Signal Processing - January 2016 - 30

three additional spikes in the 800-millisecond segment, was used
here as the poststimulus period, and a separate, spike-free segment of
equal length was used as the prestimulus period. Figure 7 shows
SAKETINI's performance estimating single spikes relative to a spike-
free prestimulus period. Figure 7(a) shows the raw sensor data for
the segment containing the marked spike. Figure 7(b) shows the
location of the equivalent-current dipole (ECD) fit to 20 spikes from
this patient. Figure 7(d) shows the SAKETINI likelihood map based
on the data in Figure 7(a); the peak is in clear agreement with the
standard ECD localization. Figure 7(c) shows the time course esti-
mated for the likelihood spatial peak. The spike at 400 milliseconds is
clearly visible; this cleaned waveform can be used by a clinician for
analyzing the peak shape. Figure 7(e) and (f) shows a source signal's
time course from a randomly selected location far from the epileptic
spike source, to show the low noise level and to show the absence of
cross talk onto source estimates elsewhere.

× 1012

1

0.5
0
-0.5
-1

1
Normalized
to Max Intensity

1

Normalized
to Max Intensity

Magnetic Field (T)

RECONSTRUCTION OF
LOW SNR MEG DATA WITH SAKETINI
The SAKETINI algorithm can be used to reconstruct MEG data
under a low-SNR condition [20]. In one MEG data set

for localization of the primary somatosensory cortex (S1), a small
diaphragm was placed on the subject's right index finger and was
driven by compressed air, and the stimulus was delivered 256
times every 500 milliseconds. If we limit the available data to only
a small subset of trials, the lower SNR issue became limiting for
most benchmark reconstruction algorithms. We first applied
SAKETINI to the average of all 256 trials to assess performance for
the high-SNR case. We then applied it to the average of only the
first five trials and across other sets of five-trial averages.
Figure 8(a) shows typical somatosensory evoked MEG data, with
the largest peak at 50 milliseconds, expected to be coming from S1
in the posterior wall of the central sulcus. Figure 8(b) shows per-
formance of SAKETINI for the high-SNR data, which can be accu-
rately localized to the contralateral S1. Figure 8(c) shows the
sensor data averaged over the first five trials of the same data set.
Therefore, SAKETINI consistently localizes S1 even in these low-
SNR data, whereas the benchmark algorithms [minimum vari-
ance adaptive beamforming (MVAB) and standardized
low-resolution brain electromagnetic tomography (sLORETA);
results not shown-see [20] for details] perform poorly in low-
SNR regimes.

0.5
0

-0.5

0

200 400 600
Time (milliseconds)

-1

0.5
0

-0.5

0

200
400
600
Time (milliseconds)

-1

0

200
400
600
Time (milliseconds)

(a)

(c)

(e)

(b)

(d)

(f)

[FIG7] Localization of interictal spike MEG data with saKEtInI. (a) MEG sensor waveforms for a 2-second segment of data from an
epilepsy patient with interictal spikes. (b) three orthogonal views of the Mr imaging (MrI) for this patient, and a cluster of dipoles fit
by traditional dipole fitting methods. (c) the time course of a voxel showing interictal spiking activity. (d) three orthogonal views of
the MrI and the overlay of source activity estimated by saKEtInI for this data. the source localization cluster is consistent with the
more traditional methods of localization. this interictal source localizes to the margins of a prior resection cavity seen in the MrI.
(e) time-course of a voxel farther away from the interictal source. the voxel does not show clear spiking activity. (f) the location of this
"control" voxel on the patient's MrI. these results demonstrate the sensitivity of saKEtInI in spike localization as well as the source
time course estimation. (Figure adapted with permission from [20].)

IEEE SIGNAL PROCESSING MAGAZINE [30] jANuARy 2016
Table of Contents for the Digital Edition of Signal Processing - January 2016
