sampling and reconstruction schemes for biomagnetic sensor arrays
TRANSCRIPT
Sampling and reconstruction schemes for biomagnetic sensor arrays
This article has been downloaded from IOPscience. Please scroll down to see the full text article.
2002 Phys. Med. Biol. 47 N239
(http://iopscience.iop.org/0031-9155/47/18/401)
Download details:
IP Address: 140.164.31.174
The article was downloaded on 01/10/2012 at 10:51
Please note that terms and conditions apply.
View the table of contents for this issue, or go to the journal homepage for more
Home Search Collections Journals About Contact us My IOPscience
INSTITUTE OF PHYSICS PUBLISHING PHYSICS IN MEDICINE AND BIOLOGY
Phys. Med. Biol. 47 (2002) N239–N248 PII: S0031-9155(02)36007-X
NOTE
Sampling and reconstruction schemes for biomagneticsensor arrays
Adele Naddeo1, Stefania Della Penna2, Ciro Nappi3, Emanuele Vardaci1
and Vittorio Pizzella2
1 Dipartimento di Scienze Fisiche, Universita di Napoli “Federico II”, Napoli, Italy2 Istituto di Tecnologie Avanzate Biomediche, Universita di Chieti “G. D’Annunzio”,
Chieti Scalo, Italy3 CNR-Istituto di Cibernetica “Eduardo Caianiello,” Pozzuoli (NA), Italy
E-mail: [email protected]
Received 19 April 2002
Published 5 September 2002
Online at stacks.iop.org/PMB/47/N239
AbstractIn this paper we generalize the approach of Ahonen et al (1993 IEEE Trans.
Biomed. Eng. 40 859–69) to two-dimensional non-uniform sampling. The
focus is on two main topics: (1) searching for the optimal sensor configuration
on a planar measurement surface; and (2) reconstructing the magnetic field
(a continuous function) from a discrete set of data points recorded with a finite
number of sensors. A reconstruction formula for Bz is derived in the framework
of the multidimensional Papoulis generalized sampling expansion (Papoulis
A 1977 IEEE Trans. Circuits Syst. 24 652–4, Cheung K F 1993 Advanced
Topics in Shannon Sampling and Interpolation Theory (New York: Springer)
pp 85–119) in a particular case. Application of these considerations to the
design of biomagnetic sensor arrays is also discussed.
1. Introduction
In the last few years many research groups from all over the world have developed large
multichannel arrays for magnetoencephalography(Hamalainen et al 1993, Pizzella et al 2001,
Neuromag Ltd. 1997, Yoshida et al 2000, Della Penna et al 2000, Pasquarelli 2000)4,5. These
systems are designed to cover the whole brain, so the key in the optimization is to know the
most efficient geometry of the pick-up coil of each SQUID sensor, and how to dispose such
sensors on the measurement surface to get the optimal reconstruction of components of the
magnetic field outside the skull and the location of the current sources in the human brain (in
the least-squares sense).
4 For a description and specifications of all 4D neuroimaging systems see webpage: www.4dneuroimaging.com.5 For a description and specifications of all CTF Systems Inc. systems see webpage: www.ctf.com.
0031-9155/02/180239+10$30.00 © 2002 IOP Publishing Ltd Printed in the UK N239
N240 A Naddeo et al
In this paper we assume the simplest geometry for the pick-up coil of each sensor, a square
magnetometer, and a planar measurement surface but the sampling points on this surface are
not regularly spaced. So we discuss the overall geometry and the minimum density of such
a sensor array in a more difficult and more general context than Ahonen et al (1993), i.e.
for a non-uniform sampling grid. In section 2 we sketch briefly the pattern and the spatial
frequency content of biomagnetic fields and then we introduce the Shannon sampling theorem
(Shannon 1949) and the interplay between white noise and aliasing. The Papoulis generalized
sampling expansion (Papoulis 1977) and its multidimensional extension (Cheung 1993) are
briefly reviewed in section 3, then we apply these tools to the design of our non-uniform sensor
array (Naddeo 2001). Finally, in section 4 a few comments are made and some applications
are discussed, then a new proposal of a measurement module for a MEG multichannel system
is presented (Naddeo 2001).
2. Pattern and spatial frequency content of biomagnetic fields
The first step to carry forward the optimization of a biomagnetic sensor array is to build up a
scheme for the faithful reconstruction of a continuous field from a discrete and finite set of its
values on the measurement surface. This problem is a general one; to solve it a useful starting
point is the Shannon sampling theorem (Shannon 1949) which we rewrite, following Ahonen
et al, in a form suitable to deal with our case: the distance between two adjacent samples
must be smaller than half of the shortest spatial wavelength present in the detected quantity
(the radial component Bz of the magnetic field). If the intersensor distance is consistent with
this constraint, then there is no aliasing and we can reconstruct the magnetic field on the
measurement surface. In the following sections we shall see how this request can be cast in
such a form to be employed in our particular case.
In addition to the Nyquist frequency there is another critical spatial frequency (Ahonen
1993), fcr, which we must consider when we make measurements with such systems; this
frequency depends on the signal-to-noise ratio and above its value the spectrum of Bz falls
below the white noise background, taken uncorrelated and of instrumental origin. So it is a
nonsense to try to avoid aliasing at spatial frequencies higher than fcr, which can be seen as
the ultimate limit for every intersensor spacing (Ahonen 1993).
In the following we assume a white noise level such as to fulfil the previous request, so
we concentrate on aliasing. An essential step forward is the calculus of the magnetic field due
to the model source assumed on a planar measurement surface and the analysis of its spatial
frequency content.
2.1. Spatial distribution of biomagnetic fields
The biomagnetic inverse problem, in general, does not have a unique solution; we can constrain
it to have a unique solution only if we choose a suitable source model and volume conductor
model. Extending the concept of the equivalent current dipole model (Hamalainen 1993), we
see that all magnetic activity of the examined tissue can be modelled as a sum of multiple
simultaneous or sequentially activated current dipoles; thus the conclusions found for a current
dipole source are true for a generic electrophysiological activity. Also the analysis of the spatial
frequency content of biomagnetic fields reduces to that of the dipolar field.
So, let us consider a current dipole �Q ≡ (Qx,Qy , 0), pointing along the straight line of
equation y = x and located at −→r0 ≡ (0, 0, d), inside a semi-infinite conducting medium with
z < 0; the formula (Sarvas 1987)
�B(�r) =µ0
4πK2· [ �Q × �a · −→ez
�∇K − K−→ez × �Q] (1)
Sampling and reconstruction schemes for biomagnetic sensor arrays N241
Figure 1. Vertical component of the magnetic field, Bz, on a plane at a distance z = D due to a
current dipole �Q embedded in a semi-infinite conducting medium.
where K = a(a + �a · −→ez ), �a = �r − −→r0 , a = |�a|, �∇K = (2 + a−1 �a · −→ez )�a + a−→ez , gives us the
magnetic field of such a dipole on a planar measurement surface z = D. In the particular
case we are interested in, the neuromagnetic one, if we take into account the thicknesses
of the dewar walls, the scalp and the skull, it is sufficient to assume a dipole depth d =−0.005 m and to compute the vertical component of the magnetic field, Bz, on the plane
z = D, where D = 3 cm. Also we choose Qx = Qy = 10 nA m, which are plausible
values for an equivalent current dipole, so formula (1) gives the spatial distribution shown in
figure 1.
2.2. Spatial frequency content of biomagnetic fields
Formula (1) gives us the componentBz of the magnetic field, so we are ready for the calculation
of its spatial Fourier transform. If we fix the height z of the measurement plane above the
plane z = 0, Bz becomes a function of x and y alone, so we have to compute a two-dimensional
spatial Fourier transform. We do it numerically, using a two-dimensional fast Fourier transform
(FFT) algorithm (Press et al 1986); the absolute value of the Fourier transformF(Bz) is shown
in figure 2, with kx and ky denoting the coordinate axes of the momentum space.
We see clearly that Bz(x, y) is a band-limited function and the support S of F(Bz) in the
reciprocal space is isotropic, i.e. a circle of radius W = FN where FN is the Nyquist frequency.
It can also be shown (Ahonen 1993) that this limiting value becomes higher if we diminish the
distance source-measurement surface, so the highest spatial frequencies in Bz(x, y) are due
to the shallowest sources. In this paper we are not interested in the exact evaluation of this
frequency, but only in the fact that there is a Nyquist frequency, so we can apply a Shannon-like
sampling theorem.
In the following we focus on a particular configuration of sampling points on the
measurement plane, a non-uniform one, so we need a generalization of Shannon sampling
N242 A Naddeo et al
Figure 2. The absolute value of the two-dimensional Fourier transform F(Bz) of Bz ; the two-peak
structure is due to numerical computation.
theorem. In the next section we briefly review such a generalization and then we apply it to
our case.
3. Non-uniform sampling and reconstruction of biomagnetic fields
In this section we start from a regular but non-uniform array of sampling points on the
measurement plane, as shown in figure 3. In each point we suppose to have the centre
of a planar, square magnetometer coupled to a dc-SQUID (Pizzella 2001) for the detection
of the normal component of the magnetic field Bz. We also know the spatial frequency
content of Bz, i.e. the Nyquist frequency FN. The advantage of this non-uniform grid relies
on the better versatility of the sensor. Indeed, it is possible to combine the output of the
single magnetometers of the 2 × 2 blocks shown in figure 3 to obtain planar gradiometers
with constant baseline. On the other hand, a uniform grid allows this only for a planar
sensor geometry, whereas in a spherical geometry the magnetometers are not co-planar and
the baseline is not constant. As a matter of fact, the superior sensitivity of magnetometers
sometimes vanishes on account of a possible large noise source (e.g. pacemaker, implanted
electrode). In this case the use of planar gradiometers may improve the overall performances
of the biomagnetic sensor.
The non-uniform sampling grid shown in figure 3 needs a suitable generalization of
Shannon sampling theorem, so we first introduce Papoulis’ GSE (Papoulis 1977) and its
multidimensional extension (Cheung 1993), and then apply it to our problem.
3.1. GSE and multidimensional GSE: a note
In 1977 Papoulis (1977) showed that a band-limited signal can be exactly reconstructed from
samples of the response of m linear time-invariant (LTI) systems if the sampling rate is 1/m
times the Nyquist rate. So, in the framework of a ‘m-order’ GSE we get the continuous
Sampling and reconstruction schemes for biomagnetic sensor arrays N243
Figure 3. Configuration of the sampling points of the component Bz on the measurement surface.
function f (x) univocally from m sets of samples {{gk(nT )}|k = 0, . . . ,m − 1}, where each
signal gk(x) is the response of an LTI system, hk(x), to the input f (x):
gk(x) = hk(x) ∗ f (x) k = 0, . . . ,m − 1. (2)
The sampling step T is such that T = mTN where TN is the Nyquist value, so the sampling rate
of each gk(x) is 1/m times the Nyquist rate: the whole rate is the Nyquist rate. If Hk(ν) is the
Fourier transform of hk(x), putting m = 1 and H1(ν) = 1 we recover the ‘classic’ Shannon
sampling theorem (Shannon 1949). Also in this case we can write a reconstruction formula
Shannon-like (Papoulis 1977) for the signal f (x):
f (x) =
m−1∑
k=0
+∞∑
n=−∞
gk(nT )yk(x − nT ) (3)
where we have an interpolation function yk(x) for each set of samples.
We clearly see that the power of Papoulis GSE lies in the freedom of choosing the m LTI
systems; so we can apply this generalized sampling scheme to a wide class of problems.
The last remark concerns the multidimensional extension of Papoulis’ approach (Cheung
1993) which we present in great detail in the following, applied to our case. Also in this
framework, the signal f (�r) in input to L LTI systems can be reconstructed by an interpolation
formula such as (3), i.e.
f (�r) =
L−1∑
i=0
∑
�n
gi( V g �n)yi(�r − V g �n). (4)
Now we are ready to discuss the non-uniform sampling of biomagnetic fields.
3.2. Sampling and reconstruction of biomagnetic fields on a square grid
We start from the regular but non-uniform sampling grid shown in figure 3: in each point
we have a measure of the component Bz which is a band-limited function with Nyquist
frequency FN. The aim of this section is to show (Naddeo 2001) that we can reconstruct the
normal component Bz of biomagnetic field from its non-uniform samples, i.e. from the values
Bz1, Bz2, Bz3 and Bz4 recorded by the magnetometers 1, 2, 3, 4 of each 4-bunch (see figure 3).
N244 A Naddeo et al
1
4
Bz1
2
3
Bz2
Bz3Bz4
Bz
h0
h1
h2
h3
Bz1
Bz2
Bz3
Bz4
Figure 4. Equivalence between a ‘4-order’ two-dimensional GSE and the 4-bunch of the given
non-uniform sampling grid.
ky
C0 S
kx
ky
kx
SC
0
(a) (b)
Figure 5. (a) Basic unit cell C0 and support S in the reciprocal space K2—square sampling
geometry; (b) Hexagonal sampling geometry (Cheung 1993).
The following remark is crucial: we can consider the four values Bz1, Bz2, Bz3 and Bz4 as
outputs of four LTI systems, i.e. as response to the input signal Bz, so the 4-bunch of samples
implements a ‘4-order’ two-dimensional GSE as shown in figure 4 (Naddeo 2001).
Now, each value Bzi, i = 1, . . . , 4, is evenly sampled on the measurement plane with a
sampling step Ti ≡ TN = 12FN
; also, the support S of F(Bz) in the reciprocal space K2 is
isotropic, i.e. a circle with radius W = FN . The simplest sampling geometry we can conceive
in this context is fixed through the diagonal sampling matrix
V =
( 12W
0
0 12W
)
. (5)
It corresponds to a square lattice with a square basic unit cell C0, 2W -sided, such that the
support S of F(Bz) is entirely embedded in it, as shown in figure 5(a). This condition is a
crucial one for the sampling scheme just outlined to be well defined (Cheung 1993).
We can divide each side of C0 into l = 2 identical partitions, so we get L = 2N=2 = 4
identical subcells, all geometrically congruent to the basic unit cell but with dimensions
scaled down by the same factor l = 2. Such new geometry, fixed through the square subcells
Cgi, i = 1, . . . , 4,W -sided, is implemented by the sampling matrix
V g = lV = 2V =
(
1W
0
0 1W
)
(6)
Sampling and reconstruction schemes for biomagnetic sensor arrays N245
so the corresponding periodicity matrix in the reciprocal space is
U g =1
lU =
1
2U =
1
2
(
2W 0
0 2W
)
=
(
W 0
0 W
)
. (7)
If we assume as a reference the subcell Cg0 located on the lower left-hand vertex of the unit
cell C0
�V e = −1
2
2∑
i=1
�bi = −1
2
{(
2W
0
)
+
(
0
2W
)}
= −1
2
(
2W
2W
)
=
(
−W
−W
)
(8)
we get the following region in the reciprocal space K2:
Cg0 ≡ {(k1, k2)|−W � k1, k2 < 0} (9)
where we have put kx = k1 and ky = k2. In order to fix the notation, in the following k1 and k2
denote the coordinate axes in K2 while x1 and x2 denote the coordinate axes in R2.
The four LTI systems which we introduce to give rise to the four sets of samples
Bzi, i = 1, . . . , 4, see figure 4, are defined through the following transfer functions:
H0(k1, k2) = ej2π(k1d11+k2d12) H1(k1, k2) = ej2π(k1d21+k2d22)
H2(k1, k2) = ej2π(k1d31+k2d32) H3(k1, k2) = ej2π(k1d41+k2d42)(10)
where di1 and di2, i = 1, . . . , 4, give us the offset of the sampling locations of each sample set
from the centre of the 4-bunch. If �k ∈ Cg0, the Fourier spectrum of the sampled function from
the output of the lth LTI system is
GBzl(�k) =1
|V g|
3∑
i=0
Hl(�k + Ug �q i)FBzi(�k) l = 0, 1, 2, 3 (11)
where FBzi(�k) = FBzi(�k + Ug �q i) is the ith partition of FBz(�k) shifted to Cg0. If Cg0 is the
reference subcell, we can express the ith subcell as
Cgi ≡ {�k|�k ∈ (Cg0 ⊕ Ug �q i)} i = 0, . . . , 3. (12)
�q i , which defines the position of Cgi , is a two-dimensional vector of integers ranging from 0
to 1, i.e. it is a binary representation of the integer i:
i =
1∑
p=0
qi,p2p = qi,0 + 2qi,1 i = 0, . . . , 3. (13)
So we can write explicitly the 4 × 4 matrix Hl(�k + Ug �q i) of formula (11):
H =
ej2π(k1d11+k2d12) ej2π((k1+W)d11+k2d12) ej2π(k1d11+(k2+W)d12) ej2π((k1+W)d11+(k2+W)d12)
ej2π(k1d21+k2d22) ej2π((k1+W)d21+k2d22) ej2π(k1d21+(k2+W)d22) ej2π((k1+W)d21+(k2+W)d22)
ej2π(k1d31+k2d32) ej2π((k1+W)d31+k2d32) ej2π(k1d31+(k2+W)d32) ej2π((k1+W)d31+(k2+W)d32)
ej2π(k1d41+k2d42) ej2π((k1+W)d41+k2d42) ej2π(k1d41+(k2+W)d42) ej2π((k1+W)d41+(k2+W)d42)
(14)
and, thus, H T .
Formula (11) becomes
G �Bz =1
|V g|H F �Bz (15)
where G �Bz =(
GBz0(�k), GBz1(�k), . . . , GBz3(�k))T
and F �Bz =(
FBz0(�k), FBz1(�k), . . . ,
FBz3(�k))T
.
N246 A Naddeo et al
Then we define the carrier vector �E ≡ (1 ej2πWx1 ej2πWx2 ej2πW(x1+x2)), so from the
following matrix equation:
H T �Y = �E (16)
we get the solution:
�Y = (H T )−1 �E. (17)
Given �Y , the four interpolation functions are obtained by
yi(x1, x2) = | V g|
∫
Cg0
Yi(�k, �r) ej2π �kT �r d�k
=1
W 2
∫ 0
−W
∫ 0
−W
Yi(k1, k2, x1, x2) ej2π(k1x1+k2x2) dk1 dk2 (18)
and our signal Bz is restored by formula (4), which we rewrite as
Bz(�r) =
3∑
i=0
∑
�n
Bzi(V g �n)yi(�r − V g �n). (19)
We can find an explicit form of the interpolation functions (18) if we keep in mind that the
matrix H T can be written asH T = H 1H 2, where
H 1 =
1 1 1 1
ej2πWd11 ej2πWd21 ej2πWd31 ej2πWd41
ej2πWd12 ej2πWd22 ej2πWd32 ej2πWd42
ej2πW(d11+d12) ej2πW(d21+d22) ej2πW(d31+d32) ej2πW(d41+d42)
(20)
and
H 2 =
ej2π(k1d11+k2d12) 0 0 0
0 ej2π(k1d21+k2d22) 0 0
0 0 ej2π(k1d31+k2d32) 0
0 0 0 ej2π(k1d41+k2d42)
. (21)
Given (H T )−1 = (H 1H 2)−1 = (H 2)
−1 ·(H 1)−1, let β ij, for i, j = 0, . . . , 3 be the (i + 1, j + 1)th
element of (H 1)−1, so (17) becomes
Yi(k1, k2, x1, x2) = e−j2π(k1di1+k2di2)
3∑
m=0
βimEm(x1, x2) (22)
where Em(x1, x2) is the (m + 1)th element of the vector �E.
The four functions (18) can be written as
yi(x1, x2) =βi0
W 2ejπW(−x1+di1−x2+di2)
sin[πW(x1 − di1)]
π(x1 − di1)
sin[πW(x2 − di2)]
π(x2 − di2)
+βi1
W 2ejπW(x1+di1−x2+di2)
sin[πW(x1 − di1)]
π(x1 − di1)
sin[πW(x2 − di2)]
π(x2 − di2)
+βi2
W 2ejπW(−x1+di1+x2+di2)
sin[πW(x1 − di1)]
π(x1 − di1)
sin[πW(x2 − di2)]
π(x2 − di2)
+βi3
W 2ejπW(x1+di1+x2+di2)
sin[πW(x1 − di1)]
π(x1 − di1)
sin[πW(x2 − di2)]
π(x2 − di2)
i = 0, . . . , 3 (23)
so we are ready to insert them in the reconstruction formula (19).
Sampling and reconstruction schemes for biomagnetic sensor arrays N247
Figure 6. Scheme of our MEG detector unit.
Also, given the value W from the analysis of the spatial frequency content of the
component Bz, the matrix V g in (6) defines a square sampling grid; the sampling step T = 1W
is equal to the spacing between the central points of the 4-bunches in figure 3. We emphasize
that T = 2TN, so the grid obtained gives us a larger sensor spacing without aliasing. We could
introduce another sampling geometry, an hexagonal one, and obtain a measurement array even
larger, but in this case the support S of F(Bz) is not entirely embedded in the basic unit cell
C0, as shown in figure 5(b); so the reconstruction scheme just outlined is not well defined
(Cheung 1993). In particular, it is impossible to recover the continuous function Bz directly
from the outputs Bzi( V g �n) through an interpolation formula like (4), so further generalizations
of this already non-simple approach would be needed (Cheung 1993). Then, the non-uniform
sampling grid we choose on the planar measurement surface is a square one.
4. Conclusions
The formal considerations just developed can be applied to the design of biomagnetic sensor
arrays. We see that, when the measurement points are arranged in a non-uniform pattern, i.e.
as 4-bunches on a planar surface, the square array is chosen: it allows the largest possible
sensor spacing without aliasing, leading to a well defined reconstruction scheme. Also it is
possible to recover a continuous function, i.e. the vertical component of biomagnetic field,
from a discrete set of its values recorded on the measurement surface. The quantitative results
we can obtain in the framework of the simplest source, volume conductor and measurement
surface model presented here are qualitatively valid for more realistic geometries, for example
a spherical volume conductor and a helmet-shaped measurement surface covering the whole
head as is in the case we are interested in, i.e. the design of a neuromagnetic sensor array.
In particular, the non-uniform sampling geometry just discussed leads us to a new proposal
of a complex detector unit for a multichannel neuromagnetometer. Our detector unit comprises
four square planar magnetometers, each having an area of about 100 mm2, lying on a square
support 3 cm sided, for the measurement of Bz, as shown in figure 6. The magnetometer area
is such as to fulfil a simple criterion: to be the best compromise between the lowest magnetic
field noise level and the best spatial resolution (Granata 2002). Also, with these magnetometer
dimensions, an array of about 500 sensors, that is a reasonable number for the next generation
neuromagnetic systems, is able to cover the whole head surface. This configuration allows
us realization of planar gradiometers through a suitable software coupling of the four pick-up
coils; so we also get the two derivatives ∂Bz
∂xand ∂Bz
∂y, adding more information which can be
useful for the solution of the inverse problem.
N248 A Naddeo et al
References
Ahonen A I, Hamalainen M S, Ilmoniemi R J, Kajola M J, Knuutila J E T, Simola J T and Vilkman V A 1993
Sampling theory for neuromagnetic detector arrays IEEE Trans. Biomed. Eng. 40 859–69
Cheung K F 1993 A multidimensional extension of Papoulis’ generalized sampling expansion with application in
minimum density sampling Advanced Topics in Shannon Sampling and Interpolation Theory ed R J Marks II
(New York: Springer) pp 85–119
Della Penna S, Del Gratta C, Granata C, Pasquarelli A, Pizzella V, Rossi R, Russo M, Torquati K and Erne S N 2000
Biomagnetic systems for clinical use Phil. Mag. B 80 937–48
Granata C, Monaco A, Di Russo C and Russo M 2002 Low critical temperature superconducting quantum interference
device with integrated additional positive feedback Physica C 368 227–31
Hamalainen M, Hari R, Ilmoniemi R, Knuutila J and Lounasmaa O V 1993 Magnetoencephalography—theory,
instrumentation, and applications to non-invasive studies of the working human brain Rev. Mod. Phys. 65
413–97
Naddeo A 2001 PhD Thesis University of Naples “Federico II” (unpublished)
Neuromag Ltd., 1997 Neuromag Vector View, Setting the Standard in Magnetoneurography—product information
Papoulis A 1977 Generalized Sampling Expansion IEEE Trans. Circuits Syst. 24 652–4
Pasquarelli A, Rossi R, Bucciarelli P, Fantacone V, Ziolkowsky M and Erne S N 2000 Argos 500: a whole-head
vector-MEG Proc. 12th Int. Conf. on Biomagnetism (BIOMAG 2000, 13–17 Aug. 2000, Helsinki, Finland)
Pizzella V, Della Penna S, Del Gratta C and Romani G L 2001 SQUID systems for biomagnetic imaging Supercond.
Sci. Technol. 14 R79–R114
Press W H, Flannery B P, Teukolski S A and Vetterling W T 1986 Numerical Recipes in C—The Art of Scientific
Computing (Cambridge: Cambridge University Press)
Sarvas J 1987 Basic mathematical and electromagnetic concepts of the biomagnetic inverse problem Phys. Med. Biol.
32 11–22
Shannon C E 1949 Communications in the presence of noise Proc. IRE 37 10–21
Yoshida Y, Arakawa A, Kondo Y, Kajihara S, Tomita S, Tomita T, Matsuda N and Takahashi Y 2000 129 channel
vector neuromagnetic imaging system Biomag 96: Advances in Biomagnetism Research ed C Aine et al (New
York: Springer) p 154–7