Information Technology Reference
In-Depth Information
to multisource scanning protocols that search through combinations of sources. Al-
though multisource scanning methods can recover perfectly synchronized sources
(which are usually missed by single-source methods), there is no agreed protocol to
scan the immense space of possible multisource configurations.
8.8 Comparison of Methods
To compare some of the methods reviewed here, we performed a retinotopic map-
ping of the four visual quadrants of one subject using the Elekta MEG system,
which contains 204 planar gradiometers and 102 axial magnetometers. Data were
processed with SSS and bandpass filtered (2-55 Hz). One hundred epochs were av-
eraged for each quadrant. A BEM forward model was used and sources were con-
strained to the cortical surface of the subject.
A schematic of the black and white checkerboard visual stimuli used is shown
in Fig. 8.1a, where color is used only to code for the retinotopy maps shown on in-
flated cortical surfaces in Fig. 8.1c, d. Figure 8.1b shows an example of the weighted
minimum-
l
2
-norm solution (thresholded at 0.1 of absolute maximum) of the event-
related field at 100 ms poststimulus onset (lower right visual quadrant). Note how
this activity is very distributed. To visualize the maps for different quadrants simul-
taneously and contrast them, we normalized these maps by their absolute maximum,
thresholded them at 0
9, and color-coded the activity based on which quadrant had
the maximal activity on each source point.
Figure 8.1c shows maps produced by distributed methods (from top to bottom):
(1) weighted minimum-
l
2
-norm (
.
κ
=
0
.
5); (2) dSPM (
κ
=
0
.
8); (3) sLORETA
(
0); and (4) matched filter. The first three had dipole orientation constraints,
but the matched filter did not. Figure 8.1d shows maps produced by sparse meth-
ods (from top to bottom): (1) SBL; (2) multiscale SBL; (3) MAP with
p
κ
=
0; (4)
multiscale MAP with Laplacian prior. In contrast to the distributed estimates, the
sparse estimates were not changed much by thresholding, as expected. The maps
produced by the different methods show some minor differences (related to depth-
bias compensation and sparsity), but all maps show the basic expected pattern for
V1/V2 retinotopy. The fact that retinotopy was discriminated with the thresholded
maps suggests that thresholding can be very useful for distributed estimates since
these have maxima with little localization bias. Interestingly, the simple matched
filter showed a clear map consistent with the V1/V2 borders.
=
8.9 Conclusion
The relative strengths of different localization algorithms offer an opportunity to se-
lect the most appropriate algorithm, constraints, and priors for a given experiment.
If one expects only a few focal sources, then dipole fitting algorithms may be suffi-
