Inside Out: Inverse Problems

MSRI Publications
Volume 47, 2003
Introduction to the Mathematics
of Computed Tomography
ADEL FARIDANI
Abstract. Computed tomography (CT) entails the reconstruction of a
function f from line integrals of f. This mathematical problem is encoun-
tered in a growing number of diverse settings in medicine, science, and
technology. This introductory article is divided into three parts. The rst
part is concerned with general theory and explores questions of uniqueness,
stability and inversion, as well as detection of singularities. The second
part is devoted to local tomography and is centered around a discussion of
recently developed methods for computing jumps of a function from local
tomographic data. The third part treats optimal sampling and has at its
core a detailed error analysis of the parallel-beam ltered backprojection
algorithm. Matlab source code for the ltered backprojection algorithm
and the FeldkampDavisKress algorithm is included in an appendix.
1. Introduction
Computed tomography (CT) entails the reconstruction of a function f from
line integrals of f. This mathematical problem is encountered in a growing
number of diverse settings in medicine, science, and technology, ranging from
the famous application in diagnostic radiology to research in quantum optics. As
a consequence, many aspects of CT have been extensively studied and are now
well understood, thus providing an interesting model case for the study of other
inverse problems. Other aspects, notably three-dimensional reconstructions, still
provide numerous open problems.
The purpose of this article is to give an introduction to the topic, treat some
aspects in more detail, and to point out references for further study. The reader
interested in a broader overview of the eld, its relation to various branches of
pure and applied mathematics, and its development over the years may wish to
consult the monographs [6; 31; 32; 36; 62; 67; 78], the volumes [21; 22; 28; 33;
34; 76; 77], and review articles [42; 49; 56; 58; 66; 84; 89].
Work supported by NSF grant DMS-9803352 and NIH grant R01 RR 11800-4.
1
2 ADEL FARIDANI
In practice only integrals over nitely many lines can be measured, and the
distribution of these lines is sometimes restricted. The following presentation is
centered around the question: What features of f can be stably recovered from a
given collection of line integrals of f? For example, we may ask what resolution
can be achieved with the available data. If a full reconstruction of f is not
possible, we may try to detect the location of boundaries (jump discontinuities
of f), or also the sizes of the jumps.
The exposition is divided into three parts. The rst part is concerned with
general theory. Its main themes are questions of uniqueness, stability and in-
version for the x-ray transform, as well as detection of singularities. The second
part is devoted to local tomography. The exposition is similar to [17] and is cen-
tered around a discussion of recently developed methods for computing jumps of
a function from local tomographic data. The third part treats optimal sampling
and has at its core a detailed error analysis of the parallel-beam ltered backpro-
jection algorithm. The article conludes with three appendices containing basic
results on wavelets, Matlab source code for the ltered backprojection algorithm
and the FeldkampDavisKress algorithm, and some exercises.
2. The X-Ray and Radon Transforms
We begin by introducing some notation and background material. R
n
consists
of n-tuples of real numbers, usually designated by single letters, x = (x
1
, . . . , x
n
),
y = (y
1
, . . . , y
n
), etc. The inner product and absolute value are dened by
'x, y` =

n
1
x
i
y
i
and [x[ =

'x, x`. The unit sphere S

n1
consists of the
points with absolute value 1. C

0
(R
n
) denotes the set of innitely dierentiable
functions on R
n
with compact support. A continuous linear functional on C

0
is called a distribution. If X is a set, X

denotes its interior, X its closure, and

X
c
its complement.

X
and

n
denote the characteristic functions (indicator
functions) of X and of the unit ball in R
n
, respectively (that is,

X
(x) = 1
if x X and

X
(x) = 0 if x X). [X[ denotes the n-dimensional Lebesgue
measure of X R
n
. However, when it is clear that X should be treated as a set
of dimension m < n, [X[ is the m-dimensional area measure. Thus
[S
k1
[ = 2
k/2
/(k/2)
is the (k 1)-dimensional area of the (k 1)-dimensional sphere.
The convolution of two functions is given by
f g(x) =

R
n
f(x y)g(y) dy.
The Fourier transform is dened by

f() = (2)
n/2

R
n
f(x)e
ix,
dx
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 3
for integrable functions f, and is extended to larger classes of functions or dis-
tributions by continuity or duality. For square-integrable functions f, g we have
f g(x) =

R
n

f() g()e
ix,
d. (21)
The integral transforms most relevant for tomography are the x-ray transform
and the Radon transform.
Definition 2.1. Let S
n1
and

the hyperplane through the origin

orthogonal to . We parametrize a line l(, y) in R
n
by specifying its direction
S
n1
and the point y where the line intersects the hyperplane

.
The x-ray transform of a function f L
1
(R
n
) is given by
Pf(, y) = P

f(y) =

R
f(y +t) dt, y

.
The Radon transform of f is dened by
Rf(, s) = R

f(s) =

f(x +s) dx, s R. (22)

We see that Pf(, x) is the integral of f over the line l(, y) parallel to which
passes through y

, and that Rf(, s) is the integral of f over the hyperplane

orthogonal to with signed distance s from the origin. In the following we
will be mostly concerned with the x-ray transform. In two dimensions the two
transforms coincide apart from the parameterization: We parametrize S
1
by
its polar angle and dene a vector

orthogonal to such that

= (cos , sin ),

= (sin , cos ). (23)

Then the points in the subspace

are given by

= s

: s R and
we have the relation Pf(, s

) = Rf(

, s). Also, when working in two di-

mensions, we will often use the simplied notation Pf(, s) or P

f(s) instead of
Pf(, s

). Occasionally we will also replace by the polar angle according

to (23) and write Pf(, s) .
We consider two examples. Let G be the Gaussian function G(x) = e
x,x/2
.
Then
PG(, y) = e
y,y/2

R
e
t,t/2
dt = (2)
1/2
e
y,y/2
, y

.
For

n
, the characteristic function of the unit ball in R
n
, we can use a geometrical
argument. We obtain P

n
(, y) = 0 for [y[ > 1 since then the line l(, y) does
not intersect the unit ball. For [y[ 1 observe that the intersection of the line
l(, y) with the unit ball in R
n
is a line segment of length 2

1 [y[
2
and that
P

n
(, y) is equal to this length.
The following relation between the Fourier transforms of P

f and f will prove

to be useful:
4 ADEL FARIDANI
Theorem 2.2. Under the hypotheses of Denition 2.1,

f() = (2)
1/2

f(),

f() = (2)
(n1)/2

f(), R
Proof. This is a straightforward computation. We demonstrate it for the x-ray
transform. Let

. Then

Pf(, ) = (2)
(1n)/2

Pf(, x)e
ix,
dx
= (2)
(1n)/2

R
f(x +s) ds e
ix,
dx
= (2)
(1n)/2

R
n
f(y)e
iy,
dy =

2

f().
As we will see below, Theorem 2.2 can be used to explore questions of uniqueness,
nonuniqueness, stability, and inversion.
Current medical scanners employ an x-ray source which moves around the
patient. To describe this type of data collection, the parameterization of lines
by S
n1
and y

is less convenient. It is more suitable to introduce the

divergent beam x-ray transform
Df(a, ) = D
a
f() =

0
f(a +t) dt, S
n1
,
which gives the integral of f over the ray with direction emanating from the
source point a.
The x-ray and Radon transforms are special cases of the general k-plane
transform, which maps a function into its integrals over k-dimensional ane
subspaces; see [42], for instance.
3. Uniqueness and Nonuniqueness
Theorem 3.1 [89; 42]. Let f L
2
(R
n
) have compact support, and suppose that
Pf(, ) 0 for innitely many . Then f 0.
Proof. The Fourier transform

f is analytic and

f() =

P

f() = 0 on the
hyperplanes ', ` = 0. Since no nontrivial entire function can vanish on an
innite set of hyperplanes through the origin, we must have

f 0.
As an application, consider the so-called limited angle problem. Let Pf(, )
be given for innitely many concentrated in a cone C. Then f is uniquely
determined, even if C is very small. However, if C = S
n1
, the reconstruction
is not stable. Indeed, the proof of the above theorem shows that reconstructing
f is equivalent to analytic continuation of

f, and analytic continuation is known
to be extremely unstable.
The uniqueness result requires an innite number of directions, while in prac-
tice only a nite number can be measured. It was already recognized by the
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 5
pioneers of CT that this entails the loss of uniqueness; see the example given
in [3]. The next theorem shows that the nonuniqueness is quite extensive, i.e.,
given Pf(
j
, ) for nitely many directions
j
, there are null functions which can
be prescribed arbitrarily on a large portion of their domain.
Theorem 3.2. ([89]) Let
1
, . . . ,
p
S
n1
, K R
n
compact, and f C

0
(K).
Let K
0
U K with U open and K
0
compact. Then there is f
0
C

0
(K),
f
0
= f on K
0
, and Pf
0
(
k
, ) 0, k = 1, . . . , p.
While this result makes it seem dicult to obtain reliable reconstructions in
practice, it is not the end of the story. It turns out that the null functions for
the x-ray transform are high-frequency functions [51; 52; 53; 60] , and that it is
possible to suppress such functions in practical reconstructions.
Theorem 3.3 [51; 52; 53]. Let f
0
L
2
(R
2
) with support contained in the unit
disk. If Pf
0
(
k
, ) 0 for k = 1, . . . , p, then

f
0
() =

m>p
i
m

1
J
m+1
()q
m
(),
where R, S
1
, J
m+1
the order m+1 Bessel function of the rst kind and
q
m
a polynomial of degree m.
Since J
l
(t) is very small for l > t, it follows that if P

f vanishes for p distinct

directions
j
, then

f() is almost entirely concentrated in the set R
n
:
[[ > p [60]. This means that measuring P

j
f determines

f() reliably for
[[ < p. However, the reconstruction problem may still be severely unstable,
e.g., when the directions are concentrated in a narrow range. In cases where
sucient stability is present, a low-pass ltered version of f may be recovered.
A loose application of Shannons sampling theorem yields that the reconstruction
will resolve details of size 2/p or greater.
Remark 3.4. It follows that the inuence of nonuniqueness may be avoided in
practice under the following conditions:
(a) A-priori information that [

f()[ is small for [[ > b is available.

(b) Data P

j
f for p > b directions
j
are measured.
(c) The reconstruction method used produces a function f
R
with [

f
R
()[ small
for [[ > b.
Nonuniqueness theorems for the divergent beam x-ray transform have been
proved in [48; 93]. A generalization to the general k-plane transform has been
given in [42].
4. Inversion and Ill-Posedness
Calderns operator is dened in terms of Fourier transforms by

() = [[ (), C

0
(R
n
).
6 ADEL FARIDANI
It is extended by duality to the class of functions f for which (1 +[x[)
1n
f is
integrable [14]. Note that

2
= , = Laplacian. (41)
For n 2, the inverse
1
of is given by convolution with the Riesz kernel R
1
:

1
f = R
1
f, R
1
(x) = ([S
n2
[)
1
[x[
1n
.
In dimension n = 1 we have f = Hf, where f denotes the derivative of f
and H denotes the Hilbert transform
Hf(s) =
1

R
f(t)
s t
dt (42)
where the integral is understood as a principal value.
We can formally derive an inversion formula for Pf by combining Theorem 2.2
and the inverse Fourier transform. For simplicity we rst consider dimension
n = 2. Using the Fourier inversion formula, Theorem 2.2, the relation (23) and
changing to polar coordinates we obtain
f(x) = (2)
1

R
2

f()e
ix,
d
= (2)
1

2
0

f(

)e
ix,

d d
= (4)
1

2
0

[[

f(

)e
ix,

d d
=
1
2
(2)
3/2

2
0

[[

f()e
ix,

d d
=
1
2
(2)
3/2

2
0

R
\
P

f()e
ix,

d d
= (4)
1

2
0
P

f('x,

`) d
=
1
4
2

2
0

R
P

f(s)
'x,

` s
ds d.
(43)
In the last step we made use of the relation g = Hg mentioned above.
For general dimension n one uses the change of variables [89, Formula (9.2

)]

R
n
g() d = [S
n2
[
1

S
n1

[[g() d d (44)
and obtains
f(x) =

2[S
n2
[

S
n1
P

f(E

x) d (45)
where E

x denotes the orthogonal projection of x onto the subspace

.
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 7
If we use the backprojection operator P

dened by
P

g(x) =

S
n1
g(, E

x) d,
then (45) assumes the compact form
f(x) =

2[S
n2
[

1
P

Pf(x).
An inversion formula for the Radon transform can be derived in a similar way.
For other inversion formulas see [62, ' II.2].
From the last line of (43) we see that computation of f(x) requires integrals
over lines far from x, because the Hilbert transform kernel has unbounded sup-
port. Note that P

f('x,

`) is the integral over the line with direction which

passes through x. Hence the inversion formula is not local. A local inversion
formula would utilize only values P

f(s) with s close to 'x,

`. We will discuss
what can be done with local formulas in a later section.
Equation (43) gives us valuable information about the stability of the in-
version. The factor [[ in the inverse Fourier integral will become arbitrarily
large. This means that the inversion is unstable. In practice measurement and
discretization errors will prevent accurate computation of

P

f() for large [[,

and these errors are then amplied by multiplication with [[. In other words,
due to the integration in P, Pf is smoother than f itself. The inversion has to
reverse this smoothing and this makes it unstable. The extent of this instability
will depend on the amount of smoothing inherent in P. This can be quantied
using Sobolev norms. For functions f with compact support we dene
|f|
H

0
=

R
n
(1 +[[
2
)

f()[
2
d

1/2
,
|Pf|

S
n1
d

d (1 +[[
2
)

f()[
2

1/2
.
Theorem 4.1 [62, p. 42]. If f C

0
is supported in the unit ball , then there
are constants c(, n), C(, n) such that
c(, n)|f|
H

0
|Pf|
+
1
2
C(, n)|f|
H

0
.
Hence the operator P smoothes by an order
1
2
measured in a Sobolev scale. In
order to see what the instability might mean in practice we assume that we have
measured data g

such that |Pf g

|
L
2
, and a-priori information about
f of the form |f|
H

0
. For > 0 this excludes highly oscillatory functions,
so this condition corresponds to condition (a) in Remark 3.4. Let f
1
, f
2
be
two candidate functions for reconstruction, i.e., f
1
, f
2
both satisfy the a-priori
condition and |Pf
i
g

|
L
2
. We are interested to know by how much f
1
8 ADEL FARIDANI
and f
2
can dier. Since |P(f
1
f
2
)|
L
2
2 and |f
1
f
2
|
H

0
2, we have the
worst case error
|f
1
f
2
|
L
2 d(, , ), with
d(, , )=sup

|f|
L
2
: |Pf|
L
2
2, |f|
H

0
2

.
A natural choice for is such that functions which are smooth except for jump
discontinuities along smooth boundaries are in H

0
. This leads to the condition
<
1
2
[62, p. 92]. For the limiting case =
1
2
the worst case error satises
d

, ,
1
2

c(n)

.
[62, p. 94]. This means that the reconstruction problem is moderately ill-posed.
We expect a gain of 2k digits in data accuracy to yield k additional accurate
digits in the reconstruction. In other words, the instability in the reconstruction
causes a loss of half the number of accurate digits.
Another approach to quantify the degree of ill-posedness is provided by the
singular value decomposition of P [60]. Here one looks at how fast the singu-
lar values converge to zero. Again, the assessment of moderate ill-posedness is
conrmed.
In order to use the inversion formula in practice we have to stabilize it. This
involves a well-known trade-o between stability and accuracy of the recon-
struction. Here we give up the goal of recovering the function f itself, and
aim instead at reconstructing an approximation e f, where e is an approxi-
mate delta function. As the computation below shows, stabilization requires the
Fourier transform e() to decay suciently fast for large [[. The price to pay for
the stabilization is limited resolution, so e must be chosen carefully, depending
on the amount and accuracy of the available measurements. Note also that a
proper choice of e helps to satisfy the condition (c) for avoiding the inuence of
nonuniqueness given in Remark 3.4.
As we will see later, it is sometimes advantageous to reconstruct
m
f instead
of f, with m > 1 an integer. The case m = 0 of course yields an approximation
to the function f itself. Using the convolution theorem (21) for the Fourier
transform we obtain, in a similar way as above,
e
m
f(x) =

R
n
e()[[
m

f()e
ix,
d
= [S
n2
[
1

S
n1

[[
m+1
e()

f()e
ix,
d d
= (2)
1
[S
n2
[
1

S
n1

[[
m+1

e()

f()e
iE

x,
d d
=

S
n1
(k P

f)(E

x) d, m 1, (46)
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 9
with the convolution kernel
k(y) = (2[S
n2
[)
1

m+1
P

e(y), y

. (47)
If e is a radial function, P

e and the convolution kernel k are independent of .

A corresponding formula for the Radon transform can be derived by using po-
lar coordinates in R
n
instead of (44). For rigorous proofs and general conditions
on e and f for which (46) is valid see [48], [90] and [59]. Of greatest interest
are the case m = 0, which gives the formulas for reconstructing the function f
itself, and the cases m = 1. Letting e yields the exact inversion formula

m
f(x) = (2[S
n2
[)
1

S
n1

m+1
P

f(E

x) d.
A desirable property would be the possibility of local reconstruction, i.e., re-
construction at a point should require only lines passing through a small neigh-
borhood of that point. Since the parameters and y

of a line passing
through a point x must satisfy the equation E

x = y, reconstruction according
to (46) will be local if the kernel k is supported in a small neighborhood of
the origin. However, for m even and

R
n
e(x) dx = 0,

k is not analytic, so k
cannot have compact support. Hence ordinary tomography is global, not local.
On the other hand, it follows from (47) and (41) that k has compact support if
m 1 is odd and e has compact support. This explains the interest in the cases
m = 1. Computing
1
f(x) consists of taking the average of all integrals over
lines passing through x. This was done in early imaging techniques preceding
CT. However, the result is a very blurry image of f which by itself is of limited
usefulness; see the bottom left picture in Figure 1. Current local tomography,
reviewed below, avoids this disadvantage by computing a linear combination of
f and
1
f.
If f is supported in the unit ball, and the source points a lie on a sphere
A with center in the origin and radius R > 1, then the approximate inversion
formula for the divergent beam x-ray transform reads as follows [90]:
e
m
f(x) = R
1

S
n1
D
a
f() ['a, `[ k(E

(x a)) d da, (48)

with m 1 and k as in (47).
We conclude this section with a few remarks on reconstruction algorithms.
The ltered backprojection algorithm is the most popular reconstruction method.
It is a computer implementation of the approximate reconstruction formulas
(46) and (48) for parallel-beam and fan-beam sampling, respectively. We will
discuss it in detail in a later section. For references on the ltered backprojection
algorithm see, e.g., [49].
The Fourier reconstruction algorithm uses the Fast Fourier transform to com-
pute
[
P

j
f() =

2

f(),

j
, j = 0, . . . , P1.
10 ADEL FARIDANI
0.5 0 0.5
0.5
0
0.5
0.5 0 0.5
0.5
0
0.5
0.5 0 0.5
0.5
0
0.5
0.5 0 0.5
0.5
0
0.5
Figure 1. Top left: Global reconstruction of density f(x) of calibration object.
Top right: Reconstruction of f. Bottom left: Reconstruction of
1
f. Bottom
right: Reconstruction of Lf = f +
1
f, with = 46.
In 2D this gives values of

f on a polar grid. These are now interpolated onto a
rectangular grid and a 2D inverse FFT is used to obtain f. This is much faster
than ltered backprojection, but the interpolation is problematic, i.e., prone to
cause artifacts in the reconstructed image. For further discussion and references
on methods to overcome these drawbacks see [66; 67].
Algebraic methods do not discretize an inversion formula or use the projection
slice theorem, but start from an ansatz f(x) =

N
i=1
c
i

i
(x) and then solve the
linear system
N

i=1
c
i
P

i
(y
k
) = g
jk
, j, k = 1, 2, . . .
for the unknown coecients c
i
. Here g
jk
= P

j
f(y
k
) are the measured data.
Often the basis functions are the characteristic functions of pixels or voxels, but
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 11
this is of course not the only choice. Indeed, the advantage of such methods lies in
their exibility, e.g., in incorporating irregular sampling geometries or available
a-priori information on f. The resulting linear systems are large and sparse
and require special (usually iterative) algorithms for suciently fast solution.
Stabilization can be achieved by limiting resolution or by stopping the iteration
before convergence is achieved (see, e.g., Figure V.12 in [62]).
Numerous other reconstruction algorithms have been developed. For a survey
see, e.g., [62, Chapter V] and [65; 66; 67].
5. Incomplete Data Problems and Detection of Singularities
Incomplete data problems arise when measurements of P

f(y) are unavailable

for a certain range of arguments (, y). In dimension 2 the most common exam-
ples are the limited angle problem, the exterior problem, and the interior prob-
lem. Assume that f has compact support contained in the unit disk. In the lim-
ited angle problem, measurements Pf(, s) are available only in an angular range
[
1
,
2
] with [
1

2
[ < . Note that because of Pf(, s) = Pf(+, s),
an angular range of is sucient for complete data. It follows from Theorem 3.1
that f is uniquely determined by the limited angle data. The problem is lack of
stability. We see from Theorem 2.2 that the data determine the Fourier transform

f() in the cone = (cos , sin ) : [

1
+ /2,
2
+ /2], R. Re-
constructing f is therefore equivalent to accomplishing an analytic continuation
of

f, and analytic continuation is severely ill-posed. A more detailed picture
emerges from the singular value decomposition of the limited angle transform
[54]. The severe ill-posedness is reected in exponentially decaying singular val-
ues. However, the spectrum splits into two parts, one with singular values close
to the singular values in the full-range case, and the other with singular val-
ues close to 0. The components of f corresponding to singular functions in the
rst part are therefore recoverable. The characterization of the unrecoverable
singular functions in [54] allows to predict and recognize typical reconstruction
artifacts.
In the exterior problem only line integrals P

f(s) with [s[ > a > 0 are avail-

able. Uniqueness holds in the measured region but stability is missing. The
singular value decomposition was given in [70; 73], and used to develop a recon-
struction algorithm [74; 75].
Finally, the interior problem is characterized by measurements in the range
[s[ a < 1. Uniqueness does not hold, not even inside the disk [x[ < a where
for each point x all integrals over lines passing through a small neighborhood
of x are measured. A singular value decomposition has been derived in [61].
Promising new methods for the interior problem also include the wavelet-based
approaches of [2; 80] and pseudolocal tomography [41]. The wavelet localization
method presented in [68] requires additional integrals over a small number of
lines not intersecting the disk [x[ < a.
12 ADEL FARIDANI
None of these problems provides both uniqueness and stability. It is now
natural to ask that if the function f itself cannot be recovered stably, what
features of f can? One answer to this question is provided by the singular value
decompositions, which tell us that components of f corresponding to singular
functions with large singular values are recoverable. Another approach is to ask
which singularities of f can be stably recovered. In many applications f can
be considered to be approximately piecewise constant with jump discontinuities
along the boundaries between dierent features. Identifying the singularities of
f thus allows to determine the shape of such features. A general answer has
been given in [75] based on the correspondence between the wavefront sets of f
and Pf. In the special case of f being smooth except for jump discontinuities
along a smooth curve , a singularity at a point x is detected stably if and
only if integrals over lines in a neighborhood of the tangent line to at x are
available. Applying this rule to the incomplete data problems mentioned above
yields that in the limited angle and exterior problems not all jumps can be stably
detected, since for any point x there are lines passing through x for which the
data are not available. On the other hand, in case of the interior problem one
can stably determine all singularities inside the disk [x[ < a. It is thus interesting
to note that the interior problem is the worst behaved of the three with respect
to uniqueness, but is the best behaved with respect to detecting singularities.
For general reconstruction methods where the reconstruction preserves the
stable singularities see [45]. Several methods have been suggested to detect
singularities directly from the line integrals without rst performing a recon-
struction [39; 75; 79].
A problem of great practical interest which still poses many open problems is
three-dimensional cone-beam reconstruction with sources on a curve. See, e.g.,
[97] for an inversion formula, [19] for a general stability result, [75] for conditions
to detect singularities, and [8; 18; 23; 57; 64] for reconstruction algorithms and
other developments. The approximate inversion formula (46) is very useful in
two dimensions, but not so in three dimensions. It needs integrals over all lines,
but in three dimensions the lines form a four parameter family, so (46) requires
far more data than should be needed to determine a function of three variables.
In practical 3D tomography an x-ray source moves on a curve, so only integrals
over lines intersecting the curve are measured. This situation is modelled by
the divergent beam x-ray transform D
a
f(), where a runs through the curve .
The conditions on the source curve for stable inversion are restrictive, so that
in most practical situations one has an incomplete data problem. Based on the
exposition in [75], we now state the precise denitions for the microlocal concepts
mentioned above and apply them to this situation. The reader interested in a
deeper treatment may wish to rst read [75] and [27], and then proceed to articles
such as [1; 24; 25; 26; 72].
The following concept of a wavefront set uses the fact that the Fourier trans-
form of a C

0
function decays rapidly. A local version of this fact can be obtained
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 13
by rst multiplying f with a C

0
cut-o function with small support, and see-
ing if the Fourier transform of the product f decays rapidly. The wavefront
set gives even more specic, so-called microlocal information, inasmuch as it
identies the directions in which the Fourier transform of f does not decrease
rapidly.
Definition 5.1. Let f be a distribution and take x
0
,
0
R
n
with
0
= 0.
Then (x
0
,
0
) is in the wavefront set of f if and only if for each cut-o function
in C

0
with (x
0
) = 0, the Fourier transform of f does not decrease rapidly
in any conic neighborhood of the ray t
0
, t > 0.
Loosely speaking, we say that a singularity of f can be stably detected from
available x-ray data, if there exists a corresponding singularity of comparable
strength in the data. The strength of a singularity can be quantied microlocally
using Sobolev space concepts:
Definition 5.2. A distribution f is in the Sobolev space H
s
microlocally near
(x
0
,
0
) if and only if there is a cut-o function C

0
(R
n
) with (x
0
) = 0
and function u() homogeneous of degree zero and smooth on R
n
` 0 and with
u(
0
) = 0 such that u()
[
(f)() L
2
(R
n
, (1 +[[
2
)
s
).
First, one localizes near x
0
by multiplying f by , then one microlocalizes near

0
by forming u

f. and sees how rapidly

f decays at innity.
For 3D tomography with sources on a curve we have the following result:
Theorem 5.3 [75, Theorem 4.1], [1], [24]. Let be a smooth curve in R
3
and f a
distribution whose support is compact and disjoint from . Then any wavefront
set of f at (x
0
,
0
) is stably detected from divergent beam x-ray data Df with
sources on if and only if the plane P through x
0
and orthogonal to
0
intersects
transversally.
If data are taken over an open set of rays with sources on , then a ray in P
from to x
0
must be in the data set for stable detection to apply. In these cases
f is in H
s
microlocally near (x
0
,
0
) if and only if the corresponding singularity
of Df is in H
s+1/2
.
We see that the corresponding singularities of Df are weaker by
1
2
Sobolev order,
but this is still strong enough to allow stable detection in practice.
Theorem 5.3 allows to analyze singularity detection in 3D tomography in the
same way as described above in the two-dimensional case.
It is now interesting to ask if the available numerical algorithms can actually
reconstruct all the stable singularities. The results for a general class of restricted
x-ray transforms obtained in [24; 25; 26] show that microlocal analysis is also
a powerful tool to answer such a question. For an introduction to these results
see [27]. Explicit calculations analysing an algorithm for contour reconstruction
proposed in [57] and some closely related methods have recently been given in
[38; 47].
14 ADEL FARIDANI
The algorithm of [57] aims to reconstruct the function
f
R
= D

Df, (51)
with
D

g(x) =

|x a|
1
g

a,
x a
|x a|

da.
The results in [24; 38; 47] show that the wavefront set of f
R
consists of two
parts. The rst part contains those wavefronts (x, ) of f for which the plane
through x and normal to intersects . The second part may introduce new
singularities, namely on the line from a source point a to x, the location of
the original singularity in f. This will happen if the plane through x and normal
to contains a and the tangent vector to at a is orthogonal to , i.e., the plane
touches but does not intersect transversally. In addition, the acceleration
vector of the curve at a should not be orthogonal to . The Sobolev strength of
these additional singularities is the same as the reconstructed part of the original
wavefront set [25; 26; 38], and they appear as artifacts in numerical simulations
[17; 35; 38].
An advantage of the formula (51) is that reconstruction of f
R
is local, i.e.,
reconstruction at a point x requires only integrals over lines close to x. In [57]
it is shown that f
R
approximates f in certain cases. Another, and apparently
the historically rst method for 3D local tomography is an adaptation of the
algorithm by Feldkamp et al. [18] developed by P.J. Thomas at the Mayo Clinic.
While the details of this algorithm have not been published, it has been used in
various papers, e.g., [94; 14].
6. Local Tomography
Often only part of an object needs to be imaged. In this case it would be
preferable if only integrals over lines which intersect the region of interest (ROI)
are needed. We know from the discussion of the interior problem above that
we dont have uniqueness. However, it turns out that the null functions are
nearly constant inside the ROI, and we know already that all singularities inside
the ROI are stably determined. Several approaches have been developed in the
literature. For example, the wavelet based method of [68] exploits the fact that
the error contains mostly low frequencies, and that these can be recovered by
supplementing the data with relatively few measurements outside the ROI. The
method of [80] which will be discussed below, extrapolates the missing data
and aims at reconstruction of f up to a constant error. Another method using
extrapolation of the missing data is described in [62, ' VI.4].
Lambda tomography, the main topic of this section, was introduced indepen-
dently in [98] and [90]. It does not attempt to reconstruct the function f itself
but instead produces the related function Lf = f +
1
f. This has the ad-
vantage that the reconstruction is strictly local in the sense that computation
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 15
of Lf(x) requires only integrals over lines passing arbitrarily close to x. Local
tomography has found applications in medical imaging [94], nondestructive test-
ing [85; 99], and microtomography [15; 16; 83; 86]. Extensions to more general
settings have been presented in [37; 45]. Other approaches include [2] and [41].
Intelligent use of Lambda tomography requires knowledge of what kind of
useful information about f is retained in Lf. Let us consider an example. The
upper left of Figure 1 shows an ordinary, global reconstruction of the density
function f of a calibration object used by the Siemens company. The data come
from an old generation Siemens hospital scanner. Units are such that the radius
of the global reconstruction circle is one. The gure displays the reconstruction
inside the rectangle [0.5, 0.5]
2
. The scanning geometry is a fan-beam geometry
(710) with source radius R = 2.868, p = 720 source positions, and 2q = 512
rays per source. The upper right of Figure 1 shows a reconstruction of f.
Reconstructions of
1
f and Lf = f + 46
1
f are shown in the lower left
and lower right, respectively. The similarity between the images of f and f is
at rst glance surprising. We expect that a good local reconstruction method
should detect the singularities of f, since these are stably determined by the
data. Indeed, since is an invertible elliptic pseudo-dierential operator, f and
f have precisely the same singular set. However, we see that f is cupped
where f is constant, and that the singularities are amplied in f. The image of

1
f by itself seems less useful, but it provides a countercup for the cup in f.
Thus, the image of Lf shows less cupping and looks even more similar to f than
the image of f. For example, the image of Lf indicates that the density just
inside the boundary of the object is larger than the density outside the object,
while this can not be clearly seen from the image of f. To achieve this eect,
a good selection of is necessary. A prescription for selecting can be found in
[15].
A more detailed understanding of images of f or Lf is obtained from study-
ing quantitative relations between f,
1
f and f [14; 15]. Some of the results
for f are as follows. For corresponding results on
1
see [14].
Theorem 6.1. ([14]) Let X and Y be measurable sets, n 2, and let (1 +
[x[)
1n
f be integrable.
(a) If f
r
(x) = f(x/r), then f
r
(x) = r
1
f(x/r).
(b)

X
(x) > 0 on X

, and < 0 on X
c
;

X
c =

X
.
(c)

X
is subharmonic (Laplacian 0) on X

, and superharmonic on X
c
.
This implies that

X
cannot have a local maximum in X

, nor a local min-

imum in X
c
.
(d) If x is outside the support of f, then
f(x) =
1 n
[S
n2
[

R
n
[x y[
1n
f(y) dy.
16 ADEL FARIDANI
(e) Near X we have [

X
(y)[ 1/d(y, X), where d(x, X) is the distance of
x to X.
Remark 6.2. The results for

X
are of practical interest, since in many appli-
cations the function f can be modeled as a linear combination of characteristic
functions.
As a consequence of (a), small features are amplied in images of f. This
is benecial for the detection of small, low contrast details. For example, in
Figure 1 the small holes in the rectangular pieces are more clearly visible in
the image of f than in the image of f.
Part (b) indicates that the jumps of f at discontinuities of f have the same
direction as those of f.
Part (c) explains why there are no oscillations which could be mistaken for
actual details in images of f.
Part (d) shows that if f has compact support, then f cannot. This means
that there are global eects in images of f in the sense that the value of
f(x
0
) depends on the values of f everywhere. However, Part d) implies that
f(x) will decay at least as O([x[
1n
) for [x[ . More rened estimates
are derived in [15].
Part (e) shows that a nite jump in f causes an innite jump in f. In a
neighborhood of X, f is not a function but a principal value distribution
[14].
While Lf retains the signs of jumps in density, it does not give direct informa-
tion about the size of these jumps. However, such information about density
dierences may be extracted in certain cases. In the following we will describe
several methods. We assume that f is a linear combination of a smooth function
and of characteristic functions of sets:
f = f
0
+

c
i

X
i
, (61)
with f
0
C

0
, [X
i
[ = 0, X
i
= X

i
, and X

i
X

j
= ? if i = j.
We are interested in estimating c
j
c
i
when X
j
, X
i
have a common nontrivial
boundary ,
= X
i
X
j
W = ?, W = (X
i
X
j
)

.
We rst discuss the method developed in [15]. It is based on Theorem 6.3
below. The theorem expresses the fact that for x suciently close to , we have
c
j
c
i
=
f(x)

X
j
(x)
+O(d), [c
j
c
i
[ =
[f(x)[
[

X
j
(x)[
+O(d
2
),
where d is the distance from x to .
Recall that a set Y has curvature 1/r along a subset Y
0
of Y if for each
point y Y
0
there are open balls B Y and B

Y
c
of radius r with y

BB

.
The distance of a point x to a set Y is denoted by d(x, Y ).
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 17
Theorem 6.3 [15]. Let f be as in (61). Fix i, j, let W = (X
i
X
j
)

and
assume that
= X
i
X
j
W = ?.
Let X
j
have curvature 1/r, r > 0, along a closed subset
0
of . Let x W`
be such that d(x, X
j
) = d(x,
0
) = d. Then

f(x)

X
j
(x)
(c
j
c
i
)

F
1
(d/r)

max [f
0
[ +C
1
max
k=j
[c
k
[
d(x, W)

d, (62)

[f(x)[
[

X
j
(x)[
[c
j
c
i
[

F
2
(d/r)

max [f
0
[ +C
2
max
k=j
[c
k
[
d(x, W)
2

d
2
. (63)
The constants C
1
and C
2
and the functions F
1
, F
2
can be given explicitly. For
example, for n = 2, we have C
1
= 2 and C
2
= 3. Furthermore,
lim
t0
+
F
1
(t) = lim
t0
+
F
2
(t) = .
The error terms on the right-hand sides of (62) and (63) indicate that in general
the estimate (63) should be more accurate than (62) when d is small. The
terms involving d(x, W) come from the inuence of other boundaries than .
Numerical implementation of (62) or (63) requires computation of recon-
structions of f and

X
j
inside a region of interest R. In the following let

f and

X
j
denote these reconstructions, rather than the functions f and

X
j
themselves. It is also assumed that f has the form (61) with sets X
i
such
that X
i
R or X
i
R = ?. This entails no loss of generality since any set X
i
violating this condition can be replaced by the two sets X
i
R and X
i
R
c
.

X
j
is computed using simulated x-ray data, after X
j
has been found from

f. In principle, either (62) or (63) can be used, but as discussed above the
method based on (63) is likely to be more accurate. This gives only [c
j
c
i
[,
but since the sign of c
j
c
i
is preserved in f, this is all that is needed.
The method consists of the following steps:
(i) Compute

f from local data inside a region of interest R.
(ii) Determine X
j
by nding X
j
from

f.
(iii) Compute

X
j
inside the region of interest from simulated x-ray data, using
the same sampling geometry as for the original data.
(iv) If x X
j
, take the ratio [

f(x)[/[

X
j
(x)[ as an estimate for the
magnitude of the density jump. It is advisable to use suitable averages of the
gradients over points near the boundary of X
j
instead of the gradient at a
single point x. This reduces eects due to measurement noise.
Following [12], we demonstrate the method with x-ray data from a medical scan-
ner. Additional applications of this method are reported in [15; 12; 83].
The top panel of Figure 2 shows again the global reconstruction of the cali-
bration object. The region of interest R is indicated by the box. The picture
in the lower left shows the local reconstruction

f inside the region of interest.
18 ADEL FARIDANI
0.5 0 0.5
0.5
0
0.5
0.130.120.110.10.09
0.01
0.02
0.03
0.04
0.05
0.130.120.110.10.09
0.01
0.02
0.03
0.04
0.05
Figure 2. Top: Global reconstruction of density f(x) of calibration object. Box
indicates region of interest R. Bottom left: Local reconstruction

f inside
region R = [0.14, 0.08] [0.008, 0.058] (contained in the small box in the
top panel). Bottom right: Result of automatic edge detector applied to the
image of

f shown in top right. Pixels where an edge is detected are white.
The goal is to determine the density dierence between the small hole and its
surroundings. Let X
j
be the characteristic function of the hole.
Finding X
j
involves edge detection. This is currently done by the user of the
method, who species the vertices of a polygon approximating X
j
. Matlabs
image processing toolbox allows to do this selection conveniently. Our software
gives the user the option to use either the reconstruction

f itself, an image of
[

f[, or the result of a standard automatic edge detector, for specifying the
polygon. Which image is most convenient diers from application to applica-
tion. In [15], where the method was applied to projection data from a human
pelvis, the gradient image was most convenient. Here the result of the auto-
matic edge detector (Matlabs edge command) applied to the reconstruction

f
is satisfactory, as can be seen from the lower right of Figure 2.
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 19
orig. data sim. data
0.4 0.6 0.8 1
1750
1800
1850
1900
t
d
(
t
)
0.4 0.6 0.8 1
0
50
100
150
200
250
300
t
N
(
t
)
Figure 3. Top left: Image of |

f| inside R, f being the density function of the

calibration object. Box indicates region R

. Top right: Image of |

X
j
| inside
R. Box indicates region R

. Bottom left: Graph of estimated density dierence

d(t) for 0.5 d(t) 0.95. Bottom right: Number N(t) of points contributing
to the averages of |

f| (solid line), and of |

X
j
| (dotted line).
The top left of Figure 3 shows an image of [

f[ inside the region of in-

terest. The box indicates the subregion R

containing the part of the boundary

which will be used to estimate the density jump. Here we average over the whole
boundary of the small hole. Matlabs imcrop command allows convenient selec-
tion of R

by the user. The corresponding image for [

X
j
[ is shown in the
top right part of the gure. This reconstruction was computed from simulated
x-ray data using the same scanning geometry as in the reconstruction from real
data. Having no specic information on the detectors, the eect of the positive
detector width was modeled by averaging line integrals over the angular distance
between two adjacent detectors.
The following averaging procedure was used to estimate the density dierence.
Let M be the maximum of [

f[ in R

. Take the average of [

f(x)[ over
all points x in R

such that [

f(x)[ > tM for some t (0, 1). The same

averaging procedure is applied to [

X
j
(x)[, with M replaced by the maximum
20 ADEL FARIDANI
t = 0.6: orig. data t = 0.9: orig. data t = 0.6: sim. data t = 0.9: sim. data
orig. data sim. data
Figure 4. Two left panels: White points are those where |

f(x)| exceeds
t max
yR
|

f(y)| for t = 0.6 (leftmost) and t = 0.9 (middle left). Two right
panels: White points indicate where |

X
j
(x)| > t max
yR
|

X
j
(y)| for
t = 0.6 (middle right) and t = 0.9 (rightmost).
of [

X
j
[ in R

. The ratio of the two averages is the estimate d(t) for the
density dierence. This estimate depends on the choice of t. If t is too close to 1,
the average is taken over very few points, while a small t will include points too
far from the boundary. So t should be chosen small enough to have suciently
many points for averaging, but large enough so that only points close to the
boundary contribute to the averages. The graph in the bottom left of Figure 3
displays the estimated density dierences d(t) for 0.5 t 0.96. The bottom
right shows the numbers N(t) of points contributing to the averages of [

f[
(solid line), and of [

X
j
[ (dotted line). If t 0.9 very few points contribute
to the average of [

f(x)[. The corresponding estimates are therefore likely to

be unreliable. On the other hand, for t < 0.6 points away from the boundary
begin to contribute to the average. The binary images in Figure 4 show the
location of the points considered for the averages in the case of t = 0.6, and
t = 0.9, respectively. Hence reasonable estimates for the density dierence are
the values of d(t) for 0.6 t 0.9. These values lie between 1782 and 1898. The
global reconstruction indicates that the true density dierence is approximately
1854. Hence all of the acceptable estimates lie between 96 and 102 per cent of
the true value. Since the x-ray data have been scaled by an unknown factor, the
reconstructed values do not correspond to Hounseld numbers.
When implementing the method described above a few parameters have to
be chosen judiciously and a few comments on how to do this are in order. If
the ltered backprojection algorithm is used the reconstruction

f will, apart
from discretization errors, be equal to e f. The point-spread function e is
assumed to be of the form e = e
r
dened in (75) and (78) below, usually with
= 11.4174; compare the appendix of [14]. Choosing the point spread radius r
entails the usual tradeo between stability (larger r) and high resolution (smaller
r) and will depend on the number of measured line integrals as well as on the
accuracy of these measurements. In the example above r = 0.0225, which means
that the minimum of the convolution kernel falls on the second detector; see [14,
' 9].
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 21
The other important parameter is the spacing h of the grid of points where
[

f[ [(e f)[ is computed. [(e f)[ varies rapidly near a boundary

and h has to be suciently small so that the maximum of the gradient at the
gridpoints is close to the overall maximum. The special case with f the char-
acteristic function of a halfspace seems to give sucient guidance for practical
purposes. If f is the characteristic function

H
of a halfspace H then both f
and e f can be computed as follows. For x H one has ([14, Theorem 4.5])

H
(x) = (

d(x))
1
,
where

d(x) is the signed distance of x from H, i.e.,

d(x) = d(x, H) for x H,
and

d(x) = d(x, H) for x H. Computing e

H
involves the Radon
transform (22) of e. Since e is radial, R

e does not depend on . Therefore the

subscript will be suppressed and Re(s) viewed as a function of the one variable
s. It now follows that
e

H
(x) = HRe(

d(x)), (64)
where H denotes the Hilbert transform as dened in (42). Observing that for
functions f of one variable f(t) =
d
dt
Hf(t) gives
[(e

H
(x))[ = [Re(

d(x))[. (65)
Inspection of the graph of Re for e as in (78) and = 11.4174 now yields
that the width of the interval where [Re(t)[ > 0.98(max
sR
[Re(s)[) is approx-
imately r/20. Hence a rule of thumb for choosing h would be to set h = r/20.
The method described above can be simplied by making a priori assumptions
about the unknown boundary X
j
, so that the polygonal approximations and
the reconstruction from simulated data are avoided. For example, X
j
could be
assumed to be a halfspace H. Replacing f and

X
j
in (62) and (63) by
e f and e

H
, and using (64) and (65) gives the approximate formulas
c
j
c
i

e f(x)
HRe(

d(x))
, [c
j
c
i
[
[(e f(x))[
[Re(

d(x))[
. (66)
These two formulas are the basis of two of the algorithms proposed in [40; 78]
for dimension n = 2; see formulas (2.17) and (2.21) in [40]. The derivation in
[40; 78] is dierent and employs an asymptotic expansion for f, where f is
smooth except for jumps across smooth boundaries.
Another method to compute jumps of a function from essentially local data
is pseudolocal tomography [41; 78]. We follow the presentation given in [4] which
allows to understand the numerical implementation of this method in the frame-
work of equations (66).
The starting point for pseudolocal tomography is the two-dimensional inver-
sion formula from (43), which we repeat here:
f(x) =
1
4

S
1
HP

f('x,

`) d =
1
4
2

2
0

R
d
ds
P

f(s)
'x,

` s
ds d.
22 ADEL FARIDANI
Now truncate the Hilbert transform integral and dene
f
d
(x) =
1
4
2

2
0

x,

+d
x,

d
d
ds
P

f(s)
'x,

` s
ds d.
It was shown in [41] that f f
d
is continuous, hence f
d
has the same jumps
as f. Recalling that P

f('x,

`) is the integral over the line in direction

which passes through x, we see that computation of f
d
(x) requires only integrals
over lines with distance at most d from x (pseudo-local reconstruction.)
In practice one has to use an approximate inversion formula and computes
f
d,r
(x) = e
r
f
d
(x) =

2
0

k
d,r
('x,

` s)P

f(s) ds d,

k
d,r
(t) =
1
4
2

t+d
td
d
ds
P

e
r
(s)
t s
ds,
where e
r
is a radial function satisfying
e
r
(x) = r
2
e
1
(x/r), e
1
(x) = 0 for [x[ > 1,

R
2
e
1
dx = 1.
Note that

k
d,r
(t) = 0 for [t[ > d +r, i.e., computation of f
d,r
(x) requires inte-
grals over lines with distance at most d+r fromx. Furthermore, lim
d

k
d,r
(t) =
(4)
1
HP

e
r
(t). Hence (46) gives that lim
d
f
d,r
(x) = e
r
f(x). Indeed,
the convolution kernel k
d,r
can be obtained from the kernel k in (47) by letting
m = 0 and truncating the Hilbert transform integral. The relation f
d,r
= e
r
f
d
was shown in [41].
It turns out that for small d (i.e., local data), f
d
is signicantly dierent
from zero only in a narrow region near a boundary [41, Figure 3], and that the
convolution with the point spread function e
r
alters these values so much that the
jumps cannot just be simply read o the reconstructed image f
d,r
. We need an
algorithm to obtain information about the jumps of f. The methods developed
by Katsevich and Ramm [41; 78] can be understood in the framework developed
for Lambda tomography. According to equations (66) we have, for x close to ,
c
j
c
i

E f(x)
HRE(

d(x))
, [c
j
c
i
[
[E f(x)[
[RE(

d(x))[
. (67)
The task now is to nd E
d,r
such that E
d,r
f = f
d,r
= e
r
f
d
.
Proposition 6.4 [78; 4]. Dene E
d,r
by
P

E
d,r
= (P

e
r
) M
d
with
M
d
(s) =
1

ln([s/d[)

[d,d]
(s).
Then
f
d,r
(x) = E
d,r
f(x).
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 23
With this result, equations (67) give
c
j
c
i

f
d,r
(x)
HRE
d,r
(

d(x))
, [c
j
c
i
[
[f
d,r
(x)[
[RE
d,r
(

d(x))[
, (68)
and we can apply the same algorithms for recovering the jumps as in Lambda
tomography.
Some remarks are in order.
1. Because E
d,r
is radial, we have HRE
d,r
(0) = 0, so f
d,r
(x) 0 for x . This
makes it dicult to use the rst relation (68) in practice, since nding

d(x)
is not easy. See the algorithm given [41] and further discussed in [4]; see also
[17]. But since [f
d,r
[ is maximal for x , one can nd the points x by
looking for the local maxima of [f
d,r
[ and then estimate the jump by
[c
j
c
i
[
[f
d,r
(x)[
[RE
d,r
(0)[
, x .
This approach has essentially been used in [78] for pseudolocal tomography
and in [40] for Lambda tomography.
2. The property that f
d
has the same jumps as f is not used in the algorithm.
3. E
d,r
(x) = 0 for [x[ > d + r. Hence our derivation of the algorithm is only
justied for d + r suciently small. In practice the method seems to work
also for much larger values of d +r.
Another method which can be used for region of interest tomography is the
wavelet-based multiresolution local tomography of [80]. It illustrates the possible
uses of wavelets to localize the x-ray transform, or, more precisely, to sepa-
rate the features which are well determined by local data from those who are
not. For readers unfamiliar with wavelets we have collected some basic facts in
Appendix A.
Consider a (two-dimensional) multiresolution analysis of nested subspaces V
j
,
j Z of L
2
(R
2
). We assume a dilation matrix M = 2I (see Denition A.2
below), where I is the identity matrix, and use the notation
f
j,k
(x) = 2f(2
j
x k) for j Z, k = Z
2
, x R
2
(compare (A1) in the Appendix). Let be the scaling function and

, =
1, 2, 3 the associated wavelets. Since the
j+1,k
, k Z
2
are a Riesz basis of the
subspace V
j+1
, a function f V
j+1
can be written as
f(x) =

kZ
2

A
j+1,k

j+1,k
(x).
The so-called approximation coecients

A
j,k
are given by

A
j,k
= 'f,

j,k
`
24 ADEL FARIDANI
where ', ` denotes the inner product in L
2
and

is the biorthogonal scaling
function (Denition A.4). Alternatively we can use the relation V
j+1
= V
j
+W
j
and obtain the expansion
f(x) =

kZ
2

A
j,k

j,k
(x) +
3

=1

kZ
2

j,k

j,k
(x).
We can interpret the rst sum as an approximation to f in V
j
V
j+1
, i.e.,
at a lower resolution. The second sum supplies the missing detail information.
Therefore the coecients

j,k
= 'f,

j,k
`
are called detail coecients. The Fast Wavelet Transform and its inverse (see
Theorem A.6) allow ecient computation of the

A
j,k
and

D

j,k
, k Z
2
from the

A
j+1,k
, k Z
2
, and vice versa.
We now observe that the approximation and detail coecients can be com-
puted directly from the x-ray data. Let f

(x) = f(x). Then

A
j,k
= 'f,

j,k
` = (f (

j,0
)

)(2
j
k),
and Similarly

j,k
= 'f,

j,k
` = (f (

j,0
)

)(2
j
k).
Hence we can use the approximate inversion formula (46) with e(x) = (

j,0
)

(x)
and reconstruction on the grid x = 2
j
k, k Z
2
, to obtain the approxima-
tion coecients directly from the x-ray data. For the detail coecients we let
e = (

j,0
)

. Alternatively one could rst compute the approximation coecients

A
j+1,k
by letting e(x) = (

j+1,0
)

(x) and choosing the ner grid x = 2

j1
k,
k Z
2
, and then use the Fast Wavelet Transform to obtain the approximation
and detail coecients at level j. Since the additional computational burden of
applying the Fast Wavelet Transform is negligible compared to the eort re-
quired for the reconstruction from the x-ray data, this alternative method seems
preferable, since only one point-spread function and corresponding convolution
kernel need to be used. However, if not all coecients on level j are needed, the
rst method will be more ecient.
The next question is how this approach allows to localize the x-ray transform,
i.e., to separate features which are determined by local data from those which
are not. It was observed in [68] that the detail coecients for suciently large
j should be well determined by local data, if the wavelets

have vanishing
moments. Lets see why.
Definition 6.5. A function f of n variables has vanishing moments of order
up to N, if

R
n
x

f(x) dx = 0
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 25
for all multiindices = (
1
, . . . ,
n
) with [[ =

i
N. Recall that the
i
are non-negative integers and that x

= x

1
1
x

2
2
. . . x

n
n
.
The nonlocality in the approximate inversion formula comes from the convolution
kernel k in (47) in case of m = 0. In two dimensions this is caused by the pres-
ence of the Hilbert transform in the formula k = (4)
1
P

e = (4)
1
HP

e.
The key observation now is that the Hilbert transform of a function with van-
ishing moments decays fast.
Lemma 6.6. ([80, p. 1418])Let f(t) L
2
(R) vanish for [t[ > A and have
vanishing moments of order up to N. Then, for [s[ > A,
[Hf(s)[
1
[s A[
N+2

A
A
[f(t)t
N+1
[ dt
The construction of wavelets with vanishing moments is well known, and it turns
out that the functions P

j,0
)

inherit the vanishing moments from the

.
Therefore the convolution kernels k = (4)
1
HP

j,0
)

will decay rapidly

outside the support of P

j,0
)

.
So we see that the detail coecients for large j, when

j,0
has small support,
are well determined by local data. This is intuitively plausible since these coef-
cients contain high-frequency information, and we know already from Lambda
tomography that high-frequency information is well-determined. So the non-
locality shows its greatest impact in the approximation coecients. Since the
scaling function satises

(x) dx = 1, its zero order moment does not vanish.
One could still choose

so that the moments of order 1 through N vanish. It
is shown in [80, p. 1419] that in such a case the resulting convolution kernel k
satises
[k(s)[ = O(s
2
) +O(s
N3
).
It seems that this does not achieve much, since we cannot remove the leading
O(s
2
) term. Nevertheless, the authors of [80] found that some scaling func-
tions having vanishing moments lead to convolution kernels with suciently
rapid decay for practical purposes. These scaling functions where found from
one-dimensional scaling functions by the method of Denition A.8. In their re-
constructions the authors of [80] also extrapolated the missing data by constant
values, thus reducing cupping artifacts. While it is suggested in [80] to rst com-
pute the approximation and detail coecients at level j and then use an inverse
fast wavelet transform to obtain the approximation coecients at level j + 1,
numerical tests in [87] indicated that the simpler approach of directly computing
the approximation coecients at level j +1 yields equivalent results. We observe
that this can be done without using wavelet theory, namely just by specifying
the particular point spread function e = (

j+1,0
)

in (46).
26 ADEL FARIDANI
7. Sampling the 2D X-Ray Transform
We rst consider the parallel-beam geometry in two dimensions. Our analysis
of sampling and resolution will use techniques from Fourier analysis. These
require both the domain of Pf as well as the sampling sets to have a group
structure. In 2D we parameterize S
1
by = (cos , sin ), and let

=
(sin , cos ). We write Pf(, s) for Pf(, s

) and consider Pf to be a
function on the group T R, where T denotes the circle group. We take the
interval [0, 2) with addition modulo 2 as a model for T.
Recall that for xed , the values of Pf(, s) for dierent s correspond to
integrals over a collection of parallel lines. We rst consider the case where Pf
is measured at points
(
j
, s
jl
), j = 0, . . . , P1, l Z.
Since for each angle
j
we measure integrals over a collection of parallel lines
l(
j
, s
jl
), such an arrangement is called a parallel-beam geometry. We would like
the set of points (
j
, s
jl
) to be a discrete subgroup of T R, and for practical
reasons we require that more than one measurement is taken for each occurring
angle
j
. We call a sampling set which satises these requirements an admissible
sampling lattice (ASL). There are several ways to parameterize such lattices
[11; 13; 16]. Here we use the parametrization given in [16]. If L is an ASL, then
there are d > 0 and integers N, P, such that 0 N < P and
L = L(d, N, P) =
(
j
, s
jl
) :
j
= 2j/P, s
jl
= d(l +jN/P), j = 0, . . . , P1; l Z . (71)
We see that P is the number of angles (views). For each view angle
j
the
values s
jl
, l Z, are equispaced with spacing d, hence d is the detector spacing.
The parameter N characterizes an angle dependent shift of the detector array.
We also see that there are P dierent lattices for given parameters d and P.
The most important lattices are the standard lattice
L
S
= (
j
, s
l
) :
j
= 2j/P, s
l
= d l, j = 0, . . . , P1, l Z ,
which is obtained by letting N = 0, and the interlaced lattice
L
I
= (
j
, s
jl
) :
j
= 2j/P, s
jl
= d(l + j/2), j = 0, . . . , P1, l Z ,
where P is even and N = P/2. We see that for the standard lattice the detector
positions s
l
do not change with the angle of view. For the interlaced lattice the
detector array is shifted by one-half of a detector spacing when going from one
angle of view to the next.
In practice one chooses P = 2p for both lattices, and for the interlaced lattice
one lets p be even. Then, because of the symmetry relation
Pf(, s) = Pf( +, s), (72)
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 27
only the angles
j
[0, ) need to be measured. It can be shown that among all
admissible lattices the standard and interlaced lattices are the only ones which
fully exploit this symmetry [16].
We now describe the implementation of the ltered backprojection algorithm,
which is based on discretizing the approximate inversion formula (46) with the
trapezoidal rule. In two dimensions (46) reads
e
m
f(x) = (
m
e) f(x) =

2
0

R
k('x,

` s)Pf(, s) ds d,
k(s) =
1
4

m+1
P

e(s).
(73)
We assume that we have sampled Pf on an admissible lattice. Discretizing (73)
with the trapezoidal rule gives
e f(x)
2
P
P1

j=0
Q
j
('x,

j
`), Q
j
(t) = d

l
k(t s
jl
)Pf(
j
, s
jl
),
with
j
, s
jl
as in (71), and

j
= (sin
j
, cos
j
). We assume that f is sup-
ported in the unit disk. Hence the sum in the discrete convolution is nite.
The reconstruction is usually computed for values of x on a rectangular grid
x
m
1
m
2
= (m
1
/M
1
, m
2
/M
2
), [m
i
[ M
i
. Since computing the discrete convolu-
tion Q
j
('x,

j
`) for each occurring value of 'x,

j
` would take too long, one rst
computes Q
j
(iH), [i[ 1/H, and then obtains an approximation I
H
Q
j
('x,

j
`)
for Q
j
('x,

j
`) by linear interpolation with stepsize H. We assume that
H = d/(N

m) with 0 < m, N

Z and N

N/P Z.
This gives H = d/m for the standard lattice (N

= 1) and H = d/(2m) for the

interlaced lattice (N

= 2). Then the eect of interpolating the convolution is

the same as replacing the kernel k with the piecewise linear function I
H
k which
interpolates k at the points Hl, l Z; see, e.g., [11, p. 84]. Hence the algorithm
computes the function
f
R
(x) =
2
P
P1

j=0
I
H
Q
j
('x,

j
`)
=
2d
P
P1

j=0

lZ
I
H
k

'x,

j
` s
jl

Pf (
j
, s
jl
) . (74)
If e is not radial, k and I
H
k will depend on , which is not explicitly reected in
our notation. Matlab source code implementing (74) for the standard lattice is
provided for illustrative purposes (see Appendix B).
In practice one chooses a basic point spread function e
1
and then controls the
resolution by using a scaled version
e(x) = e
r
(x) = r
2
e
1
(x/r). (75)
28 ADEL FARIDANI
The corresponding kernels scale as
k
r
(s) = r
2m
k
1
(s/r). (76)
Examples for point spread functions and kernels are as follows. A popular choice
for global tomography (m = 0) is the SheppLogan kernel [62, p. 111]:
k
1
(s) =
1
2
3
(/2) s sin s
(/2)
2
s
2
. (77)
This kernel is bandlimited with bandwidth 1, i.e., the Fourier transform

k
1
() =
1
2
(2)
3/2
[[
sin()

[1,1]
()
vanishes for [[ > 1. It follows that the scaled kernel k
r
(s) = r
2
k
1
(s/r) is
bandlimited with bandwidth b = 1/r, and the same is true for the corresponding
point-spread function e
r
and hence for e
r
f(x).
Since the kernel (77) does not have compact support, it is not useful for local
tomography. There we start with a point spread function
e
1
(x) =

C(1 [x[
2
)
+1/2
for [x[ < 1,
0 for [x[ 1,
C = ( + 5/2)/(( + 3/2));
(78)
see [14, p. 482]. The corresponding kernel for computing f (i.e., m = 1) is
K
1
(s) =

1
4
d
2
ds
2
P

e
1
(s) =
2(+5/2)

(+1)
(1s
2
)
1
(1(2+1)s
2
) for [s[ < 1,
0 for [s[ 1.
Here we used the fact that
2
= d
2
/ds
2
in one dimension. Now the kernel is
no longer bandlimited, but has compact support. The scaled kernels K
r
(s) =
r
3
K
1
(s/r) vanish for [s[ > r. The kernel for global tomography generated
by the point spread function (78) has a complicated analytic expression but a
quickly convergent series expansion [88].
Discretization of (48) yields the ltered backprojection algorithm for the
fan-beam sampling geometry. Recall that f is supported in the unit disk. Let
R > 1, a = R(cos , sin ), = (cos(), sin()), and xa = [xa[

cos(), sin()

. Writing Df(, ) for D

a
f(), (48) becomes
e
r

m
f(x) = R

2
0

/2
/2
Df(, ) cos()k
r
([x a[ sin()) d d. (79)
In order to evaluate the inner integral eciently, a homogeneous approximation
[46] is needed. It follows from (47) and (76) that
k
r
([x a[ sin()) = [x a[
2m
k
c
(sin()), c = r/[x a[.
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 29
The approximation consists in replacing c = r/[xa[ by a constant independent
of x and a. This gives
e
r

m
f(x) R

2
0
[x a[
2m

/2
/2
Df(, ) cos()k
c
(sin()) d d.
From here we can proceed as before by discretizing with the trapezoidal rule and
inserting an interpolation step. The standard sampling lattice for the fan-beam
geometry has the form
L
SF
=

(
j
,
l
) :
j
= 2j/p for j = 0, . . . , p 1,

l
= l arcsin(1/R)/q for l = q, . . . , q 1

. (710)
The reconstruction of
1
f is not unstable, so convolution with e
r
is not needed.
One can directly discretize the formula

1
f(x) = (R/4)

2
0
[x a[
1
Df(, ) cos d,
which comes from letting e
b
in (79).
In order to further analyze the parallel-beam algorithm we use Shannon sam-
pling theory. We begin with some denitions. We dene the Fourier transform
of a function g with domain T R by
g(k, ) =
1
2

2
0

R
g(, s)e
i(k+s)
ds d, (k, ) Z R.
The corresponding inverse Fourier transform is given by

G(, s) =
1
2

ZR
G()e
iz,
d (where z = (, s))
=
1
2

kZ

R
G(k, )e
i(k+s)
d. (711)
The reciprocal lattice L

Z R is dened as
L

(d, N, P) =

P N
0 2/d

Z
2
.
For g C

0
(T R), L = L(d, N, P) an ASL, K Z R compact dene
Sg(z) =
d
P

yL
g(y)

K
(z y), z T R,
where

K
is the inverse Fourier transform of the characteristic function of K.
We may view Sg as an approximation of g computed from sampled values of g on
the lattice L. The following classical sampling theorem gives an error estimate
for this approximation:
30 ADEL FARIDANI
Theorem 7.1. Let g C

0
(T R), L an ASL and K be a compact subset of
Z R such that the translates K +, L

are disjoint. Then

[g(z) Sg(z)[
1

(ZR)\K
[ g()[ d.
This result is an adapted version of the multidimensional sampling theorem of
Petersen and Middleton [69]. For a proof see, for example, [62, p. 62] or [11,
Theorem 2.2].
If supp( g) K, then g = Sg, i.e., g can be recovered exactly from its samples
on the lattice L.
In order for Sg to be close to g, the set K should be such that g is concentrated
in K. The sampling condition that the translates K + , L

be disjoint
requires the reciprocal lattice L

to be suciently sparse, and therefore the

sampling lattice L to be suciently dense.
A suitable set K for sampling the 2D x-ray transform was given by Natterer
[62] based on results by Lindgren and Rattey [81]:
Theorem 7.2. [62, p. 71] For b > 0 and 0 < < 1 let
K
0
(, b) =

(k, ) Z R : [[ < b, [k[ <

1
max([[, (1 )b)

. (712)
Let f C

0
(). Then

(ZR)\K
0
[

Pf()[ d
8

||>b
[

f()[ d +|f|
L
1
(, b), (713)
where (, b) decreases exponentially with b, satisfying an estimate
0 (, b) C()e
()b
with constants C(), () > 0.
Usually the parameter is chosen close to 1. The parameter b plays the role of
a cut-o frequency. If [

f()[ is suciently small for [[ > b then the right-hand

side of (713) will be small. In this case Theorem 7.1 imposes the condition that
the sampling lattice L be such that the translated sets K
0
(, b) +, L

are
disjoint. In terms of the lattice parameters d, N and P these conditions are as
follows:
For the standard lattice N = 0, and the reciprocal lattice L

equals
L

= (Pk
1
, 2k
2
/d) : k
1
, k
2
Z.
For reasons of eciency as discussed above we let P = 2p be even. The translated
sets K
0
(, b) +, L

are disjoint if and only if

d < /b and p > b/, P = 2p. (714)
For the interlaced lattice we again let K = K
0
(, b) as in (712). For this
lattice P = 2p and N = P/2 = p. We always let p be even, so that because of
the symmetry relation (72) only the angles
j
[0, ) need to be measured.
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 31
The reciprocal lattice is L

= (p(2k
1
k
2
), 2k
2
/d) : k
1
, k
2
Z. It turns out
[62; 11] that the sets K
0
(, b) + , L

are disjoint if either the conditions

(714) are satised, or if

b
< d
2
b
and p > max

2
d
,
(2 )b

, p even , P = 2p. (715)

We see that the interlaced lattice allows for a maximal detector spacing of
d = 2/b which is twice as large as the maximum allowed for standard lattice,
with only a moderate increase in p. Sampling conditions for a general admissible
sampling lattice have been given in [13]. We see that both (714) and (715) re-
quire p to be greater than the cut-o frequency b, which corresponds to condition
(b) in Remark 3.4 on avoiding the eects of nonuniqueness.
It remains to investigate if the theoretically superior resolution of the in-
terlaced lattice can be realized in practice. In principle there are two obvious
approaches: One could rst interpolate the missing data to a denser lattice and
then use any reconstruction algorithm. This approach has been successfully tried
in [10], but we will not discuss it here. The second approach would be to recon-
struct directly from interlaced data. It turns out that the ltered backprojection
algorithm is very suitable for this purpose [44; 11; 16]. We have the following
error estimate, which extends the results of [44] and [11].
Theorem 7.3 [16]. Let e be radial and suciently smooth, let f C

0
(R
2
) be
supported in the unit disk and let the sets K
0
(, b) + , for L

, be disjoint.
Then
f
R
(x) = G
H
e
m
f(x) +
4

i=1
E
i
(x),

G
H
() = (2)
1
sinc
2
(H[[/2)

1
([[/b),
[E
1
(x)[ c |

k|

||>b
[

f()[ d,
[E
2
(x)[ c|f|

||>b
[

k()[ d,
[E
3
(x)[ (2)
3/2
sup

b
b

1 sinc
2
(H/2)

k()[

lZ

+
2l
d

d,
[E
4
(x)[ c|f|

k|

(, b).
Here sinc x = (sin x)/x. The proof ([16]) is somewhat technical and will not be
given here. However, it is worthwhile to note that apart from the interpolation
step this is an estimate for the error of numerical integration by the trapezoidal
rule. The estimate for this error is based on the Poisson summation formula for
T R. This approach was rst applied in the present context by Kruse [44].
We will discuss the origin and importance of the four error terms. The error E
1
is the so-called aliasing error stemming from the fact that f is not bandlimited,
32 ADEL FARIDANI
since it has compact support. If the cut-o frequency b is chosen suciently
large, E
1
will be small. The sampling conditions then require that the number
of data available is commensurate with b. The error E
2
is present when k is not
bandlimited with bandwidth b. In global tomography, i.e., when m = 0, one can
chose e and k to be b-bandlimited, so that E
2
vanishes. In local tomography
one wishes k to have compact support, so k cannot be bandlimited.
The error E
3
is caused by the interpolation step and usually not a concern
when using the standard lattice. This can be explained as follows ([11]): Consider
the common parameter choice d = H = /b. Since

f() is assumed to be small
for [[ > b, only the term with l = 0 in the sum will be signicant, i.e., we have
for [[ b

lZ

f(( + 2l/d))

lZ

f(( + 2bl))

f()[.
Usually the density function f is non-negative so that [

f()[ has a sharply

peaked maximum at = 0 and is very small for [[ close to b. In such a case the
error E
2
will be small since the factor 1 sinc
2
(H/2) is small exactly where
[

f()[ is large.
The error E
3
is of much greater concern when the interlaced lattice is used.
Consider the choice of parameters d = 2/b, H = /b. Now the sum over l in
the estimate for E
3
may have three signicant terms for [[ < b:

lZ

f((+2l/d))

lZ
[

f((+bl))[

f((b))

f()

f((+b))

.
As discussed before the contribution of the term

f() is largely cancelled by the
factor (1 sinc
2
(H/2)). However, this is not the case for the other two terms.
E.g., let be close to b. Then, assuming again that

f is large near the origin,
[

f(( b))[ will be large and is not attenuated by the factor (1 sinc
2
(H/2))
which will be close to 1. Therefore we expect considerable reconstruction errors
for this choice of parameters. That this is indeed the case has been demonstrated
for global tomography in [44; 11]. Hence when using the interlaced lattice one
should choose H < b, so that (1 sinc
2
(H/2)) is small for [[ < b. Typical
choices in practice are H = /(16b) or smaller. Choosing H < /b has also a
cosmetic side eect. If a b-bandlimited convolution kernel is used whose Fourier
transform has a jump discontinuity at [[ = b (e.g., a scaled version of the Shepp
Logan kernel (77)), then ringing artifacts are caused by this discontinuity. In
case of the standard lattice with the parameter choice d = H = /b these arti-
facts are practically removed by the additional smoothing from the interpolation.
For smaller H this eect is lost. In this case the Fourier transform of k should
taper o continuously to zero if smooth images are desired [11].
Another eect of the interpolation is the additional ltering with G
H
. Since
this alters only the higher frequencies, it is usually not a concern. In any case,
the eect can be eliminated by choosing very small H.
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 33
The last error E
4
decreases exponentially with b, as indicated by the notation
(, b). Explicit estimates are as follows [16]. Let
=

1
2
[x[
2

3/2
.
For the standard lattice we have
[E
4
(x)[ c|f|

k|

b
e
b/
1 e

.
For the interlaced lattice we let b

= (1 )b/, and obtain with as above

[E
4
(x)[ c|f|

k|

be
b

1 +b

1 e

+
e

(1 e

)
2

.
In both cases the error decays exponetially with b, but in case of [x[ close to
1, when is close to zero, the above estimates indicate that the error should
be larger in case of the interlaced lattice due to the term involving (1 e

)
2
.
This eect has been observed in [11]. It causes a thin ring artifact in the region
[x[ 1, i.e., at the boundary of the reconstruction region. It can be eliminated
by choosing a smaller value for , i.e., by increasing the number of views p;
compare equations (714), (715).
Numerical experiments for both global and local tomography, with simulated
as well as real data [44; 11; 10; 12; 16] show that the higher eciency of the
interlaced lattice can at least be partly realized in practice. However, there
are also some drawbacks. There is a somewhat reduced stability because of
inaccurate convolutions. In case of the interlaced lattice the stepsize d = 2/b
is not small enough to allow accurate computation of the convolutions. Because
of the truly two-dimensional nature of the numerical integration, these errors
cancel out during the discrete backprojection step. The sensitivity with regard
to the interpolation stepsize H can be understood as coming from a disturbance
of these cancellations by the additional interpolation. A second drawback is a
requirement that the sampling condition with respect to the number of views
P has to be strictly observed. The aliasing caused by violating this condition
is usually quite moderate in case of the standard lattice but much more severe
for the interlaced lattice. This can be easily seen from the pattern in which the
translated sets K
0
+ begin to overlap when the sampling condition for p is
violated. Hence the interlaced lattice seems to be most useful when the detector
spacing is the main factor limiting resolution.
Good reconstructions from the interlaced lattice can also be obtained by using
the direct algebraic reconstruction algorithm [43], or by increasing the amount
of data through interpolation according to the sampling theorem [10]. Results
for the fan beam geometry can be found in [63; 64; 67]. As we have seen, the
interpolation step can introduce signicant errors in certain cases. It has also
been shown [64] that the interpolation can be avoided by chosing the points x
where the reconstruction is computed on a polar grid rather than on a rectangular
34 ADEL FARIDANI
grid, and interchanging the order of the two summations. This algorithm should
work well for the interlaced lattice [100] and is particularly benecial in case of the
fan-beam sampling geometry [64], since the method also avoids the homogeneous
approximation, whose inuence on the reconstruction is dicult to estimate.
Appendix A. Basic Facts about Wavelets
We give a brief introduction to multidimensional biorthogonal wavelets. The
discussion follows [87] and is based on the presentation in [96] for multidimen-
sional orthonormal wavelets. For more details on wavelets and lter banks, see
[7] or [95], for example.
Definition A.1. A lattice is a discrete subgroup of R
n
given by integral
linear combinations of a vector space basis v
1
, . . . , v
n
of R
n
.
Definition A.2. Let be a lattice and M be an n n matrix such that
M and that all eigenvalues of M satisfy [[ > 1. M is called the dilation
matrix. Let m = [ det M[. A multiresolution analysis with scaling function ,

(t) dt = 1, is a sequence of subspaces V

j
of L
2
(R
n
), j Z, satisfying the
following conditions:
1. V
j
V
j+1
,

V
j
= 0 and

V
j
= L
2
(R
n
).
2. f(t) V
j
if and only if f(Mt) V
j+1
.
3. f(t) V
0
if and only if f(t k) V
0
for k .
4. (t k), k is a Riesz basis of V
0
.
Convention. For notational convenience we set
f
j,k
(t) := m
j/2
f(M
j
t k), for k , j Z. (A1)
It follows from the denition of a multiresolution analysis that
j,k
(t), k
is a Riesz basis of V
j
.
Definition A.3. Consider a multiresolution analysis with lattice and dilation
matrix M. For j Z, let W
0
be such that V
1
is the direct sum of V
0
and W
0
.
Assume there are
1
, . . . ,
m1
W
0
such that

0,k
: = 1, . . . , m1, k
is a Riesz basis of W
0
. The

are called wavelets.

For j Z, let W
j
be the subspace with Riesz basis

j,k
, k , = 1, . . . , m1.
It follows that V
j+1
= V
j
W
j
, a direct sum but not necessarily orthogonal.
Definition A.4. Let V
j
,

V
j
be two multiresolution analyses corresponding to
the same lattice and dilation matrix M. Let and

, for = 1, . . . , m1,
be the scaling function and wavelets corresponding to V
j
. Let

and

, for
= 1, . . . , m1, be the scaling function and wavelets corresponding to

V
j
. The
multiresolution analyses are called biorthogonal if the following conditions hold
for j, j

Z, ,

= 1, . . . , m1, and k, k

:
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 35
(i) '

j,k
,

,k

` = (j, j

)(,

)(k, k

).
(ii) '

j,k
,
j,k
` = '

j,k
,

j,k

` = 0.
(iii) '
j,k
,

j,k
` = (k, k

).
For n = 1, M = 2, = Z and =

, multiresolution analysis becomes the
familiar one-dimensional, orthonormal case.
Since V
0
V
1
, and
1,k
, k is a basis for V
1
, there are coecients
F
0
(k), k such that
(t) = m
1/2

k
F
0
(k)
1,k
(t)
From condition (iii) above, it follows that F
0
(k) = m
1/2
',

1,k
`. Similarly, since

V
0


V
1
, and

1,k
, k is a basis for

V
1
,

(t) = m
1/2

k
H
0
(k)

1,k
(t)
where H
0
(k) = m
1/2
'

,
1,k
`. The above equations are called dilation equa-
tions. Similarly, the wavelets

must satisfy so-called wavelet equations:

(t) = m
1/2

k
F

(k)
1,k
(t),

(t) = m
1/2

k
H

(k)

1,k
(t),
where = 1, . . . , m1, F

(k) = m
1/2
'

1,k
`, and H

(k) = m
1/2
'

,
1,k
`.
The following lemma and theorems show how to decompose a function f into
its wavelet coecients and how to reconstruct f if its wavelet coecients are
known.
Lemma A.5. For j Z and = 1, . . . , m1, we have:

j,l
(t) = m
1/2

k
F
0
(k)
j+1,Ml+k
(t), (A2)

j,l
(t) = m
1/2

k
H
0
(k)

j+1,Ml+k
(t), (A3)

j,l
(t) = m
1/2

k
F

(k)
j+1,Ml+k
(t), (A4)

j,l
(t) = m
1/2

k
H

(k)

j+1,Ml+k
(t). (A5)
Proof. This follows directly from the dilation and wavelet equations.
Theorem A.6 (Fast Wavelet Transform). Let j Z and f V
j+1
. For
k and = 1, . . . , m1 dene the approximation coecients as

A
j,k
= 'f,

j,k
` = (f (

j,0
)

)(M
j
k)
and the detail coecients as

j,k
= 'f,

j,k
` = (f (

j,0
)

)(M
j
k).
36 ADEL FARIDANI
Then

A
j,k
= m
1/2

l
H
0
(lMk)

A
j+1,l
(A6)
and

j,k
= m
1/2

l
H

(lMk)

A
j+1,l
. (A7)
Proof. Consider the following expansion of f:
f(t) =

A
j+1,l

j+1,l
(t) (A8)
For

A
j,k
, take an inner product of (A8) with

j,k
. Use (A3) and biorthogo-
nality to get (A6).
For

D

j,k
, take an inner product of (A8) with

j,k
. Use (A5) and biorthogo-
nality to get (A7).
Theorem A.7 (Inverse Fast Wavelet Transform). Under the hypothesis
of Theorem A.6

A
j+1,l
= m
1/2

F
0
(l Mk)

A
j,k
+
m1

=1
F

(l Mk)

D

j,k

(A9)
Proof. Consider the following expansions of f:
f(t) =

A
j+1,k

j+1,k
(t)
=

A
j,k

j,k
(t) +
m1

=1

j,k

j,k
(t). (A10)
The second expansion comes from writing f V
j+1
= V
j
W
j
as a sum of
elements of V
j
and W
j
. To get

A
j+1,l
, take an inner product of (A10) with

j+1,l
. Use (A2), (A4) and biorthogonality to get (A9).
An easy way to obtain wavelets in R
n
is to use a tensor product construction with
the wavelets in R. We will look specically at the two-dimensional case. Dene
the lattice as = Z
2
, and the dilation matrix M = 2I, where I denotes the
identity matrix. Since [detM[ = 4 = m, one can expect 3 wavelets and 1 scaling
function. Let the spaces V
j
, W
j
be the chosen one-dimensional multiresolution
analysis with scaling function , and wavelet . The coecients for the dilation
and wavelet equation are F

(k), = 0, 1. Constructing the two-dimensional

scaling function and wavelets by taking products of the one-dimensional functions
leads to the following denition.
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 37
Definition A.8. From a one-dimensional scaling function (x) and its cor-
responding wavelet (x), two-dimensional separable wavelets are dened for
(x, y) R
2
:
(x, y) = (x)(y),

1
(x, y) = (x)(y),
2
(x, y) = (x)(y),
3
(x, y) = (x)(y).
Consider a biorthogonal pair of one-dimensional multiresolution analyses. Recall
for separable, two-dimensional wavelets, M = 2I, m = 4 and = Z
2
. Let V
j
and W
j
be the multiresolution analysis with scaling function , wavelet , and
coecients F
0
, F
1
for the dilation and wavelet equations. Let

V
j
and

W
j
be the
multiresolution analysis with scaling function

wavelet

and coecients H
0
,
H
1
for the dilation and wavelet equations.
We want to rewrite the fast wavelet transform and inverse fast wavelet trans-
form for the case of two-dimensional separable wavelets. Let F

(k), = 0, 1, 2, 3
be the coecients in the dilation and wavelet equations for the separable wave-
lets. It is easy to verify that
F
0
(k) = F
0
(k
1
)F
0
(k
2
), F
1
(k) = F
0
(k
1
)F
1
(k
2
),
F
2
(k) = F
1
(k
1
)F
0
(k
2
), F
3
(k) = F
1
(k
1
)F
1
(k
2
),
where k = (k
1
, k
2
). Similarly for the H

(k), = 0, 1, 2, 3. Thus, equation (A6)

becomes:

A
j,k
= 2

l
H
0
(l 2k)

A
j+1,l
and equation (A7) becomes

j,k
= 2

l
H

(l 2k)

A
j+1,l
for = 1, 2, 3.
The inverse wavelet transform, equation (A9), becomes

A
j+1,n
= m
1/2

F
0
(nMk)

A
j,k
+
3

=1
F

(nMk)

D

j,k

= 2

k
1
,k
2
Z

F
0
(n
1
2k
1
)F
0
(n
2
2k
2
)

A
j,k
+F
0
(n
1
2k
1
)F
1
(n
2
2k
2
)

D
1
j,k
+F
1
(n
1
2k
1
)F
0
(n
2
2k
2
)

D
2
j,k
+F
1
(n
1
2k
1
)F
1
(n
2
2k
2
)

D
3
j,k

.
Recall that the approximation coecients are a convolution of f with

j,0
:

A
j,k
=
(f (

j,0
)

)(M
j
k). We would like to consider, for suciently large j, the
approximation coecients as an approximation for f at a particular point. Notice
that for M = 2I, M
j
x = 2
j
x, and m = 2
n
. From Real Analysis, the following
lemma holds.
38 ADEL FARIDANI
Lemma A.9. Let f be continuous, g L
1
(R
n
) with

g dx = 1, g is bounded
and has compact support. Then for all x R
n
f(x) = lim
j
2
jn

f(x +y)g(2
j
y) dy = lim
j
2
j/2
f (g
j,0
)

(x)
Proof. Let t = 2
j
. Then we have a special case of [20, Theorem 8.15, p. 235].

Thus, for j suciently large, we get an approximation for f:

2
j/2
f (

j,0
)

(2
j
k) f(2
j
k)
or, with n = 2,
2
j/2

A
j,k
f(2
j
k).
Appendix B. Matlab Source Code
We have made available online at http://www.msri.org/publications/books/
Book47/faridani a number of Matlab M-les related to the algorithms described
here. This source code is provided for illustrative purposes only and comes with-
out warranties of any kind.
First there is an implementation of the parallel-beam ltered backprojection
algorithm for the standard lattice. The main le for this is fbp.m, and it is sup-
plemented by three function M-les: Rad.m computes line integrals for a mathe-
matical phantom consisting of ellipses, slkernel.m computes the discrete Shepp
Logan convolution kernel (77), and window3.m allows one to view the recon-
structed image. It is automatically called at the end of the reconstruction. (How-
ever, with the example phantomthe well-known SheppLogan phantom
given in fbp.m the picture shown does not display the most interesting details. It
is better to call window3 again with the parameters window3(0.07,0.07,roi,P).
Also included at the same address are the les fdk.m (a simple implementation
of the FeldkampDavisKress algorithm) and Divray.m (an implementation of
the transform discussed in Problem 9 below).
Appendix C. Some Exercises
Problem 1: Let f(x) be the characteristic function of an ellipse with center
(x
0
, y
0
), half-axes of length a and b, respectively, such that the axis of length 2a
makes an angle with the x-axis when measured counterclockwise starting from
the x-axis. Compute the x-ray transform Pf(, s). Parameterize the unit vector
with its polar angle , i.e., = (cos , sin ).
Problem 2: Compute the Fourier transform of the characteristic function of the
unit disk in R
2
. Hint: Use polar coordinates and formula (3.16) in [62, p. 197].
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 39
Problem 3: Using the Matlab code described in Appendix B for the ltered
backprojection algorithm (with the Shepp Logan phantom), perform the follow-
ing experiments.
(a) Run the program for the following values of p and q, leaving the other
parameters unchanged: q = 16, p = 50; q = 32, p = 100; q = 64, p = 200;
q = 128, p = 400. For each case plot a crossection along the horizontal line of
pixels closest to the line y = 0.605 which passes through the centers of the
three small ellipses. Compare the resolution for the various parameter choices.
(b) Fix the parameter b in the program at the value 64. Theory suggests that
the choice q = 64, p = 200 is a good one. Compare the images for the following
choices of p and q, again leaving the other parameters unchanged. q = 128, p =
200; q = 64, p = 200; q = 32, p = 200; q = 16, p = 200; q = 64, p = 400;
q = 64, p = 100; q = 64, p = 50; q = 64, p = 20; q = 64, p = 10. Summarize
your ndings about the inuence of chosing larger or smaller values of p or q
than the ones suggested by theory.
Problem 4: Modify the ltered backprojection program so that it reconstructs
the function P
#
Pf. Compute an image of P
#
Pf for the SheppLogan phantom
with p = 200, q = 64.
Problem 5: A fundamental question for image reconstruction is if the data
uniquely determine the original image.
(a) Convince yourself that the x-ray transform is a linear operator, i.e., P(f) =
Pf and P(f + g) = Pf + Pg. Show that for linear operators the question of
uniqueness is equivalent to the question if there are nontrivial null-functions.
I.e., Pf = Pg implies f = g, if and only if Pf 0 implies f 0.
(b) While the data Pf(, s) for all s and innitely many directions uniquely
determine the function f, it was already known to the pioneers of tomogra-
phy that this is not the case if Pf(, s) is known for all s but for only nitely
many directions . For example, in his 1963 paper Representation of a function
by its line integrals, with some radiological applications (Journal of Applied
Physics, 34 (1963), 27222727), A. M. Cormack, who later shared the Nobel
prize in medicine for his contributions to tomography, claims that if the func-
tion f(x) vanishes outside the unit disk and inside the unit disk is given by
f(x) = Acos(n), n > 0, where is the polar angle of x, then the line integrals
of f are zero along all lines perpendicular to the directions with polar angle
= (2m+ 1)/(2n), m = 0, . . . , 2n 1. Show that Cormacks claim is correct.
(c) What do you think may be the implications of the existence of such null
functions (or ghosts) for medical imaging?
Problem 6: (a) Modify the ltered backprojection program so that it recon-
structs from fan-beam data. Test it for the SheppLogan phantom with p = 200,
q = 64, MX = MY = 128, R = 2.868 and c = q/ arcsin(1/R).
40 ADEL FARIDANI
(b) Modify the program so that it can read the projection data from a le, using
the fread command. Request from the author the data le pelvis.ctd. It contains
real data from a hospital scanner. The data are stored in integer*2 format and
correspond to a fan-beam geometry with p = 360, q = 256, and R = 2.868.
The angle in the data is incremented in reverse order compared to the lecture.
Reconstruct an image (the best you can get) from these data.
Problem 7: Modify the ltered backprojection program so that it reconstructs
from fan-beam data with detectors on a line as described in [67, pp. 9395]. Test
it for the SheppLogan phantom with p = 200, q = 64, MX = MY = 128,
R = 2.868 and b = q.
Problem 8: Consider a crude method for so-called region-of-interest tomogra-
phy, as follows.
(a) Modify the parallel-beam reconstruction program so that the data outside
the circle inscribed in a square given by the parameter roi are set to zero. Test
your program with the SheppLogan phantom and roi = [0.2,0.2,0.8,0.4].
(Set the parameter circle equal to 1.) Discuss the quality of the resulting image
and compare with the reconstruction from complete data.
(b) Perform the same experiment as in part (a), only do not set the data to
zero outside the region of interest (ROI) but set them to a constant equal to the
nearest line integral intersecting the circle inscribed in the ROI.
Problem 9: (a) Consider the family of functions
f(x) =

1 |x|
2

m
+
, x R
3
, m > 1.
(The case m = 0 gives the characteristic function of the unit ball. The larger m
is, the smoother the function becomes.) Compute the transform
Df(z, ) =

f(z +t) dt, z R

3
, S
2
for these functions. The following formula may be helpful:

1
1
(1 u
2
)
m
du = 2
2m+1
((m+ 1))
2
(2m+ 2)
(Gradshteyn and Ryzhik, p. 949, section 8.380, Formula 9).
(b) Show that for g(x) = f(x x
0
), we have Dg(z, ) = Df(z x
0
, ), and for
h(x) = f(Ax) with A a non-singular matrix,
Dh(z, ) = |A|
1
Df(Az, ), =
A
|A|
.
The M-le Divray.m provided in Appendix B implements this transform for the
functions of part (a).
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 41
Problem 10: Use the source code for the FDK algorithm [18] provided in
Appendix B and familiarize yourself with its use. Input the parameters for a
phantom consisting of one ellipsoid with center at (0.2, 0.3, 0.1) and half-axes
of lengths 0.4, 0.2, 0.3 along the directions (1, 1, 0)/

(2), (1, 1, 0)/

(2) and
(0, 0, 1), respectively. In the code the rows of the orthonormal matrix OV in-
dicate the directions of the principal axes. Produce reconstructions along the
planes y = 0.3, z = 0.1, and x + y + z = 0.6, respectively. Use values p = 20
and q = 64 and indicate for each case the orthonormal vectors n, w1, w2 which
you are using. Observe which boundaries are well reconstructed and which are
blurred.
References
[1] J. Boman and E.T. Quinto, Support theorems for real-analytic Radon transforms
on line complexes in three-space, Trans. Amer. Math. Soc., 335(1993), pp. 877-890.
[2] C. Berenstein and D. Walnut, Local inversion of the Radon transform in even
dimensions using wavelets, in [22, pp. 45-69].
[3] A. M. Cormack, Representation of a function by its line integrals, with some
radiological applications, J. Appl. Phys., 34 (1963), pp. 2722-2727.
[4] K. Buglione, Pseudolocal tomography, M.S. paper, Dept. of Mathematics, Oregon
State University, Corvallis, OR 97331, U.S.A., (1998).
[5] A. M. Cormack, Sampling the Radon transform with beams of nite width, Phys.
Med. Biol., 23 (1978), pp. 1141-1148.
[6] S. R. Deans, The Radon transform and some of its applications, Wiley, 1983.
[7] I. Daubechies, Ten Lectures on Wavelets, SIAM, Capital City Press, VT, 1992
[8] M. Defrise and R. Clack, A cone-beam reconstruction algorithm using shiftvariant
ltering and conebeam backprojection, IEEE Trans. Med. Imag., MI-13 (1994), pp.
186-195.
[9] L. Desbat, Ecient sampling on coarse grids in tomography, Inverse Problems,
9(1993), pp. 251-269.
[10] A. Faridani, An application of a multidimensional sampling theorem to computed
tomography, in [28, pp. 65-80].
[11] A. Faridani, Reconstructing from eciently sampled data in parallel-beam computed
tomography, in G.F. Roach (ed.), Inverse problems and imaging, Pitman Research
Notes in Mathematics Series, Vol. 245, Longman, 1991, pp. 68-102.
[12] A. Faridani, Results, old and new, in computed tomography, in: Inverse Problems
in Wave Propagation, G. Chavent et al. (editors), The IMA Volumes in Mathematics
and its Applications, Vol. 90, Springer Verlag, New York, 1997, pp. 167-193.
[13] A. Faridani, Sampling in parallel-beam tomography, in: Inverse Problems and
Imaging, A.G. Ramm (editor), Plenum, New York, 1998, pp. 33-53.
[14] A. Faridani, E. L. Ritman, and K. T. Smith, Local tomography, SIAM J. Appl.
Math., 52 (1992) , pp. 459-484. Examples of local tomography, SIAM J. Appl. Math.,
52 (1992) , pp. 1193-1198.
42 ADEL FARIDANI
[15] A. Faridani, D.V. Finch, E. L. Ritman, and K. T. Smith, Local tomography II,
SIAM J. Appl. Math., 57 (1997), pp. 1095-1127.
[16] A. Faridani and E. L. Ritman, High-resolution computed tomography from ecient
sampling, Inverse Problems, 16(2000), pp. 635-650.
[17] A. Faridani, K. A. Buglione, P. Huabsomboon, O. D. Iancu, and J. McGrath,
Introduction to Local Tomography, Contemporary Mathematics, Vol. 278 (2001),
pp. 2947.
[18] L. A. Feldkamp, L. C. Davis, and J. W. Kress, Practical conebeam algorithm, J.
Opt. Soc. Am. A, 1 (1984), pp. 612-619.
[19] D. V. Finch, Cone beam reconstruction with sources on a curve, SIAM J. Appl.
Math., 45(1985), pp. 665-673.
[20] G. Folland, Real AnalysisModern Techniques and their Applications John Wiley
and Sons, Inc, (1984)
[21] I.M. Gelfand and S. G. Gindikin (eds.), Mathematical problems of tomography,
Translations of Mathematical Monographs, Vol. 81, Amer. Math. Soc., 1990.
[22] S. Gindikin and P. Michor (eds.), 75 Years of Radon Transform, Conference
Proceedings and Lecture Notes in Mathematical Physics, Vol. 4, International Press,
Boston, 1994.
[23] P. Grangeat, Mathematical framework of cone beam 3D reconstruction via the rst
derivative of the Radon transform, in [34, pp. 66-97].
[24] A. Greenleaf and G. Uhlmann, Nonlocal inversion formulas for the X-ray trans-
form, Duke Math. J., 58(1989), pp. 205-240.
[25] A. Greenleaf and G. Uhlmann, Estimates for singular Radon transforms and
pseudodierential operators with singular symbols, J. Funct. Anal., 89(1990), pp.
202-232.
[26] A. Greenleaf and G. Uhlmann, Composition of some singular Fourier integral op-
erators and estimates for restricted X-ray transforms., Ann. Inst. Fourier, 40(1990),
pp. 443-466.
[27] A. Greenleaf and G. Uhlmann, Microlocal techniques in integral geometry., in [28,
pp. 121-135].
[28] E. Grinberg and E.T. Quinto (eds.), Integral Geometry and Tomography, Contem-
porary Mathematics, Vol. 113, Amer. Math. Soc., Providence, R.I., 1990.
[29] C. W. Groetsch, Inverse problems in the mathematical sciences, Vieweg, Braun-
schweig, 1993.
[30] S. Gutmann, J .H. B. Kemperman, J. A. Reeds, and L. A. Shepp, Existence of
probability measures with given marginals, Annals of Probability, 19 (1991), pp. 1781-
1797.
[31] S. Helgason, The Radon transform, Birkhuser, 1980.
[32] G. T. Herman, Image reconstruction from projections: the fundamentals of com-
puterized tomography, Academic Press, New York, 1980.
[33] G. T. Herman (ed.), Image reconstruction from projections: implementation and
applications, Springer, 1979.
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 43
[34] G.T. Herman, A.K. Louis, and F. Natterer (eds.), Mathematical Methods in
Tomography, Lecture Notes in Mathematics, Vol. 1497, Springer, 1991.
[35] O. D. Iancu, Contour reconstruction in 3D x-ray computed tomography. M.S. paper,
Dept. of Mathematics, Oregon State University, Corvallis, OR 97331, U.S.A., (1999).
[36] A. C. Kak and M. Slaney, Principles of computerized tomographic imaging, IEEE
Press, New York, 1988.
[37] A. I. Katsevich, Local Tomography for the generalized Radon transform, SIAM J.
Appl. Math. 57(1997), pp. 1128-1162.
[38] A. Katsevich, Cone beam local tomography, SIAM J. Appl. Math, 59(1999), pp.
2224-2246.
[39] A.I. Katsevich and A. G. Ramm, A method for nding discontinuities of functions
from the tomographic data, in [76, pp. 115-123].
[40] A.I. Katsevich and A. G. Ramm, New methods for nding jumps of a function
from its local tomographic data, Inverse Problems, 11 (1995), pp. 1005-1023.
[41] A.I. Katsevich and A. G. Ramm, Pseudolocal tomography, SIAM J. Appl. Math.,
56, (1996), pp. 167-191.
[42] F. Keinert, Inversion of k-plane transforms and applications in computer tomog-
raphy, SIAM Review, 31(1989), pp. 273-298.
[43] W. Klaverkamp, Tomographische Bildrekonstruktion mit direkten algebraischen
Verfahren, Ph.D. thesis, Fachbereich Mathematik der Universitt Mnster, Mnster,
Germany, 1991.
[44] H. Kruse, Resolution of reconstruction methods in computerized tomography, SIAM
J. Sci. Stat. Comput.,10(1989), pp. 447-474.
[45] P. Kuchment, K. Lancaster and L. Mogilevskaya, On local tomography, Inverse
Problems, 11 (1995), pp. 571-589.
[46] A. V. Lakshminarayanan, Reconstruction from divergent x-ray data, Suny Tech.
Report 32, Comp. Sci. Dept., State University of New York, Bualo, NY, 1975.
[47] I. Lan, On an operator associated to a restricted x-ray transform, Ph.D. thesis,
Dept. of Mathematics, Oregon State University, Corvallis, OR 97331, U.S.A., (1999).
[48] J. V. Leahy, K. T. Smith, and D. C. Solmon, Uniqueness, nonuniqueness and
inversion in the x-ray and Radon problems, submitted to Proc. Internat. Symp. on
Ill-posed Problems, Univ. Delaware Newark, 1979. The proceedings did not appear.
Some results of this article have been published in [93].
[49] R. M. Lewitt, Reconstruction algorithms: transform methods, Proc. IEEE, 71
(1983), pp. 390-408.
[50] R. M. Lewitt, R. H. T. Bates, and T. M. Peters, Image reconstruction from
projections: II: Modied back-projection methods, Optik, 50 (1978), pp. 85-109.
[51] B. F. Logan, The uncertainty principle in reconstructing functions from projec-
tions, Duke Math. J., 42 (1975), pp. 661-706.
[52] A. K. Louis, Nonuniqueness in inverse Radon problems: the frequency distribution
of the ghosts, Math. Z. 185 (1984), pp. 429-440.
[53] A. K. Louis, Orthogonal function series expansions and the null space of the Radon
transform, SIAM J. Math. Anal., 15(1984), pp. 621-633.
44 ADEL FARIDANI
[54] A. K. Louis, Incomplete data problems in x-ray computerized tomography I.
Singular value decomposition of the limited angle transform, Numer. Math. 48(1986),
pp. 251-262.
[55] A. K. Louis, Inverse und schlecht gestellte Probleme, B.G. Teubner, Stuttgart,
1989.
[56] A. K. Louis, Medical imaging: state of the art and future development, Inverse
Problems 8(1992), pp. 709-738.
[57] A. K. Louis, and P. Maass, Contour reconstruction in 3-D x-ray CT, IEEE Trans.
Med. Imag., MI-12 (1993), pp. 764769.
[58] A. K. Louis and F. Natterer, Mathematical problems in computerized tomography,
Proc. IEEE, 71 (1983), pp. 379-389.
[59] W. R. Madych, Summability and approximate reconstruction from Radon trans-
form data, in [28, pp. 189-219].
[60] P. Maass, The x-ray transform: singular value decomposition and resolution,
Inverse Problems, 3 (1987), pp. 729-741.
[61] P. Maass, The interior Radon transform, SIAM J. Appl. Math. 52(1992), pp. 710-
724.
[62] F. Natterer, The Mathematics of Computerized Tomography, Wiley, 1986.
[63] F. Natterer, Sampling in fan-beam tomography, SIAM J. Appl. Math. 53(1993),
pp. 358-380.
[64] F. Natterer, Recent developments in x-ray tomography, in [76, pp. 177-198].
[65] F. Natterer, Algorithms in Tomography, Preprint, (1997). Available online at
wwwmath.uni-muenster.de/math/inst/num
[66] F. Natterer, Numerical methods in tomography, in Acta Numerica, Vol. 8, 1999,
Cambridge University Press, New York, pp. 107-141.
[67] F. Natterer and F. Wuebbeling, Mathematical Methods in Image Reconstruction,
SIAM, Philadelphia, 2001.
[68] T. Olson and J. de Stefano, Wavelet localization of the Radon transform, IEEE
Trans. Sig. Proc., 42 (1994), pp. 2055-2067 .
[69] D. P. Petersen and D. Middleton, Sampling and reconstruction of wave-number-
limited functions in N-dimensional euclidean space, Inf. Control, 5(1962), pp. 279-
323.
[70] R. M. Perry, On reconstructing a function on the exterior of a disc from its Radon
transform, J. Math. Anal. Appl. 59 (1977), pp. 324-341.
[71] D. A. Popov, On convergence of a class of algorithms for the inversion of the
numerical Radon transform, in [21, pp. 7-65].
[72] E.T. Quinto, The dependence of the generalized Radon transform on dening
measures, Trans. Amer. Math. Soc. 257 (1980), pp. 331-346.
[73] E. T. Quinto, Singular value decompositions and inversion methods for the exterior
Radon transform and a spherical transform, J. Math. Anal. Appl., 95 (1983), pp.437-
448.
[74] E.T. Quinto, Tomographic reconstructions from incomplete data numerical
inversion of the exterior Radon transform, Inverse Problems, 4 (1988), pp. 867-876.
INTRODUCTION TO THE MATHEMATICS OF COMPUTED TOMOGRAPHY 45
[75] E. T. Quinto, Singularities of the x-ray transform and limited data tomography in
R
2
and R
3
, SIAM J. Math. Anal., 24 (1993), pp. 1215-1225.
[76] E.T. Quinto, M. Cheney, and P. Kuchment (eds.), Tomography, Impedance Imag-
ing, and Integral Geometry, Lectures in Applied Mathematics, Vol. 30, Amer. Math.
Soc., 1994.
[77] E.T. Quinto, L. Ehrenpreis, A. Faridani, F. Gonzalez and E. Grimberg (eds.),
Radon Transforms and Tomography, Contemporary Mathematics, Vol. 278, Ameri-
can Mathematical Society, Providence, RI, 2000.
[78] A. G. Ramm and A. I. Katsevich, The Radon Transform and Local Tomography,
CRC Press, Boca Raton, 1996.
[79] A. G. Ramm and A.I. Zaslavsky, Reconstructing singularities of a function given
its Radon transform, Math. and Comput. Modelling, 18(1993), pp. 109-138.
[80] F. Rashid-Farrokhi, K. J. R. Liu, C. A. Berenstein, and D. Walnut, Wavelet-based
multiresolution local tomography, IEEE Transactions on Image Processing, 6 (1997),
pp. 1412-1430.
[81] P. A. Rattey and A. G. Lindgren, Sampling the 2-D Radon transform, IEEE Trans.
Acoust. Speech Signal Processing, ASSP-29(1981), pp. 994-1002.
[82] M.G. Raymer, M. Beck, and D.F. McAllister, Complex wave-eld reconstruction
using phase-space tomography, Phys. Rev. Lett., 72(1994), pp. 1137-1140.
[83] E. L. Ritman, J. H. Dunsmuir, A. Faridani, D. V. Finch, K. T. Smith, and P. J.
Thomas, Local reconstruction applied to microtomography, in: Inverse Problems in
Wave Propagation, G. Chavent et al. (editors), The IMA Volumes in Mathematics
and its Applications, Vol. 90, Springer Verlag, New York, 1997, pp. 443-452.
[84] L. A. Shepp and J. B. Kruskal, Computerized tomography: the new medical x-ray
technology, Amer. Math. Monthly, 85 (1978), pp. 420-439.
[85] E. A. Sivers, D. L. Halloway, W. A. Ellingson, and J. Ling, Development and
application of local 3-D CT reconstruction software for imaging critical regions in
large ceramic turbine rotors, in Rev. Prog. Quant. Nondest. Eval.:, D.O. Thompson
and D.E. Chimenti (eds.), Plenum, New York, 1993, pp. 357-364.
[86] E. A. Sivers, D. L. Halloway, W. A. Ellingson, Obtaining high-resolution images
of ceramics from 3-D x-ray microtomography by region-of-interest reconstruction,
Ceramic Eng. Sci. Proc., 14, no. 7-8, (1993), pp. 463-472.
[87] J. Skaggs, Region of interest tomography using biorthogonal wavelets, M.S. paper,
Dept. of Mathematics, Oregon State University, Corvallis, OR 97331, U.S.A., (1997).
[88] K. T. Smith, Reconstruction formulas in computed tomography, Proc. Sympos.
Appl. Math., No. 27, L. A. Shepp, ed., AMS, Providence, RI, (1983), pp. 7-23.
[89] K. T. Smith, D.C. Solmon, and S. L. Wagner, Practical and mathematical aspects
of the problem of reconstructing objects from radiographs, Bull. Amer. Math. Soc.,
83(1977), pp. 1227-1270. Addendum in Bull. Amer. Math. Soc., 84(1978), p. 691.
[90] K. T. Smith and F. Keinert, Mathematical foundations of computed tomography,
Appl. Optics 24 (1985), pp. 3950-3957.
[91] D. T. Smithey, M. Beck, M. G. Raymer, and A. Faridani, Measurement of the
Wigner distribution and the density matrix of a light mode using optical homodyne
46 ADEL FARIDANI
tomography: application to squeezed states and the vacuum, Phys. Rev. Lett., 70
(1993), pp. 1244-1247.
[92] D. C. Solmon, The x-ray transform, J. Math. Anal. Appl., 56(1976), pp. 61-83.
[93] D. C. Solmon, Nonuniqueness and the null space of the divergent beam x-ray
transform, in [28, pp. 243-249].
[94] W. J. T. Spyra, A. Faridani, E. L. Ritman, and K. T. Smith, Computed tomographic
imaging of the coronary arterial tree - use of local tomography, IEEE Trans. Med.
Imag., 9 (1990), pp. 1-4.
[95] G. Strang, T. Nguyen, Wavelets and Filter Banks, WellesleyCambridge Press,
Wellesley, MA, (1996).
[96] R. Strichartz, Construction of Orthonormal Wavelets, in Wavelets: Mathematics
and Applications, pg. 23-50, Editors: L. Benedetto, M. Frazier, CRC Press, Inc,
Baton Rouge, FL, (1994).
[97] H. K. Tuy, An inversion formula for cone beam reconstruction. SIAM J. Appl.
Math. 43(1983), pp. 546-552.
[98] . I. Vainberg, I. A. Kazak, and V. P. Kurozaev, Reconstruction of the internal
three-dimensional structure of objects based on real-time internal projections, Soviet
J. Nondestructive Testing, 17 (1981), pp. 415-423.
[99] . I. Vainberg, I. A. Kazak, and M. L. Faingoiz, X-ray computerized back projection
tomography with ltration by double dierentiation. Procedure and information
features, Soviet J. Nondestructive Testing, 21 (1985), pp. 106-113.
[100] R. C. Vaughn, The parallel-beam ltered backprojection algorithm with reconstruc-
tion on a polar grid, M.S. paper, Dept. of Mathematics, Oregon State University,
Corvallis, OR 97331, USA, 1994.
Adel Faridani
Departmentof Mathematics
Oregon State University
Corvallis, OR 97331
United States
[email protected]
http://oregonstate.edu/faridana