You are on page 1of 31

ULTRASONIC IMAGING 1, 154-184 (1979)

uLTmsoNIc REFLECTIVITY TOMOGRAPHY: RECONSTRUCTION WITH


CIRCULAR TRANSDUCER ARRAYS

Stephen J. Norton' and Melvin Linzer

Center for Materials Science


National Bureau of Standards
Washington, DC 20234

An analysis is presented of backprojection methods for reconstructing


cross-sectional images of ultrasonic reflectivity from scattering measure-
ments. A circular array of transducer elements is considered, using three
basic modes of data acquisition and image reconstruction: (1) the same
element serves as transmitter and receiver and data is backprojected
along circular paths centered at the element; (2) distinct transmitter and
receiver with fixed separation and backprojection along elliptical paths
with the elements at the foci; and (3) distinct transmitter and receiver
with varying separations and backprojection along corresponding elliptical
paths. The point spread function (PSF) for each of these cases is derived
and is shown to depend on the shape of the acoustic pulse used. PSF's are
evaluated for three different pulses -- a narrowband, wideband, and an
analytically-derived optimum pulse which yields the best sidelobe response
and a mainlobe width equal to 0.3)1,, where 1, is the wavelength correspon-
ding to the cutoff frequency of the pulse. When backprojection is performed
along elliptical paths, the mainlobe width is shown to be broadened by a
factor proportional to the cosine of half the angle subtending the transmit
and receive elements at the center of the array. The close analogy between
the techniques used here to reconstruct reflectivity and the convolution/
backprojection method used in computerized x-ray tomography is discussed
in detail. Salient properties predicted by the analytically-derived PSF's
are confirmed in computer simulations. The characteristics of the PSF's
are also examined as a function of the number of array elements, the
location of a reflecting point in the object and the shape of the ultrasonic
pulse.

Key words: Backprojection; computerized tomography; image reconstruction;


imaging; reflectivity; scattering; ultrasonics.

I. INTRODUCTION

The resolution of conventional imaging systems is dependent on the


extent of the aperture over which rays to and from the object are inter-
cepted as well as on the distance of the aperture from the object. These
parameters effectively determine the maximum angle of energy transmission to
and reception from a point in the object. An infinite aperture will inter-
cept rays over 180° and will yield wavelength-limited resolution. Similar-
ly, a circular transceiving array surrounding an object will intercept all
rays within the plane of the circle and thus should also provide optimum
resolution within the plane. This geometry is particularly suitable for
imaging objects with roughly circular symmetry, such as the human breast.

1 This work was performed while S. J. Norton was a National Research Council
Postdoctoral Research Associate at the National Bureau of Standards. He is
now on the staff of NBS.

0161-7346/79/020154-31$02.00/O
Copyright @ 1979 by Acndrmic Press. Inc.
All rights of‘rrpwduction in any .fornt rrsenvd. 154
ULTRASONIC REFLECTIVITY TOMOGRAPHY

In this paper, we present a comprehensive theoretical analysis of a


range of ultrasonic imaging systems incorporating such circular arrays.
The arrays consist of point transducers evenly distributed around a circle,
each of which may act as transmit and receive elements. On transmit, a
point transducer will emit circular waves within the plane of the array.
On receive, the output of an element at any particular time is proportional
to the integral of the reflectivity from all scatterers lying along either
a circle or ellipse, depending on whether the same or a different element
is used for transmit. By applying backprojection techniques, analogous
to those used in computerized x-ray tomography [l], one can reconstruct a
reflectivity image of the interior of the circle.

The problem of using pulse-echo data to reconstruct a cross-sectional


picture of acoustic reflectivity has recently been considered by a number
of authors. Johnson, et al. [2,3], for example, have published their
own analysis of this problem and have obtained reconstructions using
experimentally-acquired pulse-echo data. The theoretical approach taken
by Johnson, et al., in a recent paper [41 is fundamentally algebraic
and emphasizes iterative techniques. They, however, were among the first
to point out the close analogy between the solution to the broadband reflec-
tivity reconstruction problem (which they have termed "synthetic focusing")
and the well-known convolution/backprojection method used for many years
in the field of x-ray tomography. Others have also noted this analogy,
among them Corl, et al. [5] and Norton and Linzer [6].

The analogy to convolution/backprojection invoked in the published


work referred to above is generally presented in a heuristic manner with
little analytical development or justification. In this paper we attempt
to develop from basic principles an analytical foundation for the convolu-
tion/backprojection approach to the problem of reconstructing acoustical
reflectivity with broadband pulses. In fact, it is the symmetry of the
circular array system that makes such an analytical approach feasible. As
we shall see, to proceed with a tractable development, some approximations
will be required which are analogous to the paraxial approximation in
optics. The approach developed in this paper is, however, both straight-
forward and intuitive and leads directly to a convolution/backprojection-
like technique for reconstructing reflectivity over a circular domain.

Using this approach we derive an analytical expression for the


convolution/backprojection point spread function (i.e., the reconstruction
of a point reflector). The expression for the point spread function (PSF)
is useful because it provides a good indication of the maximum theore-
tical resolving capability of the reconstruction technique. For example,
we will see from the form of the PSF that resolving power is bandwidth-
limited only; it does not depend on the physical dimensions of the array,
in contrast to imaging systems of a more conventional type where resolution
is generally aperture-limited. As explained above, this is a consequence
of the fact that the recording "aperture" extends completely around the
object and, as a result, is sensitive to all possible spatial frequency
components generated by the object. We also show that the minimum width
of the PSF (and hence maximum resolving power) is achieved in the case of
backprojection along circular paths. Moreover, in the case of elliptical
path backprojection, the width of the PSF increases (and hence resolving
power decreases) as the reciprocal of the cosine of half the angle subten-
ding the transmitter and receiver at the center of the array. In the
design of an imaging system, this property will play a role in defining a
practical limit on the maximum separation between pairs of transmitting
and receiving elements.

15.5
NORTON AND LINZER

Finally, the results of computer simulations are presented which show


how the characteristics of the reconstructed PSF depend on the form of the
acoustic pulse employed. Three different pulses are considered: narrow-
band, broadband, and an analytically-derived optimum pulse which yields a
PSF with sidelobe levels that are clearly superior to the other cases and a
mainlobe width equal to 0.3A , where A, is the wavelength corresponding to
the cutoff frequency of the &lse. The difference in the PSF mainlobe width
corresponding to circular versus elliptical path reconstructions, as noted
above, is also confirmed in the simulations. A PSF figure-of-merit is
defined in terms of the ratio of the peak to average sidelobe level of the
PSF (somewhat analogous to an antenna gain). This is plotted as a function
of the number of array elements for the optinum PSF so as to provide a
measure of the deterioration in sidelobe response as the element number is
reduced. In addition, we present a PSF reconstruction for a point off-center
which shows that while the width of the mainlobe increases to some degree,
the shape of the mainlobe remains relatively unaffected by moving the target
away from the center of the array.

Recently, the authors reported on the derivation of an exact, closed-


form analytical solution to this reconstruction problem for the case of
circular integration paths [6]. The analytical inversion formula obtained
in that work is exact in the sense that the development requires no paraxial-
like approximations. This formula, which was derived using a different mathe-
matical approach (circular-harmonic decomposition), can be shown to reduce
to the simpler backprojection-like solution obtained in this paper
in the limit when the paraxial approximation (to be discussed in
a later section) becomes exact. The detailed derivation of the exact
solution will be published elsewhere [7]. The exact inversion formula is,
of course, of interest because, as noted, no approximations were invoked in
its derivation; however, two advantages of the present approach are its
simplicity and its amenability to an intuitive interpretation of the recon-
struction process. This is important because these traits help make the
backprojection technique a particularly easy one to implement on a computer.
Furthermore, the backprojection solution presented here is easily extended
to three dimensions (reconstruction over a spherical rather than circular
domain) and to corrections for velocity inhomogeneities. These modifica-
tions have already been carried out in our laboratory and will be reported
in later papers in this series.

II. BASIC APPROACH

The geometry of the imaging systems considered here is shown in


figure 1. A circular array of evenly-spaced transducer elements encloses
the object of interest. The interior of the circle bounded by the array
defines the cross-sectional region through the object to be imaged. Each
array element can be utilized in both a transmit and receive mode. The
element beam patterns are assumed omnidirectional in the plane of the
array but collimated in the direction perpendicular to this plane (i.e.,
the element patterns exhibit a "fan beam" shape, diverging within the
interior of the array but flat normal to the interior).

Ultrasonic echo data can be generated by the transducer array in


three basic modes which we consider in turn below. More complex modes
can be constructed using combinations of these three.

The simplest mode of acquiring reflectivity information from the object


consists of using each element individually to generate an A-scan (fig. la).
The A-scan is produced by emitting a spatially-diverging pulse from one
array element and recording the resulting echoes with the same element as
a function of time. This pulse-echo process is repeated for each element

156
ULTRASONIC REFLECTIVITY TOMOGRAPHY

time time

0 2plc (P, +p,vc

Fig. 1 (a) Operation of circular array with same element serving as


transmitter and receiver (mode 1). Each point in the
A-scan is the sum of echoes arising from scatterers along
a circular arc centered at the active element. (b) Operation
of circular array with separate transmitter and receiver
elements (modes 2,3). Each point in the A-scan is the sum
of echoes arising from points lying along an elliptical arc
whose foci are the transmitter and receiver.

in turn; the result is N A-scan-like traces, where N is the number of


elements in the array. In this mode of operation (designated “mode l"),
only a single element is active at a time. Because the pulse is allowed to
spread out uniformly within the plane of the array, a given point on one
A-scan trace corresponds to the sum of echoes originating from all scat-
terers lying along a circular arc whose center is the active element. This
pulse-echo process, when performed for all N elements, can thus be inter-
preted as generating intersecting line integrals of reflectivity defined
over families of concentric circular arcs centered at each of the N
elements. Using the line-integral measurements generated in this fashion,
one can then attempt to reconstruct the unknown reflectivity function for
the object. Because of the circular nature of the integration paths, this
method of acquiring the pulse-echo data leads to the simplest reconstruction
analysis of the three data acquisition modes considered here.

The second basic mode of acquiring data ("mode 2") consists of trans-
mitting from element T and receiving the resulting echoes from element R
(distinct from T) with fixed angular separation 2B, as shown in figure lb.
This process may be regarded as generating line-integrals of reflectivity
over a family of concentric elliptical arcs whose foci coincide with the
transmitting and receiving points T and R. The pulse-echo process is
repeated N times for each pair of array elements separated by the angle
2B, i.e., for fixed 6, each element in the array is used once as a trans-
mitter and once as a receiver. Again, the total process generates N
A-scan-like functions of time which constitute the raw input data used in
the reconstruction procedure.

The third mode of data acquisition considered here ("mode 3") consists
of transmitting and receiving with pairs of elements of varying angular
separations. If all possible elements taken in pairs, one as transmitter
and the other as receiver, are used, this process would be equivalent to
transmitting from element k and receiving simultaneously with all N elements
(including element k), and then allowing k to range between 1 and N. The
same result would be obtained by allowing 26 in a mode 2 system to range

157
NORTON AND LINZER

between -IT and 71, and then coherently superimposing the resulting set of
mode 2 images. Howeves one chooses to view it, this data acquisition mode
generates a total of N A-scan-like waveforms for subsequent processing,
instead of N waveforms, as is the case for modes 1 and 2. This method may
be thought of as generating line-integral data over all possible elliptical
paths intersecting the object in the plane of the array with foci corres-
ponding to all possible pairs of transducer elements on the array.

In the following sections, the same image reconstruction algorithm


is shown to result for both modes 1 and 2, subject to certain reasonable
approximations needed for a tractable analysis. The derivation of the
reconstruction mathematics for modes 1 and 2 is straightforward and, as
discussed in Sections III and IV, bears a very close relation to the
familiar method of convolution/backprojection used in the classical pro-
blem of reconstruction-from-projections [l]. Finally, the third mode of
data acquisition may be regarded simply as a linear and coherent super-
position of multiple mode 2 images and, as a result, the same linear re-
construction algorithm derived for the mode 2 (and mode 1) problem is
applicable in this case as well.

To illustrate the approach to reflectivity imaging used here, suppose


we consider a mode 1 reconstruction of the reflectivity at a given point in
the interior of the array whose distance from a particular element is p. To
obtain a measure of the reflectivity at this point, we evaluate the A-scan
recorded by this element at time 20/c, where c = speed of sound. This value
represents the sum of echoes arising from scattering points lying along an arc
of radius p and centered at the element in question. This process is repeated
for each point in the A-scans recorded from all N elements (where p is always
the distance between the point under reconstruction and a given element), and
the results coherently added to yield the reflectivity value at the image
point. Each A-scan will thus contribute one value to each image point.

To reconstruct all of the points in the image in this fashion, the


amplitude of the A-scan evaluated at time 2p/c is uniformly distributed
or "smeared" back over an arc of radius p in image space. This procedure
is repeated for every point in each of the N recorded A-scans, and at each
point in image space the results are coherently added.

This process of "smearing back" and coherently summing the A-scan


traces may be termed "acoustic backprojection", in analogy, for example,
to the technique used in computerized x-ray tomography where an unknown
object is reconstructed from its x-ray projections. One essential differ-
ence between backprojection in x-ray tomography and the "acoustic back-
projection" described here is that in the latter case the recorded (reflec-
tivity) data are backprojected over circular (or elliptical) paths rather
than along straight lines. We also note that in x-ray tomography, the
(x-ray) projection data are subjected to a filtering process prior to
backprojection. This prior filtering is required, in principle, to eliminate
the sidelobe artifacts introduced by the backprojection process. These
sidelobes arise from the contributions of all the other (density) elements
lying along the integration paths which intersect the particular point
under reconstruction. In the acoustic version of this problem, we shall
show that a superior reconstruction results after the application of an
entirely analogous filtering operation on the echo data which, as a result
of the temporal nature of the "line-integrating" process, can be performed
directly in the time domain. Thus, to achieve an optimum reconstruction,
one needs to filter the raw rf A-scan data recorded from each transducer
element (or element pair) with a suitably-derived transfer function prior
to backprojection. An alternate and equivalent strategy, as we shall show,
ULTRASONIC REFLECTIVITY TOMOGRAPHY

consists of transmitting an optimally-shaped pulse waveform, i.e., a pulse


whose temporal frequency content matches the transfer function of the
prebackprojection filter. Thus, the same optimum reconstruction can in
principle be achieved in the absence of such a filter by employing the
optimum acoustic waveform and then coherently summing (backprojecting with
no prior filtering) the resulting A-mode data recorded from each element.
This "optimal" acoustic waveform is described in later sections.

There is another important difference between acoustic reflection


tomography and conventional x-ray tomography worth emphasizing. In the
former case, the line-integration process takes place automatically along
parallel concentric arcs normal to the direction of propagation of a single
outwardly-diverging pulse. A entire family of line-integrals are generated
in this way by just one transmit/receive element. Thus, with a stationary
array of elements, a great deal of separate line-integration measurements
may be generated in a nearly-instantaneous fashion. On the other hand,
x-ray tomography (or, more generally, transmission tomography of any kind,
including acoustic time-of-flight and attenuation tomography) requires the
time-consuming motion of multiple sources and detectors to acquire essen-
tially an equivalent quantity of information. Further discussion of the
analogy between reflectivity and transmission tomography is presented in
Section II-D, below.

In this paper we shall be concerned with the mathematical problem of


reconstructing reflectivity only. For the sake of a tractable analysis,
we consider a highly idealized image reconstruction problem based on the
following assumptions:

(1) The object is weakly reflecting such that the energy contained
in the propagating sound wave is much larger than the total energy back-
scattered over the propagation path (Born approximation). Second-order
reflections are insignificant.

(2) The absorptivity of the medium is essentially uniform; thus,


attenuation due to absorption can be systematically compensated for by vary
ing the receiver gain in an exponential fashion. (According to assumption
(11, attenuation due to scattering is considered negligible; consequently,
absorption is the only important contribution to the attenuation).

(3) The velocity of sound is essentially constant over the region


of interest. Specifically, there are no variations in the sound velocity
large enough to produce errors in the round-trip time delay between a
transducer and any elemental reflecting point that are a significant
fraction of the reciprocal of the signal bandwidth. This assumption
insures that the integration paths are essentially circular or elliptical,
depending on the mode of operation.

(4) The object may bemodeled as a collection of isotropic scatter-


ers, each acting as a Huygens-like source. Under this assumption, the
reflectivity of the medium, which describes the relative amplitude of a
signal backscattered by a point reflector, may be modeled as a scalar
function of space, i.e., its value at a point in space is assumed indepen-
dent of the direction from which the sound impinges upon the point. In
the limit of the Born approximation, this assumption can be shown to be
valid for the case of backscattering [8], but becomes progressively poorer
as the angle subtended by the transmitter and receiver at the target
increases [9]. This problem is currently under further study [lo].

In a real medium, all of these assumptions will, of course, break


down to some degree and some image degradation will result, depending on

159
NORTON AND LINZER

omnidirectional

Fig. 2 Single transmit/receive element


in a circular array enclosing
the object with reflectivity
function f(r,B).

the departures from ideality. However, it should be pointed out that


almost all existing acoustic imaging systems, whether using fixed lenses,
electronic focusing, computer focusing, or holography, operate implicitly
on these very same assumptions. Attenuation due to scattering is usually
neglected and time-gain-compensation circuitry employed to compensate
for an assumed medium absorptivity function. Distortions due to velocity
inhomogenieties in the medium are also not corrected for. Finally, since
all imaging systems must have some finite aperture in order to achieve
lateral resolving power, scattering is almost always observed over a range
of angles and thus leads to an ambiguous measure of the true reflectivity
of the object when isotropy does not hold.

The computer reconstruction approach discussed here actually has


decided advantages over other systems with respect to the breakdown of
these assumptions. Since the reflectivity data are digitized, they may,
in principle, be combined with tomographic data on velocity and attenua-
tion in an iterative fashion to produce highly-corrected images of all
these parameters. Even the frequency-dependent nature of the medium may
be taken into account using computer techniques. Finally, assumption (4)
is rigorously valid in the case of mode 1 imaging (in the limit of the
Born approximation), since only backscattering is observed. In the case
of mode 2 and 3 imaging systems, a reflectivity image effectively averaged
over a certain range of scattering angles would be produced. However,
since the angular separation between any pair of transmitting and receiving
elements would typically not exceed 45O (for other reasons which will
become apparent in Sec. IV), the departure from true backscattering is
small and, thus, its effect may not be noticeable.

III. IMAGE RECONSTRUCTION FROM REFLECTIVITY DATA GENERATED OVER CIRCULAR PATH

A. Acoustic Backprojection

Let the reflectivity of the object in the plane of the circular


array be given by f(r,e), where the origin of the polar coordinate system
is the center of the array, and R is the array radius. Suppose the trans-
ducer element located at angle q~ on the array emits a short pulse of sound
which diverges outward in an isotropic fashion into the interior of the
circle, as shown in figure 2. We will denote by g (t) the resultant omni-
directional echo data recorded by the same element'as a function of time.
For conciseness, we define E = tc , where c = speed of sound (thus, time
delay will be measured in units of distance rather than time; this elim-
inates the symbol c from our equations). For analytical convenience, we
assume g (E) is generated and recorded for all 4 between 0 and HIT. We now
examine t he process of backprojection (summation) as a method of reconstruc-

160
ULTRASONIC RGFLECTIVITY TOMOGRAPHY

transducer medium

(a) (b)

Fig. 3 (a) Linear system consisting of a transmitting/receiving


element and an isolated point reflector. <b) Schematic
of convolution relationship p,(z-20) = pT(t)*h(t-2P) for
linear system shown in (a).

ting the un@own reflectivity function f(r,8). If this reconstruction is


denoted by f(r,8), the backprojection of the signals g (t), without prior
filtering, can be expressed mathematically as 4

;(r,B) = 7" g4(2P(4;r,8))d4 , (1)


where 0

l/2
p($;r,e) = [R2 + r2 - ZRrcos(~-$)I (2)
is the distance between the object point (r,e) and the location of the
transmit/receive element at (R,$).

Thus, to reconstruct the reflectivity at a point (r,e), the waveform


recorded at position Q on the array is evaluated at time 2p($;r,e)/c
(which is the round-trip delay between the element and the point); this
same process is repeated fog all elements (all 4) and, for a given (r,B),
the results added to yield f(r,i!l).

B. Point Spread Function Obtained by Backprojection

The reconstruction, or image, of a point object is called the


point spread function (PSF) of the imaging system. The PSF may be regarded
as the two-dimensional analog of the impulse response of a one-dimensional
linear system. The PSF is usually characterized by a sharp lobe of non-
zero width corresponding to the position of the point object, and side-
lobes which trail off some distance from the mainlobe. The PSF is a very
useful measure of the overall performance of the imaging system because it
provides information about the inherent resolving capability of the system,
as well as insight into the problems of artifact formation and signal-to-
noise ratio.

In the following, we derive the PSF associated with the backprojection


process. To begin, we find it uSefu1 to express the shape of a received
pulse reflected from a point target in terms of the temporal impulse res-
ponse of a linear system which consists of the transmit/receive transducer,
the point target, and the intervening medium. Let h(t) denote the impulse
response of this lingar system , which is illustrated in figure 3. Suppose
a voltage impulse 6(t) is applied to the transmitter terminals, as shown
in the figure. Then, in response to the resultant echo reflected
back from a target at a distance p from the element, the received (voltage)

161
NORTON AND ISNZER

waveform will be h(E-2p). The shape of the impulse response, h(E), is in


general defined by the characteristics of the transducer and the propagating
medium. However, under the set of assumptions describing the idealized
medium in the last section, it is, in principle, possible to design a sys-
tem such that the functional form of h(t) remains essentially independent
of target position -- in particular, independent of target azimuth due to
the omnidirectionality of the source/receiver, and independent of range by
employing time-gain-compensation at the receiver to compensate for echo
attenuation due to geometric beam spreading and absorption. Thus, we shall
assume that while the shape of the function h(f) can be made invariant with
respect to target position, the amplitude of h(T) is proportional to the
target reflectivity, which is the parameter of interest here. Under these
conditions, the system is both linear and time-invariant and thusL for an arbi
trary transmitted pulse p,(t), the received pulse is given by pR(t-2p), where

pR(f-2p) = p,(f) * h(:-2p) , (3)

and * denotes convolution. Hence, knowledgE of h(t) allows one to prede-


termine the shape of the received pulse p,(t), in principle, by exercising
control over the shape of the transmitted waveform pT(t), or, alternatively,
by manipulating h(t) itself by inserting a filter at the receiver.

We now employ eq. (1) to reconstruct the image of a point reflector


of unit reflectivity which lies at (r ,6 ). From this equation we will
arrive at an expression for the pointOsp?ead function of an imaging system
based on the method of backprojection.

In the manner described earlier, the system generates echo data by


using each element in the array in a pulse-echo mode and recording the
resultant signals reflected from the target as a function of time. Suppose
the element at (R,@) is excited by a transmit (voltage) pulse p,(f); then
from eq. (3), the signal received at this element is given by

g+(E) = pR[t-2p(+;ro,80)] (4)

where 2p($;r ,0 ) is the round-trip distance between the target at (r ,e )


and the arra$ element at (R,@), and t = 0 is defined as the time the &lge
is transmitted. Then putting eq. (4) in eq. (1) gives

;(r,8) = 7" P,DP($ ;r,0) -2p(@;ro,00)ld$. (5)


0

Note that when f(r,e) is evaluated at (r ,8 ) (i.e., when the system’s


"focus'1 coincides with the position of the po?ntOtarget), eq. (5) reduces to

2(ro,eo) = 2xprpR(0) . (6)

This result suggests that it-is important to choose a pulse shape p,(t)
that possesses a maximum at t = 0.

We now derive an expression for the point spread function ?(r,e) for
arbitrary (r,8). Using eq. (2), the argument of p,(*) in eq. (5) can be
written as

2P(O;r,e) - 2P(G;ro,00) = 2 [R2+r2-2Rrcos(e-$)]"2 (7)


- 2[R2+ro2 -21Uocos(eo-~) I+

162
ULTRASONIC REFLECTIVITY TOMOGRAPHY

To carry the analysis further, we shall assume for the present that
(r/R) 2 << 1 and (r /RI2 << 1. Subject to this restriction, which we shall
later evaluate in Pore detail, the resultant PSF will yield an approxima-
tion to the true PSF which is excellent when both the target and image points
are close to the center of the array, but will generally be poorer as these
points move further away from the array center and closer to its periphery.
However, even when r and r are not much less than R , we will see later that
the functional form of theOPSF, particularly in the vicinity of its mainlobe
(where r g r-,I, does not deviate noticeably from its form under conditions
where (r/R) <cl and (r /R)' << 1. As a result, the analysis should still yield
considerable insight in ? o the resolution (mainlobe width) and other important
PSF properties (such as sidelobe behavior) that characterize the theoreti-
cal performance of the imaging system.

Expanding the square roots in eq. (7) and dropping terms higher than
second order in r/R and ro/R, yields
r2-r 2

Zp($;r,e) - 2p($;ro,eo) ; - 2rcos(b$) + 2rocos(00-$) ++

r2
- $ COS~C~-+) + $J cos2(eo-~) .

The approximation made here is quite analogous to the paraxial approxima-


tion used in optics. After some straightforward trigonometric manipula-
tion, this can be rewritten in the equivalent form

2p(+;r,e) - 2p($;ro,00) = 2xcos($-Y) + y1 + y2cos2($w) I (8)

where

X = [r2 + r2 - 2rorcos (e-eo) I 1’2 (9a)


0

rosin9 - rsine
0
tan Y = (9b)
rocOse - rcose
0

Yl = z l [ r2-r 3 (lOa)

Y2 = z ’ [rz + r4 - 2r2rtcos2 (e-0 ) 11’2 (lob)


0

r>in2e0 - r'sin28
tan a = (1Oc)
ricos2B0 - r'cos20

Note that X as given by eq. (9a) is the distance between the reflector,
and the point under reconstruction, at (r,8); this is illus-
Substituting eq. (8) into eq. (5) gives

211
Z(r,e) = f pR12xcos(4k-Y) + y1 + YzcosZ(&a)ld+. (11)
0
Now let P,(G) denote the Fourier transform of p,(t), where for conve-
nience we define V = V/c (i.e., frequency is measured in units of
reciprocal length). Thus,
03 --
p,(t) = 1 PR(?)ezanut d3 .
-0D

163
NORTON AND LINZER

Fig. 4 Reconstruction geometry. (r ,Bo)


is the position of the point0
reflector and (r,B) is the
point being reconstructed.

Putting eq. (12) into eq. (11) gives

h 2iT 02
i2~[y~+~~cos2(~~~lei4~Xcos(~-YJ (13)
f(r,f3) = 1 d$ ,f dv P,(;)e .
0 -03

To simplify this expression , we expand the first exponential in a power


series; it is convenient to divide the result into two parts, as follows:

;(r,B) = g,(r,0) + B(r,e) , (14)


where
2Tr m
i4GXcos(@-Y)
io(r,6) = / d.4 1 d< PB(\))e I (15)
0 -02

2iT O3
s(r,e) = J d$l d; PA(;)[i26c + a(i216<)2 +...le i4nGXcos($-Y)
, (16)
0 -m

and 5 = y1 + Y~cos~($-0).

The function denoted by fo(r,8) on the right in eq. (14) results from the
zero-order term in the expansion of the first exponential in eq. (13). This
term, defined by eq. (15), can be $mplified by interchanging orders of
integration with respect to 0 and V, and making use of the Bessel identity

Jo(p) = (1/2lT)jnd$ eipcos('-q) . (17)


0
The result is

i?o(r,B) = 2~r / a3 P,(j)Jo(4sX) . (18)


-co

We will now argue that the term 2 (r,(3) in eq. (14) generglly repre-
sents a much larger contribution to thg point spread function f(r,0) than
the term denoted by ~(r,el. For this reason, we will find it convenient
to approximate f(r,e) by f (r,(3), in which case E(r,e) may be regarded
as an error term. To jus_t?fy this approximation, we must examine the
assumption JE (r,e) 1 << Jf (r,(3) 1, and establish under what conditions,
particularly over what dogain of (r,8), it is valid.

164
ULTRASONIC REFLECTIVITY TOMOGRAPHY

We begin by noting from eq. (16) that

E (ro,eO) = 0 ,

since, in view of eqs. (lo), 5 = 0 when r = r and 6 = f3 ; i.e., E(r,e) is


identically zero at the peak of the mainlobe 8f the PSF.O Furthermore, eqs.
(gal, (lOa) and (lob) show that the terms x , y1 and YZ are all small
compared to R when the point under reconstruction (r,(3) is relatively
close to (r ,8 ), the point at which the actual reflection occurs. Thus,
the express?on"for ?,(r,B) should provide a good approximation to the
shape of the true PSF, f(r,e), at least in the immediate vicinity of
its peak.

The relative size of E and ^f as X becomes large compared to the main-


lobe width (i.e., when (r-,6) is f% from (r ,,8 )) is more difficult to
ascertain from simple inspection of eq. (16); Rowever, further analysis of
shis problem is important to show how well, and under what conditions,
f (r,!) approximates the asymptotic behavior of the sidelobes of the true
P!!F, f(r,e), relatively far away from the mainlobe. This analysis is
sarried out in the Appendix, where asymptotic expressions for e(r,e) and
f (r,e) are derived starting from eqs. (29) and (16). The derivation is
g?ven for a particular choice of P (3), which is developed in the
next section. A comparison of theRmagnitudes of E and z reveals that
under certain conditions E can become comparable to or l%rger than zo.
However, this only can happen for X very large relative to the mainlobe
width, in which case the sidelobes contribute only a negligible amount
to the total PSF energy.

To conclude, it is interesting to note that 1 (r,e), as expressed by


eq. (18), is independent of R. Also, it can be se& from eq. (16) that

Zo(r,B) = Lim i(r,8) ,


R-=Q
since Lim E(r,e) = 0 as R-)00, in view of eqs. (lOa) and (lob). In other
words, as the radius of the array becomes large compared to the size of
the object, the curved integration paths which intersect tha object
approach straight lines; and, in this straight-line limit, fo(r,8)
becomes the true PSF.

C. Point Spread Function ?,(r,8) for Different Pulse Shapes

Using the expression (18) for ^f (r,8), the form of the PSF can be
computed for a variety of pulse shapes. OIn particular, we consider a nar-
rowband pulse, a wideband pulse, and finally an optimum pulse which minimizes
sidelobe energy.

(1) Narrowband pulse

Suppose p,(f) iz very narrowband about a center frequency go.


Then, approximately, PR(V) = 6(g-go) and upon substitution in eq. (18),
we get

g,(r,e) = 2aJo(4tiOX) .

The distance X between the peak of the PSF and its first zero is
often used as a mea&re of the width of the mainlobe (this is also known
as the Rayleigh distance and represents a traditional measure of resolving

165
NORTON AND LINZER

power in diffraction-limited systems with circular symmetry). From eq. (19),


this distance is given by
0.19
=-I
xO
$0

or, in terms of the wavelength A0 corresponding to the center frequency Go,

X 0 = 0.191, .

This result is noticeably less than the 0.5x value which is sometimes cited
as the minimum possible diffraction-limited Besolution for a cw system with
wavelength X0. Although the mainlobe width of the narrowband PSF given by
eq. (19) is impressively narrow, the sidelobes of the J (*) function are
very large in the vicinity of the mainlobe and exhibit % extremely slow
asymptotic fall-off of Xm1i2. In practice, however, a narrowband pulse will
have some finite length, so at sufficiently large distances from the mainlobe,
the actual response will begin to depart from the analytical form given by
eq. (19). In particular, the J (0) function should accurately predict the
PSF form out to a distance of azproximately one-half the length of the
narrowband pulse.

(2) Wideband pulse

Suppose p,(f) is very wideband. Then, approximately, PR('3) = P =


constant, and eq. (18) now yields

io(r,O)
= 2sP ( Jo(4~X)d~ = p . (20)
-m 1x1
Thus, the wideband PSF falls off inversely with the distance from the target
location. This function exhibits the characteristic "volcano shape" of the
PSF obtained when the projections of a point are backprojected and summed
with no prior filtering. This behavior is well-known in the field of
computerized x-ray tomography, and is customarily regarded as undesirable
owing to the slow X-l fall-off of the PSF. However, this is an improvement
over the narrowband case, as given by eq. (19), with respect to the sidelobe
leve s both near the mainlobe as well as at large distances.

(3) Optimum pulse

We now seek the function P,(v) which upon substitution into


eq. 18) yields a delta function PSF; we will call this function pEpt(v).
That is, for a perfect reconstruction, we equate

zTt(r,e) = 2'Tr 7 d; P~t(~)Jo(47r~X)


-m
= k 6(x-xory-Y,) I (21)
where

x = rcose x = r co.540
y = rsine = rEsin0 0
y: 0 '
and k is a constant. It can be shown [ll] that the two-dimensional delta
function 6(x-xo,y-y,) can be expressed in polar coordinates as
6 IX)
6 fx-xory-Y,l
=iqYf ’
(22)
l/2
where we recall X = [r*+rt-2rrOcos(+eo)] *

166
ULTRASONIC REFLECTIVITY TOMOGRAPHY

We now invoke the Bessel function relation [12]

Since Jo(*) is an even function, the above integral can be written as

Jo(rc) lvldv = &-d(r) (23)

Replacing r with X in eq. (23), and comparing the result to eqs. (21)
and (22), we see that setting

(24)

gives the desired delta function PSF.

Any real imaging system will have a finite temporal bandwidth, and,
as a result, p"pt(c) must fall to zero beyond some frequency cutoff, say,
; 1 To take i!to account the finite bandwidth of the system, we multiply
eg. (24) by a frequency-apodizing function, denoted by A(c), which is
assumed essentially constant within the bandpass of the system but falls
rapidly to zero outside. Thus, we write

Pipt(;) = I:IA(;) (25)


where

A(;)) g 1 for lgl~ 3


C
= 0 for I!/>> cc .

A simple choice for A(i the rectangle function, rect(</2;c),


where

rect(u) = 1 for JulGl/2


=o otherwise.

In this case, the optimum pulse spectrum takes the form

PFt(G) = j3/rect(G/23c) , (26)

which, upon inverse Fourier transformation, yields the form of the


optimum impulse response:

pipt(F) = 2<csinc(2Gct) - Scsinc2(Gct) , (27)

where sin(Iru)
sine(u) = y .

opt -
The optimum pulse pR (t) and its spectrum Pi"(c), defined by eqs. (27)
and (26), are depicted in figure 5. The significance of these waveforms
will be discussed in greater detail in the next section.

Substituting eq. (25) into eq. (18) gives

?Ft(r,fJ) = 2~7 d< A(;) I<IJo(41T\;X) .


-co

167
NORTON AND LINZER

(a)

Fig. 5 (a) The optimum pulse, pEOpt(t): given by eq. (27);


(b) Fourier transform of pipt(t), given by eq. (26).

If A(\)) is an even function of j, then this can be written as


co
zTt(r,8) = 4711 A(<)Jo(4xjX)<d< , (28)
0
which we recognize as the Hankel transform of the function A(?).

Putting A(<) = rect(</2cc) in eq. (28) and integrating, results in

gc Jl(4ticX)
FFt(r,O) = x (29)

The distance X0 between the peak of this function and its first zero is
given by

= 0. = Q-30)1
V)_ xO c -
where x is the wavelength corresponding to the cutoff frequency gc. It
is integesting to compare the Rayleigh width given by eq. (30) to that
calculated earlier for the narrowband PSF. If we equate the upper cutoff
wavelength 1 with the narrowband wavelength A , we see that the mainlobe
width of theCoptimum PSF is 'U.5 times broader'than the value of 0.191,
computed for the narrowband case. However the envelope of the Jl(*)/(*)
function attenuates asymptotically as X-3/1 , which is faster than for
both the narrowband and the wideband cases, and thus represents the best
sidelobe response of the three PSF's.

It is significant to note that the width of the mainlobe (and hence


resolving power) is a function only of signal bandwidth; it does not depend
on some measure of the aperture size , as in the case of conventional imaging
systems. Thus, resolution is temporally limited only. This property is a
consequence of the fact that the system "aperture" extends completely around
the object and, as a result, the system does not behave as a spatial frequency
low-pass filter in the manner of conventional finite-aperture imaging systems.
We see also from eq. (28) that the PSF possesses circular symmetry and spatial
invariance.

D. Analogy to the Classical Convolution/Backprojection Solution

It is significant that formulas identical to eqs. (26) and (27)


were first obtained by Bracewell and Riddle [13] in their derivation of the
solution to the classical problem of reconstructing an unknown function
from its projections. In conventional tomography, the formula (27) repre-
sents one of several effective filter functions which are convolved with

168
ULTRASONIC RHFLECTIVITY TOMCXXAPHY

(a) ib)
parallel straight parallel circular
projection paths

Fig. 6 Schematic of single transmission projection (a) of conventional


tomography compared with a "reflectivity projection" (b) for the
circular case.

the projection data prior to backprojection (summation). The close rela-


tion here to the well-known technique of convolution/backprojection can be
understood intuitively from the following interpretation of the pulse-echo
process. Consider, for example, the omnidirectional transducer element
at point T/R in figure 6. As mentioned in Section II, the output of this
transducer plotted versus time (where time is proportional to range) is a
one-dimensional function proportional to the line integrals of the two-
dimensional reflecting object defined over concentric arcs centered at T/R.
This line-integral interpretation of the recorded echo data leads intuitively
to the idea that the omnidirectional echo data generated by the transducer
at, say, T/R, can be regarded as a one-dimensional "projection" of the two-
dimensional object (the function f(r,e)) from line AT/R to line BT/R. In
comparison with an x-ray projection of an unknown density function, however,
this "acoustic projection" of the unknown function f(r,6) appears distorted
because the lines AT/R and BT/R are not parallel, or equivalently, because
the integration paths through the reflecting object are curved (due to the
curvature of the diverging wavefront) rather than straight. However, if the
object is small compared to R, the lines AT/R and BT/R are nearly parallel
and the paths that intersect the object are nearly straight. Under these
conditions, one expects the solution to the reflectivity reconstruction
problem to reduce to the classical convolution/backprojection solution.

TO show this explicitly , we review here briefly the technique of


convolution/backprojection. Suppose we measure a set of one-dimensional
projections p,+(x) at all angles $ through a two-dimensional function F(r,6).
One projection of F(r,8) at angle 4 is illustrated in figure 7. Then the

Fig. 7 Schematic of single projection


of the function F(r,B) at angle $.

169
NORTON AND LINZER

process of backprojecting over straight lines, without prior convolution,


may be expressed by the formula

g(r,e)
Tp =
(rcos(&$))d@ . (31)
0 4
Thus, the number g(r,8) is obtained by adding together the values of all
the line integrals whose integration paths intersect at the point (r,0).
(In the field of reconstruction-from-projections, this process is some-
times referred to as simple summation).

Consider once again eqs. (1) and (2):

z(r,e)
= Tg [2(R2+r2-2Rrcos(8-$))"2]d~ , (1')
0 4
where g@(E) is the pulse-echo data recorded as a function of time by the
transmit/receive element located at the angle $. This formula describes
the process of backprojection over a family of concentric circles centered
at (R,+). To demonstrate the relationship between eqs. (1') and (31), let
us assume that the maximum spatial extent of the object, centered at r = 0,
is much less than RI. the radius of the array. Expanding the square root in
the argument of g (t) in eq. (1') and dropping terms higher than first order
in r/R yields +
2lT
;(r,e) = 1 g [2R-2rcos(&$)ld$ . (32)
0 @

Thus, if we interpret g (t) as "projection data" and equate


4
gGIWWl = P (x) ,
+
we see that the formulas (31) and (32) are identical, aside from an unim-
portant scale factor and a shift in origin of R in the acoustic "projec-
tion" g+(t). This result is not surprising since the approximation implied
by expanding the square root in the argument of g+(O) and keeping only the
linear terms in r/R corresponds to replacing the circular integration
paths with straight lines. As is well-known in computerized tomography,
however, the process of backprojecting the unprocessed projection data
will not in general yield a satisfactory reconstruction. Analysis shows
that a superior reconstruction can be achieved by first convolving (fil-
tering) each projection p@(x), prior to backprojection, with a suitably
derived filter function c(x). An example of such a filter function, first
derived by Bracewell and Riddle [13] and later by Ramachandran and
Lakshminarayanan [141, takes the form

c(x) = ZMsinc(2Mx) - Msinc2(Mx) , (33)

where M is a constant. As a result, the ideal reconstruction is obtained


by backprojecting the filtered projections, which we denote p"pt(x), where
@
opt (x) = c(x) I (34)
P@ * P+(X)

instead of the unfiltered p (x), as indicated


projections in eq. (31).
0
It is noteworthy, as mentioned earlier, that eq. (27), the optimum
acoustic pulse shape, has the same form as the Bracewell and Riddle convo-
lution kernel c(x) given above. This result could have been hypothesized
by analogy, which many authors have done [2,3,6), although this is strictly
justified only under the assumption that the integration paths are straight
lines, or nearly so. In addition, in the acoustic version of the recon-

170
ULTRASONIC REFLECTIVITY TOMOGRAPHY

struction-from-projections problem, it should be emphasized that the


received waveform g (f) in eq. (1') is actually the true "acoustic projec-
tion" (i.e., the re ii ult of line-integrating the function f(r,81 along
parallel-circular paths) convolved with the pulse waveform pR(t). (Recall
that p,(t) is the response of a system consisting of a transmitter,
receiver and a single point reflector). In other words, the received
waveform g+(t) generated with a pulse p,(t) is related to the true
d&zizibution of elemental scatterers (distributed with respect to range),
g+ 6) r by

gp = P,(t) * gtrue(f) , (35)


4
where
true -
(t) = f(r,e)ds .
g9 i
C($,E)

and C(Glt) denotes the circular integration path centered at ($,R) with
radius t/2.

Comparing the convolution relations (34) and (35) shows that, for an
optimum reconstruction, we should choose an "ideal" acoustic pulse whose
functional form resembles the convolution kernel c(x). This "ideal" pulse
was in fact obtained in the last section and was denoted there by
pFt(F) (see eq. (27)). Thus, when we replace pR(E) in eq. (35) with pR Opt ct,
t is relation takes the same form as eq. (34), and we get

gT%) = p;pt(f) * gT(F) . (36)


opt -
We conclude that if the optimum pulse is used, the function g+ (t) can
be backprojected, without any further filtering, to obtain the desired
optimum reconstruction.

We now consider the case where an arbitrary, but known, pulse p (E) is
used in generating the set of functions g (E). This set can then beRsub-
jetted to a suitable post-filtering opera 2.ion to obtain the desired form
opt(E), given by eq. (36), prior to backprojection. (Of course, for the
g4
filtering operation to be effective without seriously degrading the signal-
to-noise ratio, the original pulse should have sufficient energy over the
bandpass of the filter). That is, we can write
opt -
g,+ (t) = +) * hop%) , 137)

where h opt (t)- is to be determined. By Fourier transforming eqs. (35) and


(36) and dividing, it follows from eq. (37) that

h Opt(;) = fl{HoPt(;)}- , (38)


t
where

P opt (V)
-
Hopt(;) = R =- l;lA(<) (39)
PR (5 P,(G '

F-l{*}- denotes an inverse Fourier transform, and we have substituted eq.


(25) fir p"pt(;)
R -

Thus, when an arbitrary pulse p,(f) is employed and an optimum recon-


struction is desired, the filter whose transfer function is given by eq.

171
NORTON AND LINZER

Fig. 8 Single pair of separate transmit


and receive elements in a circular
array enclosing the object.

(39) may be applied to the received signals before thg backg$gjEction pro-
cess is carried out. (Note, of course, that when p,(t) = p,lx (t) is used,
no prefiltering is needed, as eq. (39) indicates). A

We recognize that the analysis leading to eq. (39) was as simple as


possible. For-example, the multiplication of the transfer function
I+(V) by P (v)-1 represents a classical inverse-filtering operation; the
possibility $f zeroes in the spectrum P (<) and the existence of noise in
the bandpass I;[<; have not been consi 3 ered. A more sophisticated analys is
that attempts to a&ount for the influence of noise in this region should
yield a modified optimum transfer function HOBO (e.g., a Wiener filter)
although in the limit of large signal-to-noise ratio over the bandpass of
interest, the modified transfer function should reduce essentially to eq.
(39). It is also customary in practice to choose an apodizing function
A(V) that falls more smoothly to zero at the boundaries of the bandpass
than _the rectangLe function in eq. (26). One example of such a function
is A(V) = exp(-lv(/; ) , which we shall use in some of the computer simu-
lated reconstruction: of Section VI.

IV. IMAGE RECONSTRUCTION FROM REFLECTIVITY DATA GENERATED OVER ELLIPTICAL


PATHS

In this section we consider the case where the transmitter and recei-
ver are separate elements which lie at different points on the array cir-
cumference, as shown in figure 8. For a given round-trip time delay, the
receiver integrates the echoes arising over an elliptical path whose foci
coincide with the transmitter and receiver (if the round-trip delay is
denoted by T, the length of the semi-major axis of this ellipse is easily
shown to be TC/~). Thus, when a short pulse is emitted at point T and
data are continuously recorded as a function of time at point R, line-
integrals of reflectivity are generated over an entire family of ellipses
whose foci are R and T.

Suppose that the transmitter and receiver have a fixed angular separa-
tion 2@, as shown in figure 8. Let the line that bisects this angle (i.e.,
angle TOR) have angle 4 measured from the positive x-axis. (Thus, from
the geometry of figure 8, the minor axis of any ellipse with foci R and T
is oriented at the angle 0). Now suppose pulse-echo data are generated
for all values of $ keeping 8 fixed; that is, while keeping the separation

172
ULTRASONIC REFLECTIVITY TOMOGRAPHY

between the transmitting and receiving elements constant, suppose this pair
of elements is shifted around the circumference of the circle so that $
ranges over 360'. Using the line-integral data obtained in this fashion,
one can attempt to reconstruct the unknown reflectivity function f(r,e).

Let g+ B(t) denote the received echo data recorded as a function


of time for'a T/R pair separated by the angle 2f? and oriented at angle $J,
as shown in figure 8 (for B = Or this problem reduces, of course, to
the problem of Section III). When this function is recorded and backpro-
jetted along elliptical paths for all $I, the result expressed mathematically
is

:(B;r,B) = Tg [p($+B;r,8) + p(W;r,e)la$ , (40)


* 4JrB
and, as in eq. (2),
l/2
o(u;r,@) S [R2 + r2 - 2Rrcos(&u)] . (41)

Equation (40) is entirely analogous to eq. (1); here, however,


f3(9+B;r,0) + P(4-B;r, 8) represents the round-trip delay between a point
reflector and two separate transmitting and receiving elements. Thus, the
locus of points for which the integrand in eq. (40) is a constant is now
an elliptical path instead of a circular one.

We now compute the PSF produced by a point reflector at (ro,8,). The


derivation of the PSF will closely parallel the development in Section III.
As before, suppose the received wavegorm produced by an echo originating
from a point target has the form p,(t). Then for a point target at (ro,eo),
we get

g+,e(t) = pR(t - p(@+B;r 0 ,eo) - p(@-8;r 0 ,eo)l, (42)

which is the analog of eq. (4), Section III.

Putting eq. (42) into eq. (40) gives


21T
?(B;r,B) = ( p [A b-,B;ro,eo)la@ , (43)
o R +,@

where

(r,8;ro,e0)
Ab8
E P(++P;r,e) + p(M;r,e) - p(~+B;ro,eo) - p(@-f3;r ,eo).
0
Again, we shall assume (r/R)2 << 1 and (r /R)2 << 1. Then eq. (41) can be
expanded to second order in r/R and ro/R,Ogiving

+ -r2 - 4R
r2
ph;r,e) g R - rcos(fSu) cos2(e-u)
4R
Therefore,

-rcos (e-+-f3 - rcos(e-$+f3) + roc0s(80-$H3) + roc0s(80-++B)

+& r-r2c0s2(e+f3) - r'cos2 (e-++B) + ri ~0~2 (e,-+e) + rtcos2 (e,-++f3) 1.

173
NORTON AND LINZER

After considerable trigonometric manipulation, one can rewrite the above


expression in the following form:
A+ B(r,8;ro,00) = 2Xcos&os($-Y) + y1 + y~cos2~cos2(~-a) ,
,
where X, Y, yl, ~2, and c1 are defined in eqs. (9) and (10).

This last expression is similar to eq. (8), Section III. The only
difference is the presence of the constant factors COST and cos26 multi-
plying the first and third terms on the right, respectively. Paralleling
the derivation of Section III, we substitute the above expression for
A+ B into eq. (43) and use the Fourier transform relation (12) to arrive
at'an expression analogous to eq. (13). This result can again be con-
veniently expressed as the sum of two parts:

i(B;r,B) = 20(8;r,e) + E(B;r,e) , (44)

where
m
i,(B;r,B) = 2Tr ( d: PR(:)Jo(4~<XcosB) , (45)
-m

and &(B;r,e) is an error term entirely analogous to eq. (16). Using an


argument similar to the one employed in Section ;I1 and the Appendix, it
can be shown that E(B;r,e) is small compared to f (fi;r,8) under essentially
the same circumstances that E(r,B) is small cornpaged to ^f (r,8). Therefore,
in the same Fanner as in Section III, we neglect E(B;r,e)Oand approximate
f(B;r,e) by fo(B;r,8).
h
Note that f (B;r,8) is identical in form to f (r,8), defined by eq. (la),
except for the pgesence of the COSB factor in the zrgument of the Bessel
function. Because of this similarity, the same optimum pulse spectrum
(25), is applicable here as well. Thus, putting pFt(G)=
. (451, and integrating, yields

Gc Jl[4~cXcos@]
Zyt(f3;r,c3) = (46)
XCOSB
The distance X between the peak of this function and its first zero
(the Payleigh dista&e) is given by

(471

Thus, the minimum mainlobe width occurs for B = 0. (Note when B = T/2, the
transmitting and receiving elements are diametrically opposite; in this
extreme situation, the difference between the incident and scattering
angles approaches zero degrees when the object lies close to the center of
the array, i.e., the scattering is measured in the forward direction).

The reason that the mainlobe width is proportional to the factor (cosB)-l
can be more easily understood by reference to figure 9. For simplicity,
we consider an elliptical path intersecting a point at the array center
(r = 0). We now consider the inherent resolving capability of the two
trznsducers in the r-direction, which is defined as shown in the figure.
From the geometry in the figure, it is not hard to show that the rate
of change of the roundtrip delay (over the path TOR) with respect to
r decreases as c0s.B. This means that, for a fixed time-delay resolution,
if a reflecting target moves an increment Ar along the r-axis, the
minimum possible hr that can be resolved increases as (cosB)-l. Thus,

174
ULTRASONIC REFLECTIVITY TOMOGRAPHY

elliptical arc with

Fig. 9 The rate of change of the distance


r with respect to the time delay
over the path TOR increases as
(cos%)-l. This effect produces
a broadening of the mainlobe
width of the resulting PSF by a
factor of (cos%)-l.

the system achieves its maximum inherent resolving capability in the


r-direction when the transmitter and receiver coincide (% = 0). When
the recorded signals from all the elements are backprojected over the
paths distributed symmetrically around 360°, the same fundamental (cos $3)-l
dependence reveals itself in the expression for the reconstructed PSF
(eq. 45)). These considerations suggest that a practical upper bound
on the value of % used may be about 45O; in this case, a receiving
element would never be further from the transmitter than one quarter
of the circumference of the array. This restriction would imply a
mainlobe broadening that would never exceed a factor of fi.

V. IMAGE RECONSTRUCTION WITH MORE THAN ONE ELEMENT RECEIVING SIMULTANEOUSLY

In a practical imaging system design, it is natural to contemplate


using more than a single receiving element at a time if they are available.
This can be achieved by transmitting from a single element and receiving
with many elements simultaneously. We denote this method of data acquisi-
tion as "mode 3". The advantage, of course, of using more than a single
element as a receiver is a major increase in the rate at which ultrasonic
data can be generated and recorded. In medical imaging, where data are
subject to degradation due to patient motion, fast data acquisition may
well be of crucial importance. When measurement speed is not an important
factor, the use of simultaneous receiver elements will also provide an
improvement in signal-to-noise ratio for a fixed measurement time.

When data are recorded in this multiple-receiver fashion, line-


integrals of reflectivity are generated over elliptical paths exhibiting a
variety of shapes and sizes. To reconstruct a reflectivity image, each
waveform generated from a given element pair is filtered and then backpro-
jetted over the appropriate family of elliptical paths (i.e., whose foci
coincide with the given element pair). This process is carried out for the
entire set of element pairs used to record the data. It is useful to inter-
pret the resulting mode 3 image as the result of coherently superimposing
many mode 2 images obtained for different values of %. Thus, the PSF
obtained for this process can be computed by effectively averaging f(%;r,B),
given by eq. (45), over some suitable range of %. Considerations from
Section IV suggest that I%1 < a/4, for example, is an acceptable range and
should give reasonable results. The resulting PSF can then be expressed as
follows:

?b-,B) = j'w(%)i(%;r,B)d% , (48)


-71

175
NORTON AND LINZER

Id

Fig. 10 Received waveforms used in computer simulations: (a) narrowband


pulse; (b) wideband pulse; and (c) optimum pulse pipt(t) (eq. 27).

where, for example,

w(B) = 1 for If31 < n/4

=o otherwise.

VI. COMPUTER SIMULATIONS

A computer program was written that simulates echo data generated by


reflecting a given pulse from a single isolated point target. These data
were then used to reconstruct an image (PSF) of the target. Except for
one set of simulations considered below, we have, for simplicity, assumed
that the target is located at the center of the circular array (r = 0).
Three different pulse shapes were employed and are shown in figurg 10. The
first pulse is narrowband and consists of a gated cosine wave of 16 cycles
duration (figure lOa). A wavelength x = 0.75 mm was used, corresponding
to a 2 MHz center frequency. The seco:d pulse (figure lob) is broadband,
and was constructed by adding together two Gaussian functions of opposite
sign, one peaked and the other broad. The Gaussian functions were chosen
so that the area under the pulse in figure lob is zero, thus eliminating
a DC component. The width of the sharply-peaked positive-rising portion
of the pulse was also adjusted so that approximately 80% of the spectral
energy of the pulse is contained within the bandpass IV1 ( A-'. The third
pulse (figure 10~) is pgpt(t), as given by eq. (27).
frequency, defined by eq. (26) as vc, was set equal to A Over 95% of
the energy in the digitized version of pgpt(t) used in e simulation is
contained in the spectral region IV1 h A,'l. (For the ideal pgpt(t), whose
transform is given in eq. 126), all of the energy is contained below the
cutoff vc. However, pEpt(t), as defined in eq. (27), extends to infinity
in both directions requiring, in its practical implementation, truncation
at some distance on either side of its maximum value; thus, some spectral
energy will be distributed beyond the cutoff).

The received waveforms produced by using each of these pulses were


backprojected over circular paths to obtain reconstructions (PSF's) of the
point target. Since the target is located at the center of the atray, the
reconstructions will have circular symmetry, and radial plots of the PSF's
resulting from the three pulses are shown in figure 11. In the simulation,
the array was assumed to have 64 elements evenly spaced around its circum-
ference, and the radius R was 100 mm.

In figure lla, the narrowband PSF closely resembles a truncated J (0)


Bessel function (see eq. (19)). Note that the number of visible sidelgbes

176
ULTRASONIC REFLECTIVITY TOMOGRAPHY

Fig. 11 PSF reconstructions (ro = 0, 0 = 0) using (a) narrowband pulse;


(b) wideband pulse; and (c) optimum pulse.

I (a)

Fig. 12 PSF reconstructions (r. =_ 0, 6 = 0) using (a) narrowband pulse


filtered with Iz[exp(-x,Iv/); (b) wideband pulse filtered with
l~lexP(-~ol?l); and (c) optimum pulse filtered with exp(-Ao\vl).

in this reconstruction is essentially one-half the number of cycles in the


gated cosine wave (figure lOa). Figure llb shows the reconstruction using
the broadband pulse (figure lob), which clearly exhibits the l/r asymptotic
fall-off from its center, as predicted by eq. (20). Figure llc shows the
reconstruction using the optimum pulse pgpt(S). In the vicinity of the
mainlobe, this function closely resembles the Jl(*)/(*) form, as anticipated
from eq. (29). For the narrowband and optimum PSF's, the analytically-
calculated mainlobe widths show nearly exact agreement with the simulated
results.

It is interesting to apply the optimum filter, whose transfer function


is given by eq. (251, to the received narrowband and wideband waveforms
before backprojection is carried out. For this purpose, the choice of the
spectrum apodizing function

A(;) = e-xol’l
was substituted in eq. (25). The reconstructions which then result from
applying the filter (25) to the recorded data and then backprojecting are
shown in figure 12.

177
NORTON AND LINZER

-----A :

-0.251
t -0.251.
-4 -2 0 2 4 21 23 25 27 23
r h-4 rlmm)

Fig. 13 Mode 2 PSF reconstruction Fig. 14 PSF reconstruction of off-


center target (r = R/4,
0
f3 = 0).

As seen in figure 12a, the narrowband reconstruction is not noticeably


changed by the prior filtering operation, since most of the spectral energy
of the narrowband pulse (figure 10a) is essentially zero outside of a very
narrow region, and thus filtering does little to alter this spectral dis-
tribution. On the other hand, the spectrum of the wideband pulse (figure
lob) is distinctly altered by the filtering process, particularly by the
presence of the factor Iv/ in the transfer function. The resulting recon-
struction, shown in figure 12b, may be compared with the corresponding
unfiltered PSF, shown in figure llb. Notice the more rapid asymptotic decay
away from the peak of the filtered PSF. Finally, in figure 12c, we show
a PSF obtained by using the optimum pulse, pgEt(t), filtering the resul-
ting echo data with the apodizing function A(V), as given by eq. (49), and
then backprojecting. This_ result may be compared with figure llc, which
was obtained using pgPt(t) without any spectral apodization prior to back-
projection. In figure 12c, a slight broadening of the mainlobe is visible,
but the sidelobes are less accentuated and appear to decay more rapidly
to zero.

A mode 2 reconstruction was simulated in which the pulse-echo data were


generated by transmitting and receiving with a distinct pair of eiements
separated by an angle of 90° (6 = 45O). The optimum pulse, pgpt(t), was
used in this reconstruction. Backprojection (now over elliptical paths)
of the resulting echo data gives rise to the PSF shown in figure 13. When
this PSF is compared with that of figure llc (mode 1 or backprojection
along circular paths), we see that the former reconstruction is of similar
shape but broadened by exactly a factor of (~0~8)~' = fi, as predicted by
eq. (47).

In all of the foregoing PSF reconstructions, we have used ro= 0. We


have also simulated mode 1 PSF's for the target positioned away from the
array center. These reconstructions were-obtained by employing pROpt ct, ,
filtering the resulting echo data with A(V) as.given by eq. (49), and then
backprojecting over circular paths. The result, for r. = R/4, is shown in
figure 14. The plot in this figure is along a radial line intersecting the
origin and the peak of the reconstruction. This result supports our earlier
hypothesis that the basic shape of the PSF remains relatively constant as r.
moves away from the center of the array. There is, however, a perceptible
broadening of the mainlobe which increases as r. increases. There is ah0
some "filling in" of the sidelobe nulls around the mainlobe. Unfortunate-
ly, this phenomenon is not predictable on the basis of any analysis that

178
ULTRASONIC REFLECTIVITY TOMOGRAPHY

0.6

Fig. 15 Width at half-maximum-


amplitude for several
reconstructed PSF's
(S = 0) plotted as a
function of the off-center
distance r . The wideband
pulse is ffltered with
Ivlexp(-X Ii~l) before
backprojegtion.

1 I 1 I

10 20 30 40 50
OFF CENTER DISTANCE
f. (mm1

uses the "straight line" approximation, i.e., the approximation obtained by


neglecting the curvature of the integration paths. One can argue, in fact,
that this off-center broadening effect is partly attributable to the differ-
ence in curvature of the arcs over which the backprojection is performed,
since the radius of the intersecting arcs corresponding to different array
elements vary when the image point is not at the exact center of the array.
also, as a result of this asymmetry, the intersecting arcs are not symmetri-
cally distributed with respect to the center of the point under reconstruc-
tion. The combination of these effects are probably responsible for the
visible mainlobe broadening and the smearing of the nulls in the sidelobe
pattern close to the mainlobe, as can be seen in the figure. In figure 15,
we indicate how the broadening of the mainlobe width increases as a function
of the off-center distance r. for reconstructions obtained using the three
different pulses shown in figure 10. This figure shows the mainlobe width
of the PSF, measured at half-maximum amplitude, plotted versus ro. It is
interesting to note that the width of the narrowband PSF appears to change
the least as a function of r
0’

Some general remarks may be made regarding reconstructions generated


with a finite number of array elements N. For example, in figure lla, the
Jo(*) sidelobes seem to disappear after about 7 oscillations from the
mainlobe. However, just beyond this point, some sidelobe activity reappears.
This artifact results from N being finite. In particular, this is a
result of the failure of the N backprojected signals to properly overlap
and "cancel out" for image points beyond a certain distance from the
mainlobe. This effect becomes more pronounced and occurs closer to the
mainlobe as the number of elements in the array (i.e., the number of
backprojections) decreases. For a relatively sparse array, the backpro-
jetted signals may fail to overlap entirely for some points within the
image space. This phenomenon gives rise to a "worst-case" signal-to-noise
ratio (i.e., mainlobe-to-sidelobe level ratio) approximately equal to NHp,
when N is defined as the total number of backprojections erfonned in the
recons BPruction (note that NHP = N for mode 1 and 2; NHP L NB for mode 3).
For a point reflector located at the center, this value is derived from the
fact that all of the backprojections are summed together at this point,
thus constituting the "signal", while in the region of zero overlap, a single
backprojection may exist, thereby contributing the "noise" term. Moreover,

179
NORTON AND LINZER

Fig. 16 Schematic illustration of


overlap region when a small
t number of backprojections are
i performed in image space.

one can estimate the approximate radius beyond which the backprojected
signals fail to overlap at all. For a pulse of duration L, one can show
from simple geometrical considerations (figure 16) that this radius,
measured from the mainlobe peak, is approximately NRPL/X. Thus, from the
standpoint of improving both the sidelobe response as well as the overall
system sensitivity, one should make NRP as large as practical.

We conclude by considering a PSF figure-of-merit, or measure of the


mainlobe peak to average sidelobe level, defined by

GN =
2lT r
If(ro, eo) I
I (50)

where rmax is some suitable


1 I maxIf(r,8)12rdrd9
0
upper limit
0
beyond which f(r,8)
1l/2

is assumed
negligible. To provide an indication of how rapidly the sidelobe response
of the PSF degrades as the number of array elements is decreased, GN was
computed for the mode 1 case where N was allowed to range between 10 and
100 in increments of 10. The backprojected signals were generated using
Popt(t), after filtering with the apodising function A(<) given in eq. (49).
rp*smg r = 25 mm and the same V as in the other simulations, the resulting
values gf GN are plotted in fig&e 17.

Fig. 17 The figure-of-merit G


plotted versus elemen r
number N for PSF recon-
structions (r = 0, B = 0)
using the optzmum pulse.

10 20 30 40 00 00 70 00 90 loo
ELEMENT NUMBER N

180
ULTRASONIC REFLECTIVITY TOMOGRAPHY

APPENDIX

Here we compare_ the relative magnitudes of the "zero-order" approxi-


mation to the PSF, f,(r,S), and the "error" term, E(r,8), given by eqs. (29)
and (16), respectively, under conditions where X is large compared to the
mainlobe width. We begin with eq. (29):

TJcJl (41TScX)
i,(r,e) = (Al)
X

and eq. (16):

27l m
s(r,B) = / d$(d<P,(;) ri2Tr$+ %(i2'+)2+...]e iln~Xcos($-Y) (AZ)
I
0 -co

where 5 = y1 + y2cos2($-C%). We shall assume that the functions P,(G)


of interest are even in the variable 3. Then, terms that are odd in eq.
(A2) with respect to J integrate to zero; as a result, we shall consider
only the two lowest order even terms. Upon changing orders of integration,
eq. (A2) then becomes

E(r,e)

The first term in the integrand can be shown to integrate to zero when
the integration is performed with respect to $. The second term can be
integrated to yield
co
E(r,e) = -2lT3/ d;PR(;);2[(2y,2+y22)Jo(4?%)
-co

-4yly2coS2(Y-~)J2(4~~X) + y$zosl(Y-a) J4(4msX)l . (A4)

For conciseness, we define

2 = 4TIx

la-1 = 2y12+y22
(A5)
r2 = -4y~y~cOS2(k'-Cl)

r3 = y:cos4(Y-a) .

In eq. (A4), we will substitute PR(3) = I?lrect(~/2~ ), the optimum pulse


spectrum derived in Section III-C (see eq. (26)). Thk,
V
C
E(r,0) = -4Tr' ;3[r1Jo(s;) + r2J2(zL)) + raJ4(s<)]d< . (A6)
I
0

181
NORTON AND LINZER

We wish to evaluate an integral of the form


; C
I
n
=
0
I :3Jn(z;)d, .

By employing the method of stationary phase 1151, the asymptotic form


for this integral can be obtained,

I = ;2 22; Gsin(zCc-nn/2+/4) ,
n C
[az3 I
which is valid for z >> 5 - 1 .
C

Using this in eq. (A6), and recalling that z = 4ax, we get

E(r,e) = - (;) j2y4


vc F [rl-r2+ralsin(4GcX-7T/4). @7)

Using the asymptotic form of the Bessel function,

J,(x) = k ‘cos (x-m/2-T/4) ,


II 1
in eq. (Al), results in

[I 2; 4
20(r,f3) = & $ sin(4fiCx-71/4). (A81

It is convenient now to define the ratio

Upon substituting eqs. (A7) and (Aa), we obtain

c(r,O) = 7?3p-r2+r2i
Using eqs. (A5), and after some trigonometric simplification, this
reduces to
5(r,f3) = 2.rr23~tyI+y2cos2(Y-c1)1* , (A91
whereyit Y2t Y and c1 are defined in eqs. (lOa), (lob), (9b) and (lOc),
respectively.

This formula is cumbersome to evaluate for arbitrary (r,8). However,


eq. (A9) can be simplified for points lying along the radial line defined
bye=8 and along the arc r = r
0 0’

First, we consider points lying along the arc r = ro. Putting r = r.


in eq. (lOa) gives Yl = 0.

From eq. (lob), Y2 = (ri/R) jsin(&BO) 1 .

From eq. (9b), Y = (112) (eo+e-7v2) .

From eq. (ioc), CL = eo+e-r/2 .

182
ULTRASONIC REFLECTIVITY TOMOGRAPHY

Substituting these in eq. (A9), we have

<(ro,B) = ($)$ [$]z(cos200-cos20)z

r2 2
< 2n%J2
CR
0
*
11 (A101

Next, we consider points lying along the radial line 8 = 0 Putting


0’
0=8 in eq. (lob) gives
0

y2 = (1/2R)(ri-r2j .

From eq. (9b), Y=8 0’

From eq. (lOc), a = 28 0’

After substitution into eq. (A9), we have

c(r,f3 0
1 = 27rzj2
c cI
r2-r2 2
0
R
lsineol

( 2K52 C c 1r2-r2 2
0 R
. (All)

It is convenient to use the relation (30) to express the cutoff


frequency cc in terms of XoI the mainlobe width, i.e.,

; =- 0.30
C
xO
Putting this into eqs. (AlO) and (All) yields finally

5(ro,8) -< 1.78 I22


&12 0
. (AU)

c(r,eo) 5 1.78 [ 1
r2-r2
0
XOR
2
. (Al3)

For example, from eq. (A12), we see that the condition X R > 1.33r 2
is needed to guarantee <(r ,8) < 1. This is, unfortunately, P fairly s&ere
constraint to impose on thg relative sizes of X and R. If, for example, we
set r 0 = R/S, then for C(ro,fl) ( 1 to hold, we gust have x o L R/20.

One can conclude from eqs. (A12) and (A13) that for a high resolution
imaging system where a large temporal bandwidth is used, and where r and r 0
are not restricted to values that are a small fraction of R, the contribu-
tion that E(r,e) gakes to the sidelobe level far from the mainlobe will
dominate that of f,(r,e). We note from eq. (A7), however thay2E(r,e) falls
off asymptotically at the same rate as ?,(r,O), i.e., as X . Thus,
although the sidelobe level due $0 E(r,h) may under certain conditions be
larger or comparable to that of f (r,e), it nevertheless attenuates rapidly
at 1aEge distances from the mainl8be. What is perhaps more important is
that fo(r,8) succeeds in accurately predicting the detailed shape of the
PSF close to the mainlobe as a function of the pulse spectrum P,(v).
This has been clearly borne out by the computer simulation results
presented in Section VI.

183
NORTON AND LINZER

REFERENCES

1. IEEE Trans. Nuclear Science, NS-21, June, 1974, Special Issue


on 3-D Image Reconstruction from Projections.

2. Johnson, S. A., Greenleaf, J. F., Tunaka, M., Rajagopolan, R., and


Bahn, R. C., Reflection and Transmission Techniques for High Resolu-
tion Quantitative Synthesis of Ultrasound Parameter Images, in 1977
IEEE Ultrasonic Symposium Proceedings, (IEEE Cat. No. 77 CH1264-ISU).

3. Johnson, S. A., Greenleaf, J. F., Rajagopalan, B., Bahn, R. C., Baxter,


B ., and Christensen, D., High Spatial Resolution Ultrasonic Measurement
Techniques for Characterization of Static and Moving Tissue, in Ultra-
sonic Tissue Characterization II, M. Linzer, ed., National Bureau of
Standards Special Publication 525, pp. 235-246, (U.S. Government
Printing Office, Washington, DC, 1979).

4. Johnson, S. A., Greenleaf, J. F., Rajagopalan, B., Tanaka, M., Alge-


braic and Analytic Inversion of Acoustic Data from Partially or Fully
Enclosing Apertures, in Acoustical Holography, Vol. 8, A. Metherell,
ed., (Plenum Press, New York, 1979).

5. Corl, P. B., Kino, G. S., DeSilets, C. S., Grant, P. M., A Digital


Synthetic Focus Acoustic Imaging System, in Acoustical Holography,
Vol. 8, A. Metherell, ed., (Plenum Press, New York, 1979).

6. Norton, S. J. and Linzer, M., Tomographic Reconstruction of Reflec-


tivity Images, in Abstracts of Third International Symposium on
Ultrasonic Imaging and Tissue Characterization, June 5-7, 1978,
National Bureau of Standards, Washington, DC.

7. Norton, S. J., Reconstruction of a two-dimensional reflecting medium


over a circular domain: exact solution, J. Acoust. Sot. Amer., (in
press).

8. Morse, P. M., and Ingard, K. LJ., Theoretical Acoustics, Chapter 6.


(McGraw-Hill, New York, 1968).

9. Aks, s. a., and Vezzetti, D. J., Improvements in the analysis of


ultrasonic scattering from inhomogeneities in tissue, Ultrasonic
--Imaging 1, 44-45 (1979).
10. I&s, s. a., private communication.

11. Bracewell, R. N., The Fourier Transform and Its Applications, p. 85,
(McGraw Hill, New York, 1965).

12. Papoulis, A., Systems and Transforms with Applications in Optics,


Chapter 5, (McGraw Hill, New York, 1968).

13. Bracewell, R. N., and Riddle, A. C., Inversion of fan-beam scans in


radio-astronomy, The Astrophysical Journal, E, 427-434 (1967).

14. Ramachandran, G. N., and Lakshminavayanan, A. V., Three-dimensional


reconstruction from radiographs and electron micrographs: applica-
tion of convolutions instead of Fourier transforms, Proc. U.S. Natl.
Acad. Sci., -68 2236-2240 (1971).

15. Papoulis, A., Systems and Transforms with Applications in Optics,


Chapter 7, (McGraw-Hill, New York, 1968).

184

You might also like