem is very interesting. He fits the points on the contour in two slices from the different modalities that lie on the line formed by the intersection of both slices. It is advisable to implement such a method, since the reported results are good, and local minima are avoided.
Once the matching parameters have been established, any slice in the MRI can be recalled. The corresponding PET coordinates are then calculated by inverse rotation and translation, and the PET scan is overlaid in colour over the gray scaled MR1 scan. The program has to be switched to the PET/MRI mode, and if in this mode an MRI scan is displayed, with or without a dipole, the computer will take the MIRI coordinates for each pixel in the slice, and transform it back to PET coordinates. Since PET pixels are larger than MRI pixels, the resulting coordinates are divided by the size of the PET voxels, and rounded to the nearest integer. The value at these coordinates in the PET scan determine the colour to be mixed with the colour of the voxel of the MIRI scan at that position.
Since the number of colours on our display is limited to 256, of which 128 are already in use for the grey scale needed for displaying the MRI scans, it is not possible to use new colours necessary for laying the PET scan over the MRI scan. The choice that was made was to make the PET scan 50% transparent, thereby allowing the viewer to see the underlying MRI scan simultaneously. A solution to the problem of the restriction on the number of available colours is to alternate pixels from the PET and the MRI scans, like a chessboard. This works well for most displays of the MRI scan, since 1 pixel in the MRI is displayed as either 3x3 or 2x2 pixels on the screen. In the case of the sagittal view of the MRI scan, each pixel in the MRI is mapped to one pixel on the screen. In this case 50% of the MRI pixels are replaced by PET pixels. This method has the advantage that one is still able to manipulate the colours used to display the PET data.
An example of the overlay of the transmission scan with the MRI is shown in figure VII.2. Notice the head rest visible in the transmission scan, and how well the scans fit on top of each other.
Figure VI.2: PET transmission scan overlaid on an MRI slice
Figure VI.2 shows a result obtained by the method described above of matching MRI and PET. The matching is accurate within a PET voxel. Data from the MEG and EEG measurements, the MRI data and the PET data from one subject can now be combined. Due to the relatively low time-resolution of the PET scans, the obtained value is an integral over several seconds. Therefore several active areas of the brain can typically be observed in the scan. In MEG experiments, however, one records the brain activity on a millisecond scale. The highest value of the integral of the brain activity computed over a given time lapse as is indicated in the PET scan need not coincide with the location of the area where the highest activity at any time instant took place as localised from the MEG. Furthermore, in these PET experiments it is assumed that areas of the brain that are more active during stimulation than during the resting condition have a higher blood-perfusion, and therefore a higher [15O] content. Therefore, neuronal activity is measured only indirectly. Indeed both MEG/EEG and PET measure different phenomenon. Until more research in this field has been performed, it is by no means certain that active areas of the brain, obtained from both modalities should necessarily coincide.
To test the matching procedure and to explore the possibilities in presenting the results of combining all these modalities together the area of the brain which is activated by median nerve stimulation of the wrist was localized, using MEG, EEG and PET in the same subject. In order to measure the activity with a PET scanner, the subject was injected with radioactively labeled water, H2O15, which has a half-life of 123 seconds. The radioactive dosage was 60 mCu per injection. The water was injected into the lower arm opposite to the side which was stimulated. Immediately after injection, the electrical stimulation was started. A square pulse was used with a duration of 400 ps with a strength just below the thumb twitch threshold (7.510 mA) and with an interstimulus interval of one second. The acquisition of the PET data was started after a rise in count rate from the PET camera, indicating that the tracer had reached the brain. The total measurement time for each condition was 40 seconds, divided into 15 frames. The subject was positioned in a head mould to fix the head. Apart from the scan obtained during stimulation, a scan was also taken without stimulating the subject. The resting period between the different sessions was about 15 minutes, which allowed the radioactivity in the body to be reduced to below the background radiation level.
The MEG was measured in 30 positions over the right hemisphere of the head. The EEG was measured at 23 electrode positions, placed according to the 10-20 system, using an electrocap. The electrode positions were also chosen mainly on the right hemisphere of the head of the subject. The measurements were corrected for eye movement artefacts. The interstimulus interval was again 1 second, and the subject was continuously stimulated for at least 150 seconds. Each measurement was carried out twice. The signals were averaged over at least 230 responses.
An example of three MEG channels is shown in figure VI.3. The localization of an equivalent current dipole was carried out on the first peak, between 25 and 35 ms after the stimulus in the MEG and between 30 and 40 ms in the EEG. This localization was carried out separately for the MEG and the EEG measurements. The volume conductor was simulated by a four-sphere model. A different sphere was taken in both cases, which clearly shows that a best fitting sphere is not uniquely defined. The results from the MEG are projected in the corresponding MRI slice in figure VI.4, with the PET data overlaid, using translation and rotation parameters derived with the method described above.
Figure VI.4: The equivalent current dipole, localised from the MEG,
with the MRI and PET data
The MEG dipole seems to be located just posterior of the central sulcus, in the somatosensory cortex. The PET data was derived by subtracting the PET images taken without stimulating the subject from those taken during stimulation. The used sphere is also indicated. As can be seen, the dipole overlaps with an area of the cortex that shows an enhanced cerebral bloodflow in the PET scan. The almost symmetric frontal activity seen in the PET scan was due to movement artefacts.
Figure VI.5: The equivalent current dipole, localized from EEG,
displayed with MRI and PET data
The corresponding equivalent dipole localized from the EEG data can be seen in figure VI.5. The position of the MRI slices containing the MEG and the EEG dipole are indicated in figure VI.6. There is a difference with the dipole localized from the MEG. The dipole is also found in right posterio-temporal region of the brain. The EEG dipole is found about 2 cm lower in the brain. However, this difference ran be explained in terms of a localization error. Also in the case of the EEG results, brain-activity can be seen in the PET scan in the same area as the dipole location.
Notice
that the activity seen in the PET data is not seen in the left hemisphere of the
brain, which is as expected when the left wrist is stimulated. The PET
data, however, does not give any information about a possible erroneous
localization of one of these dipoles. From Figures VI.4 and VI.5 it is
still very difficult to estimate the location of the dipoles and the PET
activity within known neurophysiological areas. Therefore another
projection was created. Figure VI.7 shows a three-dimensional view of the
surface of the right side of the brain, with PET data projected on this surface,
as well as the positions of both dipoles. The information from the PET
indicates much activity over the right hemisphere. A large area shows a
higher level of activity than in the rest condition. Various factors may
account for this abundance of brain activity see the PET. First, as
mentioned earlier more areas of the brain may be activated by stimulation than
just that area that is active when a peak in the magnetic and electric responses
is measured. The activity of those other areas is also registered by the
PET scan. This would suggest that all the activated areas seen in the PET
scan can be explained not only in terms of primary sensory processes but also of
secondary processes, possibly cognitive, that take place during the time a PET
scan is being made. Second, the stimulation may not be precise enough to
stimulate only cortical areas where the thumb is represented. The current
may also stimulate nerve fibers from the other digits. A larger cortical
area would then be active. In this case, the usefulness of a single
equivalent dipole localization may be questioned, since an area with a
significant extent would then be active at the same time and render a single
dipole approximation invalid. This could explain the poor goodness-of-fit
that was found in these localizations. Third, the PET images have a poor
signal-to-noise ratio, which deteriorates when two such images are
subtracted. Furthermore, the PET scan of our subject shows significant
movement artefacts in the frontal region. Together with the low dynamic
range of these images, this produces incorrect results for some areas.
Figure VI.7: 3D-View of the brainsurface with PET data and dipoles
projected onto it
An important aspect to note about Figure VI.7 is that neither the PET activity nor the dipoles are present in the area of the brain where they are most expected, namely the somatosensory cortex for the hand. To verify this observation, a match between the MRI and the PET scan of a second subject under the same conditions was made. The 3D view can be seen in figure VI.8. It is clear that the distribution of PET activity is the same, with the exception of the absence of the frontal activity due to the movement artefacts of the previous subject. This would indicate that the PET image expresses mainly secondarily activated regions of the brain. The somatosensory cortex may be stimulated for a time-period too short to show up in the PET scan under these conditions. However, also the MEG and EEG indicate the same region. Generally, one dipole is fitted to the data from magnetic response measurements from median nerve stimulation. However, Baumgartner et al. (1991) used two dipoles to explain the magnetic data. They found two peaks during the first 40 ms in their response after 500 averages and accounted for them with two dipoles which were simultaneously active. One of them may be located in the somatosensory cortex. A principal component analysis of the MEG data showed that it was possible to fit two dipoles to our data, but one of the dipoles was then localized outside the brain, although still in the same region of the head. However, principle component analysis of the EEG data showed that is was impossible to fit also two dipoles to the EEG data. Since only one dipole could be found from the EEG data, this suggests that only a single dipole is responsible for the responses between 25 and 35 ms in the MEG. Because, from theory we know that radial dipoles would be magnetically silent, although they would yield a signal in the EEG, the number of dipoles found using the EEG should be equal or more than the number of dipoles found from MEG. Apparently, a better signal-to-noise ratio is required to detect the activity associated with the somatosensory cortex.
Figure VI.8: Combination PET and MRI data, subject 2, in 3D view of
brain
Clearly, the results are not as trivial as might have been expected. The dipoles would normally have been rejected as inconsistent with neurophysiological knowledge, but the combination of the MRI and the PET data in the 3D view changed this point of view. In our opinion, this already demonstrates the usefulness of having available a method of fitting MRI, PET, MEG and EEG data together. That the use of this method leads to new questions maybe the best result that could have been hoped for. Clearly, further research using these methods is necessary.
(c) MEG, EEG and the integration with Magnetic Resonance Images, H.J. Wieringa, 1993
[<<Chapter V][Contents][Home][Chapter VII: Errors in source localisation>>]