This thesis deals with various aspects of the measurement of biomagnetic signals, and the incorporation of the results in MRI scans including a method for further evaluation by using PET scans.
We have shown a convenient 64 channel data-acquisition system, which allows continuous sampling at sample rates of 1000 Hz per channel. Although this sample rate is sufficiently high for most evoked response measurements, demands on the instrumentation increase. In the future, higher-performance systems will be developed in which the sample rate can be varied between 100 and 5000 Hz per channel. This enables the system to measure, for instance, early somatosensory evoked responses, but also to sample at lower rates for less demanding experiments, thereby reducing the amount of storage space needed. This development also places a higher demand on the filters, and therefore the use of programmable digital filtering is preferred. To enable DC-measurements, a 21 bit AID converter is under development. Demands on the acquisition system will also increase in terms of speed, number of acquisition channels, on-line display possibilities and storage space. Already our acquisition computer has been equipped with a 1 Gb. hard disk. Further increase in the amount of data will have consequences for the speed and storage-capacity of the computers used for analyzing the data and the devices used to archive the measurements. Programs will have to be adapted, not only to deal with an increase in the number of channels or data points, but also in the presentation of the data to the investigator. The huge amount of data needs to be easy accessible to the investigator, with comprehensive views to enable it to be analyzed accurately. Although these problems are not uncommon to users of super-computers, in the field of biomagnetism this is still an uncovered area.
Two methods to reduce the noise in biomagnetic signals were shown. The first, electronic balancing, works very well outside magnetically shielded rooms, yielding improvements in the signal-to-noise ratio of a factor of 10 to 20, but does not work properly inside shielded rooms. The second method, active shielding, actively compensates disturbing magnetic fields from the environment. This method works only in combination with a shielded room. The improvements are very impressive for magnetometers, in which the noise is suppressed by up to 40 dB, but only yield an improvement of a factor of 5 for gradiometers. These noise-suppression schemes may have their value when biomagnetic sensing devices are being used in the clinic in an inherently noisy environment. They could very well be a cheap alternative to additional shielding with p-metal and/or aluminum.
The methods for combining MEG/EEG results with MRI, shown in this thesis, enable an experimentator to view an equivalent dipole calculated from MEG and/or EEG measurements in MR] scans. Various views are available, and the investigator should be able to manipulate these views to get a general idea about the location of a dipole, to correlate it with other methods like PET, or to compare it with views taken from functional brain atlases. At this moment, viewing the data, and creating new views takes a relatively long time. Although this is, in principle, not so much a problem in a research environment, researchers should be stimulated to use these new techniques to push beyond the current standards. Therefore, the process of generating the views has to be speeded up. This can be done in an easy and expensive way by updating the computer used to current standards, increasing memory and on-line storage capacity at the same time. However, the dataset we use is very large, and does not fit completely in memory. This generates a lot of overhead by having to read from disk all the time. By reducing the resolution of the MRI scan from I mm3 to 2 mm3, the amount of data is reduced with a factor of 8. This reduced data set fits into memory completely, so it only has to be read once. Furthermore all operations on the data set, like reconstructing different planes or creating 3D views, will be reduced also by a factor of S. The time necessary to render the data on the screen will be reduced by another factor of 4. The result should be that most operations can be carried out within 1 minute. High resolution views can always be obtained if deemed necessary. If further optimizations are made, like selecting a view by pre-showing a low-resolution image, the program can be made even more interactive, inviting the investigator to try out different views. Other improvements which should be considered are:
To improve the accuracy of the inverse solution, more realistically shaped multicompartment models of the head can be used, which can be generated from the MRI scans. To enable the generation of these models, and to be able to generate 3D views of head and brain, the MRI scans needs to be segmented. We have shown a practical and successful method to do this completely automatically in chapter IV. Automatic segmentation of medical images is a problem which has not been solved in general. Our procedure appears to work well in most case, partly due to the high resolution of the scans used. This high-resolution makes it possible to neglect any partial-volume effects, where voxels derive their value from more than one tissue type due to their finite extent. The segmentation procedure is also specific for the head. Further improvements could be made with minor modifications, for instance to extract the ventricles and to discriminate between the grey and white matter of the brain. This should facilitate the construction of more realistic models. The segmentation procedure has to be carried out only once for each dataset. The segmentation is stored in the original file, without modifying the original voxel-values. Therefore, it is still possible to re-segment the dataset after improvements to the algorithm have been made. To completely segment a 256 slice dataset in high resolution takes currently about twelve hours. Optimizations to the algorithm are difficult to implement, since the procedure has to be robust for the many different situations encountered. However, this does not usually present a problem, since the segmentation can take place during the night.
From simulation studies it emerges that the accuracy of the source estimation is strongly dependent on the model of the head. The information on the conductivity distribution as found in the literature is not very accurate. Information on the geometry can be obtained from MRI as described in chapter IV. Usually a set of concentric spheres is used as a model of the head. The spheres are fitted to that part of the boundary of the head which is nearest to the source. This is not an adequate description as is illustrated in figure VIII.1 in which the head is described by a four sphere model which is obtained by fitting the spheres to the occipital region of the head. It is obvious that the head is not a sphere and neither are its compartments. When the electrodes used to measure the potential distribution cover a large part of the surface of the scalp some measuring points will not be located at the outermost sphere of a multi-sphere model. Consequently, a realistically shaped scalp surface should be incorporated into the model of the volume conductor for the EEG.
Inverse procedures for realistically shaped models are presently all based on the boundary element method. Thus, a model consisting of realistically shaped compartments has to be constructed, where all compartments have closed surfaces and a homogeneous, linear and isotropic conductivity. The MEG leads itself to the use of a simpler model than the EEG. This is due to the fact that the skull is poorly conducting, so the main contribution to the magnetic field outside of the head is from currents flowing in the brain tissue, and because the secondary sources at the outside of the skull and the ones at the outside of the scalp counteract each other. Consequently, in the case of the MEG the head can be described by a homogeneous model shaped as the inner surface of the skull. We have described in chapter IV a practical and successful method to extract such a model completely automatically from MRI-data. As mentioned above, multi-compartment models are needed in the case of the EEG. The generation of the (triangulated) outside surfaces of the brain and the scalp is fully automated. The triangularization of the outside surface of the skull sometimes requires human interference. In the healthy subjects studied the layer of CSF was in the order of I mm. From simulation studies of concentric sphere models, De Munck (1989) found that when the CSF layer of 3 mm was omitted the depth errors were very small, but the error in the tangential direction could be 6 mm. Therefore, it seems that for elderly subjects, who may have large spaces filled with CSF, this layer should be taken into account. The description of the surface of this layer is easy to obtain from brain surface data by using a scaling factor. From simulations it followed that the magnetic field will very much influenced
by the presence of a lesion and indeed could even be reversed (Ueno, 1992; Peters and Wieringa, 1993). Since the magnetic field is dependent on the distribution of the potential, it is to be expected that the EEG is also influenced by such a lesion (if the lesion is more conductive than the surrounding tissue we have the so called Brody effect). As the ventricles are filled with CSF the influence of the ventricles may be substantial for sources nearby. The ventricles can also be obtained from MRI-data with the method described in Chapter IV. An example is shown in figure VIII.2. Another geometrical aspect of the volume conductor which is expected to have a large influence on the EEG and MEG are breaches and openings in the skull (Bertrand et al., 1992). These structures will be difficult to obtain from MRI and certainly these cannot be extracted in a fully automated procedure. Although the generation of triangulated surfaces is automated, the total procedure (i.e. MRI measurements, segmentation, inverse solution using the boundary element method) is still very time consuming. The time for the calculation of a source can be reduced dramatically (a reduction from hours to seconds) if a scaled standard head model can be used in the inverse procedure.
To properly assess the shown dipole location in the MRI, an estimate of the error in this position is necessary. In chapter VII it is shown that this error depends on all the steps taken to reach this result. A global estimate is made for three situations,
i.e. for the case in which a realistically shaped model of the head is used, for the case in which a sphere model is used, where the sphere is fitted in the MRI scan, and for the case in which a sphere model is used which is fitted to points on the surface of the real head. The estimated error-estimates in the location of the dipole were for these three cases were 6.7 mm, 10 mm and 12 mm respectively. From these error-estimates we can conclude that the best results for the absolute localization of an equivalent current dipole in an MRI scan are obtained using a realistically shaped model of the head derived from the same MRI scan. Using the sphere model gives worse results, especially in the case that the sphere has been fitted to the head externally, and is afterwards transferred to the MRI. From these results it can be concluded that the errors due to the use of spherical head models are the largest. Therefore the reduction of these errors is of first priority. All studies using models which are more complex than the concentric spheres model are not of a statistical nature because the number of parameters that can be varied is too large. Consequently, they cannot provide the mean error but just a possible error. The simulations will show us the tendency of the error due to certain aspects of the volume conductor. We conclude that the accuracy of the functional imaging based on MEG and EEG will be enhanced when the shape of compartments plus breaches in the skull are taken into account. However, it is difficult to validate results.
The MEG and the EEG are the only two methods to study the functional organization of the brain with a temporal resolution of I ms. These methods are complementary rather than competitive as argued in chapter V, which shows that all six parameters of the dipole can be obtained with a higher accuracy if both MEG and EEG are used. The location of the dipole should be obtained from MEG, its direction from EEG, and the strength from both MEG and EEG together. The strength of the dipole can then be used to estimate the activated patch of cortex, using knowledge gained from electrophysiology. The method, as suggested by Williamson at al. (1991), of deriving the direction of the dipole from the tangential component found from MEG and the local curvature of the cortex instead of using the EEG seems impractical since it requires an unrealistically high localizing accuracy. This is particularly so as they use a sphere model the fit of which is not uniquely defined and since the position of the sphere influences the tangential component found in MEG. Therefore the direction of the dipole can not accurately be determined in this way.
An important part of future modelling research is experimental verification of the models and localization methods. The results have to be compared with anatomical and electrophysiological knowledge and tests with implanted sources in the human brain are also of importance. Localization studies to estimate EEG and MEG accuracy of localization by means of implanted electrodes are all based on spherically symmetric head models. It is of importance that more realistically shaped models are used. Another method to provide a critical assessment of MEG and EEG is to compare the results with those obtained with other methods i.e., functional imaging with MRI (fMRI); PET and Single Photon Emission Computer Tomography (SPECT). PET and SPECT provide a measure of chemical specifity and although the basis of fMRI is currently speculative, it probably shows the changes in blood volume and other vascular-related phenomena. All these methods have a time resolution which is much longer than I ms; the time resolution of oxygen-15 based PET is 40 s and of fMRI 2 s. In order to be able to compare the results obtained with these various modalities, it is necessary to have a method at one's disposal for multimodality matching of brain images as described in Chapter VI. It is difficult to decide at this moment whether the results of comparisons have to be displayed in 3D-images or in slices combined with the coordinates.
The clinical use of source localization based on MEG and/or EEG is still rather limited: the localization of the epileptic focus during presurgical diagnosis (Stefan et al., 1991); the delineation of the somatosensory cortex in patients submitted to neurosurgery, who will undergo a resection adjacent to the central fissures which can help avoid the accidental disabling of important brain areas (Gallen et al., in press). For these measurements it seems of the utmost importance that the measurements are carried out fast and that the total localization procedure leads to a reliable and precise result, presented in a clear and unambiguous way. Standard realistically shaped head models which can be scaled would provide an inverse procedure which would meet these requirements. If, for instance, the development of the brain is studied as a function of age the researcher is not interested in the absolute location of the sources but in the relative one (Ossenblok, 1992). Neuropsychologists are interested in processes such as cognition and language processing. They may, for instance, be interested to know the effects of attention on the cerebral generators that underlie the fields and potentials to both attended and unattended stimuli. In this type of research the interindividual differences in the exact positions of the sources are of less importance than knowledge about the relative position of active regions (Wijers et al., 1992). There is no need for individualized head models in inverse solutions in cases where researchers are not interested in the absolute position but in the relative one.
In future MEG, EEG and MRI may strengthen each other further. First, findings obtained with fMRI can be compared to findings obtained with MEG and EEG. Second, localization results from MEG and EEG can be used to choose the volume in the brain where Magnetic Resonance Spectroscopy will give additional information. Also, the structural information in this region can be enhanced by scanning the selected region at sub-millimeter resolution. Third, the results obtained from MEG and EEG can be evaluated by electrophysiologic knowledge which will be made available through atlases of the brain. Of course, the position in the brain has to be indicated in such a way that it is possible to know where the corresponding location is in an individual head. If an algorithm for scaling is available, it will also be possible to compare MEG's and EEG's with findings by PET which are averaged over a number of individuals. In this way, we can hope that the brain will reveal more of it's secrets to us, and that it will enable us to battle the various brain-diseases successfully.
(c) MEG, EEG and the integration with Magnetic Resonance Images, H.J. Wieringa, 1993