1. Introduction

The investigation of the role of Supplementary Motor Area in voluntary movements is indubitably linked to the evidence of Readiness Potentials (or Bereitschaft Potentials, BP) preceding volitional movements as demonstrated in EEG by L. Deecke (see Deecke et al., 1969 for instance).  In his seminal work, Deecke shown that BP waves could be decomposed in two distinct components: i) BP I, starting as early as 1.5 s to 1 s (typically 1.2 s) prior to movement onset with a symmetric and bilateral topography on the scalp, and ii) BP II, starting typically 0.5 sec before movement onset, usually distributed above the contralateral primary motor area (MI). Rich experimental evidence was gathered since then, indicating that BP I and II might involve different functional systems (Lang et al., 1991).

Bilateral neural activities were found in the precentral gyrus with both ECoG and MEG, even for unilateral movements, but proved to be little compatible with BP I spatiotemporal pattern. Experiments in PET and unicellular recordings in monkeys suggested that parts of the mesial frontal cortex, and typically the SMA, may be involved in BP I generation.

Unfortunately, source modeling of the SMA activation proved to be difficult. A possible reason for this may spring from the morphology of the mesio-frontal cortex where both hemispheres come very close to each other. Further, dipole modeling of the crown of the SMA gyrification would produce two tangential dipoles facing each other. This would imply at least the partial annihilation of the magnetic field in the case of simultaneous and bilateral activation of the SMA area. This hypothesis has been partly confirmed in a case study from a patient with a lesion in the right SMA region resulting in detrimental contralateral spontaneous movements (Lang et al., 1991). Source estimation of the Movement Evoked Fields (MEF) resulting from brisk opposition of the right thumb was done on the [-1200, -200] ms time window before the movement onset (detected from EMG monitoring). A single-dipole model was fitted with an 86% goodness-of-fit (GOF, based on the relative mean-square error between the data and the model) on the [-800, -600] msec time window. The dipole location was consistent with a possible localization in contralateral SMA. The dipole then jumped from SMA to the contralateral MI hand region in the following time window ([-600, -200] ms) with a poorer GOF of 62%; making the authors assume a sustained and concurrent activity in the SMA. But, two-dipole source models proved to be ineffective on this data.

More recent studies as in (Bötzel et al., 1993) failed to produce reliable dipole modeling in SMA and MI for the BP I and II waves respectively. Non-invasive identification of SMA and MI with MEG and EEG source modeling in healthy subjects proved to be problematic as confirmed by intracranial recordings of movement related readiness potentials in epileptic patients described in (Rektor et al., 1994). The results were interestingly complementary of Deecke’s group. Following hand and foot flexion, the study revealed that BP I and II did not seem to have clear SEEG counterparts either in bilateral SMA or contralateral MI but both of these areas were found as being activated before movement. No clear temporal scenario between these two areas was identified and almost no activity in ipsilateral MI was found.

In reply to this study, Deecke and Lang suggested that the so-called SMA region recorded in this experiment was too posterior to the actual SMA region to indicate a clearer temporal pattern of activation between SMA and MI.

By 1996, it was still not clear according to many researchers whether SMA could possibly be activated by tasks involving simple movements, or at least would precede MI activation. As reported in (Deecke and Lang, 1996), SMA activation would be much more expected during the preparation of complex movements involving a higher level of temporal coordination (e.g. reproduction of rhythm patterns) rather than spatial only. In fact, it appeared that simple finger opposition tasks, such as the ones used in the very first source localization attempts, could be successfully completed by patients with lesions of the SMA. Some bimanual finger-tapping task recorded in EEG revealed that SMA activation would be involved only in the early stages of motor preparation (i.e. before M I) prior to simple movement patterns, whereas more complex patterns revealed some sustained SMA activation concurrent to M I (Deecke & Lang, 1996). Actually, the distinction between a possible role of the SMA in either the temporal or spatial development of the movement is illusory as both of these variables (time and space) are linked within the very physics of movement execution (Uhl et al., 1996). 

The current consensus is that SMA is strongly involved in the decision for the initiation of movement, the coordination between limbs and the participation in motor learning.

This is not until very recently that dipole source modeling of the magnetic counterparts of BP I and II – BF I and II – has been completed successfully in a 3-subject study (Erdler et al., 2000). Equivalent current dipole (ECD) fits were completed on both the BF I and BF II waves recorded on a whole-head array MEG system. A single current dipole was found in the SMA area at about –1.9 s to –1 s prior to the movement onset (thus corresponding to BF I). A 2-dipole model could be adjusted to a time window also including the BF2 field, until –500 ms prior to movement detection. The BF I onset was all the earlier as the movement was simpler: a single dipole located in the SMA region was fitted as early as –1 s prior to movement for the subject performing a simple finger-tapping task.

These much-anticipated findings presented the very first source models for BF waves recorded in healthy subjects. The authors concluded that the whole-head sensor coverage brought a major improvement in the spatial resolution of the source modeling. This with the fact that SMA was non-invasively localized from tasks involving very simple movements, represented an authentic achievement for whole-head MEG systems.

The main purpose of our study was to replicate these findings. Beyond that, our goal was to compare MEG source localization results and the concurrent fMRI analysis obtained form the same subject in the same task conditions. Our assumption was that maybe partial sensor coverage of the head used in previous experiments were not the only reason of their “failure”. Going back to the anatomo-functional properties of the SMA in some subjects (possible cancellation effects of opposite dipoles, see above), it is still very likely that BF waves would be very small in case of bilateral and synchronous SMA activation, thus not ensuring robust source localization with dipoles. We have used new higher-order MEG source models to assess whether these could be more successful in localizing brain areas where ECD modeling would fail (Mosher et al., 1999). This approach, together with co-registration of the MEG with MRI and fMRI represents an advance in the evaluation of source modeling methodology and the non-invasive localization of SMA activation in healthy subjects.

The rest of the paper is organized as follows: in section 2, we briefly review the rationale of multipole source modeling in MEG; section 3 describes the experimental set-up and the resulting data obtained in fMRI and MEG. Results are discussed in section 4.

2. Methods

2.1        MEG Source Modeling using Multipolar Expansions

Single or multiple current dipole fits using Least-Squares or MUSIC-based approaches are certainly the most popular source modeling techniques in MEG. But one of the key limitations of current dipole source models is that they may prove to be inefficient in accounting for the synchronous activation of broader cortical areas. The ECD may fit properly to the forward field produced by such an activation, but it is very likely that its location would be far from the centroid of the true activated patch, thus making the appraisal of the localization results rather problematic.

This statement has motivated most of the early approaches to distributed cortical source imaging (Hämälainen et al., 1993). But these latter are highly underdetermined, and thus necessitate the introduction of priors for current density patterns; typically with smooth variations on the cortical surface – like with Minimum-Norm (MN) approaches  – or with sparse focal clusters of cortical activation (Phillips et al. 1997), (Baillet & Garnero, 1997). The limiting factors of these two types of approaches consist of either rather coarse spatial resolution for MN estimates or numerical issues for the more computationally intensive techniques with sparse focal priors.

Our intention was to develop parametric source models that could account for larger cortical areas while still being usable with fast source fitting algorithms. We addressed this problem by the use of higher-order terms in the multipolar expansion of the magnetic field produced by an arbitrary current source within the head (Wikswo and Swinney, 1984, Katila and Karp, 1983). The technique used here is detailed elsewhere (Mosher et al., 1999). Basically, the truncated Taylor series expansion of the magnetic fields in MEG naturally gives rise to multipolar components (dipole, quadrupole, octupole,…). For distributed sources confined to a patch of cortex some distance from the sensor, the contributions to the magnetic field from octupolar and higher order terms drop off rapidly with distance, so that restricting sources to dipolar and dipole + quadrupole is probably sufficient to represent most plausible cortical sources. But as the extent of the source grows, more terms are required in the expansion to adequately represent the external magnetic field. Discrimination can be done between (i) point sources that are exactly represented as point current dipoles, (ii) highly focal sources that can be represented by a magnetic dipole moment model, and (iii) locally distributed sources that can be represented by a first-order multipole model.

2.2        Source Localization using RAP-MUSIC

The extensions of the MUSIC approach to the multipolar models are direct. Each higher-order model is still described by three nonlinear location parameters, but with different and/or more linear moment coefficients. As in the current dipole model, the MUSIC search is carried out only over the nonlinear parameters.

We have used the RAP-MUSIC extension to MUSIC (Mosher & Leahy, 1999), which avoids the “peak picking” issues of the basic MUSIC approach. RAP-MUSIC proceeds sequentially , estimating the parameters of a first source model which subspace correlation with the signal subspace estimated from the data exceeds some threshold. Then both the subsequent subspaces spanned by the next source candidates and the data are projected away from the subspace spanned by the previous source(s). Hence, RAP-MUSIC does not require any prior choice of  the number of sources in the model.

The different source models available (from current dipole to quadrupole) are evaluated sequentially by the algorithm starting from the simplest (current dipole) and moving to the magnetic dipole moment and finally the quadrupole if the intermediate subspace correlation checks do not fall above the authorized minimum value. This version of RAP-MUSIC using higher-order source model can be found in BrainStorm, an integrated Matlab toolbox developed among our groups and freely available on the Internet (http://neuroimage.usc.edu/brainstorm).

3. Results

3.1 Data Acquisition

We describe the case of a right-handed male subject with no neurological antecedents who was investigated with MEG and fMRI on a self-paced brisk forefinger flexion paradigm. Movement execution was monitored so that the subject would keep a slow pace of about 3 to 4 s between each flexion.

fMRI acquisition was done on a General Electric Signa 1.5T scanner. We followed a classical block design of 4 cycles (movement vs. rest) for the acqui­sition of 30 scans per block (2 slices, 8mm thick embedding the primary motor areas and the SMA, TR=1.2s, Te = 45ms). 

MEG was acquired with a Neuromag-122 system with 122 planar gradiometers distributed at 61 locations in a whole-head array. The MEG paradigm consisted of 250 trials of the same self-paced movements. An optical switch was used to trigger the MEG averaging –1.5 s to 1 s about the movement onset. MEG data were band-pass filtered between 0.01Hz and 10Hz.

3.2 Data Analysis

3.2.1 fMRI

fMRI data analysis was performed with SPM99 using a typical “boxcar” design. Significant BOLD effects (p<0.001) were found principally in the contralateral post and pre-central areas and SMA (see Fig.1).

Figure 1. SPM analysis of the BOLD effects revealed by fMRI (p< 0.001) indicates clear increase of perfusion in the contralateral pre and post-central regions and SMA.

3.2.2 MEG

Typical MEG gradiometer waveforms recorded above the contralateral sensorimotor cortex and the central superior region are displayed in Fig.2. A gradually increasing slow wave is visible over the central region, starting about –1.5 s peaking just before the detection of movement onset. The central contralateral sensors indicate a smaller premotor wave starting about at –1.2 s followed by the strong sensorimotor response peaking at about 60 ms.

(a)	(b)

Figure 2. MEG gradiometer waveforms. (a) Evoked field recorded above the contralateral sensorimotor area; (b) ERF above the central superior region.

RAP-MUSIC source modeling was done on the [-600, 250] ms time window. Rank 24 was chosen as the dimension of the signal subspace following a singular value decomposition of the data matrix. A threshold of 98% was chosen when computing the subspace correlation between the source forward fields and the signal subspace. A 3-source model came out of the RAP-MUSIC scan with the following parameters: a current dipole in the contralateral central sulcus with a 99.6% subspace correlation level (source 1); no other current source could be found with a sufficient level of subspace correlation. One quadrupolar source was found in the superior part of the central fissure (source 2, 98.5% correlation), and another one possibly in the ipslateral premotor region (source 3, 98.3% correlation), (Fig. 3).

Figure 3. MEG source locations superimposed to fMRI analysis.  Numbers correspond to time series displayed in Fig. 4.

Comparison with the fMRI clusters reveals a good match in localization between the two modalities for the first 2 sources. Source 1 is 12.7 mm away from the maximum of the closest fMRI cluster with the largest error being along the infero-superior direction; this source is located just below the most inferior fMRI slice of the two. Source 2 is located 4 mm away from the fMRI cluster identified as the SMA. Source 3 appears not to have any fMRI counterpart.

The source time series are displayed in Fig. 4. Source 1 and 2 have ongoing and sustained amplitudes since the very beginning of the time window. The ispsilateral source amplitude increases about –550 ms before movement onset. The source amplitudes increase gradually until just before movement detection when the contralateral source switches to a large biphasic wave. The ipsilateral premotor source dies out at about that time. SMA activation keeps on increasing gradually (in absolute value) until about 60ms after movement onset. The ipsilateral premotor source probably accounts for motor control of the left hand supposed to remain still when moving the opposite finger. This constraint was explicitly requested from the subject.

Figure 4. Source time series corresponding to source locations displayed in Fig. 3

Cortical Remapping of MEG Sources

Because the representation of parametric source models in the head volume is focal, it makes the true spatial extent of the corresponding neural populations difficult to appraise. We have proposed recently to find equivalent distributed cortical source models for each of the original parametric sources (Mosher, Baillet & Leahy, 1999). This “cortical remapping” of parametric sources allows a more straightforward comparison of the MEG analysis with fMRI.

For each of the sources found in the original parametric source model, cortical remapping consists in finding the smallest cortical patch which forward field matches the original source’s. The criterion we have used is the subspace correlation between forward fields, assuming that the cortical patch is made of sources with uniform amplitude. Cortical sources are constrained at the vertices of a cortical tessellation following segmentation of the white/gray matter interface obtained from anatomical MRI (Shattuck and Leahy, 2000) using the BrainSuite software (University of Southern California, (http://neuroimage.usc.edu)

For each of the parametric sources found, 10 cortical locations with the best subspace correlation with the current source were selected as seed points from which a source patch is grown by adding connected cortical sources until subspace correlation of the forward fileds reaches 99% with the original source. Details of the method are described elsewhere (Mosher, Baillet & Leahy, 1999).

The cortical patches obtained are displayed in Fig. 5 and overlaid with the fMRI clusters. Very good match is achieved between the areas found in both modalities.

 

Figure 5. Overlay of the fMRI clusters (in yellow and red) with the cortically  remapped MEG sources (in green)

4. Conclusion

We have presented a joint MEG and fMRI investigation of Supplementary Motor Area activation preceding voluntary finger movements in a healthy subject. MEG source analysis was done without any prior information from the fMRI. Owing to the use of higher-order source models, SMA and ipsilateral premotor activations have been demonstrated with this subject during movement preparation and execution (for SMA only). This latter source was not found with fMRI.

Locations of the contralateral and SMA sources were found rather close to the maximum of the closest fMRI clusters (12 mm and 4 mm, respectively). Cortical remapping of the focal parametric source model was performed and the cortical clusters corresponding to the contralateral and central sources indicated a good match in location with the fMRI regions.

We believe higher-order source models are useful to account for the activation of larger cortical areas, while enjoying the benefits of a fast and robust estimation of their parameters with scanning methods such as RAP-MUSIC. The inadequate focal representation of these sources in the head volume can be subsequently overcome by  cortical remapping techniques, which estimate the spatial extent of the equivalent cortical areas involved.

Acknowledgements

The authors are grateful to Cheryl Aine, Robert Christener and Ming-Xiong Huang at the MEG facility of the VA Medical Center, Albuquerque, New Mexico, for their assistance in collecting the MEG data.

References