Biomedical Engineering Reference
In-Depth Information
Table 3.2.
Algorithm comparison: percentage of success and mean error
Algorithm
Number
Number
Correct position (%) to within
Mean error
of trials
of dipoles
5 mm
10 mm
15 mm
(mm)
PT
30
2
90
100
100
3.0
PT/RJ
30
1-4
73
92
98
4.5
SA
30
2
53
83
92
6.3
Metropolis
10
2
35
55
70
14.5
Table 3.3.
Influence of noise level on PT and SA
Total SNR
SA
PT
SA
PT
(dB)
valid runs
valid runs
mean error
mean error
(%)
(%)
(mm)
(mm)
15.8
100
90
8.7
7.0
10.1
70
90
9.5
9.4
7.8
45
95
10.6
11.5
4.1
0
15
-
17.7
dipoles were simulated with both white noise (20 dB SNR) and neuromagnetic
noise (5, 10, 15, and 20 dB SNR). Sources were restricted to the brain volume,
and constrained to the cortex (approximated by the outer 5 mm of the brain)
by a 1 to 1000 probability ratio. Solutions explaining at least 95% of the
total variance were considered as valid. Although SA and PT gave equivalent
results for a low noise level (16 dB SNR), SA results degraded quickly when
the noise level increased (45% of valid runs for 8 dB SNR). On the contrary,
PT was little affected by the noise level, and it was only for the very high
noise level (4.1 dB SNR) that the PT performance suddenly degraded.
We also applied the PT algorithm to real MEG data obtained in diffe-
rent conditions (auditory evoked fields, somatosensory evoked fields [47,56])
and with signal-to-noise ratios ranging from 5.4 dB to 20 dB. For AEF, ma-
gnetic measurements were recorded by 148 axial magnetometers (baseline =
50 mm) and averaged on 100 runs. Measurements were analyzed at a latency
of 92 ms. The noise standard deviation was equal to 15 fT, for a SNR of
20 dB. The prior probability for the source positions was the same as in
the previous study. The results clearly showed two dipoles located in the
left and right primary auditory cortices, and that explained 95.7% of the
measurement variance (Fig. 3.37). In the second experiment, SEF data were
obtained after electrical stimulation of the right index, below the movement
