http://www.paper.edu.

cn

Image Reconstruction for Fan Beam X-Ray

Tomography Using General Hankel
Transform Pair (2)

Zhao Shuang-Ren
Doubletask, Toronto, Canada

Abstract: A new transform pair has been introduced for the

fan-beam image reconstruction[10]. This new integral transform
plays the same role as the Hankel ( Fourier Bessel) transform in the
parallel beam case and includes the Hankel transform as a special
case. In this paper we develop this method for the equal-angle fan
beam geometry. We call this new pair the General Hankel transform
which gives transform and inverse transform from an object
θ) to fan beam projections Rβ(γγ). One member of the
function f(r,θ)
pair can be used to calculate projection data for simulation; the
other can be used as an algorithm for fan beam reconstruction
without interpolation in polar coordinates. Compared with other
algorithms for fan beam reconstruction the general Hankel
θ), the image to be
transform method is especially useful when f(r,θ),
reconstructed, has only low frequencies in the θ angle direction and
only few projections are available.
 http://www.paper.edu.cn

1. Introduction

The problem of reconstructing an object from a set of its

projections arises in among other fields, computer-aided tomography
CAT, radio astronomy, electron microscopy, and spotlight-mode
synthetic aperture radar. Traditional algorithms of reconstruction for
parallel or fan beam projections is the convolution backprojection
method[1-6], the Fourier method[7], the Hankel transform method [8].
The convolution-backprojection method is fast and widely used. The
Fourier method is faster than the convolution-backprojection method, but
the reconstruction quality is poor because interpolation in the Fourier
domain introduces additional artifacts[9]. For a special case where the
object to be reconstructed has no high frequencies along the θ angle, the
Hankel transform method can be used and good quality of reconstruction
can be achieved by few projections. However, the Hankel method is only
used for parallel beam projections. Ref.[10] has introduced a new integral
transform pair which extended the Hankel method for the fan beam
equal-space case. This new integral transform pair called General
Hankel transform pair plays exactly the same role in the fan-beam case as
the Hankel transform pair in the parallel-beam case. In this paper we
developed this method for the fan beam equal-angle geometry.

2. Two integral formulas and General Hankel

Transform pair for fan beam tomography

The spirit of the Hankel transform method for parallel tomography

is first to obtain the integral transform from the object function f(r,φ) to
the projection Pθ(u) and the inverse transform from the projection Pθ(u)
to the object function f(r,φ), then expands these into Fourier series. We

2
 http://www.paper.edu.cn

follow this spirit to obtain the general Hankel Transform pair which can
be applied to fan beam geometry.
From Ref. [11] the method of back projection for the case of equi-
angle detectors is available
2π 1 γm
⌡ L2 ⌠
f(r,φ)= ⌠ ⌡cos(γ)Rβ(γ)g(γ'-γ) Ddγdβ (1)
0 -γm
where
1 γ 2
g(γ)=2 ( ) h(γ) (2)
sin(γ)
∞
1 1
h(γ)= 2 ⌠
⌡|ξ|exp(jξγ)dξ
D 2π
-∞
∞
1
= ⌠|η|exp(jηDγ)dη (3)
2π ⌡
-∞
2
L= 1+ρ +2 ρ sin(β−φ) (4)
ρ cos(β−φ)
γ'=tan-1( ) (5)
1+ρ sin(β−φ)
and f(r,φ) is the image function to be reconstructed in polar coordinates
(r,φ). Rβ(γ) denotes a fan beam projection as shown in Fig.1 where γ is
the fan beam span angle corresponding to the detector bank and β is the
rotation angle; R'β(γ) is a modified fan beam projection, D is the distance
r
from the source point Φ to the origin O. ρ =D , If we define

Rβ(γ)=0 if γ is not in [-γm, γm] (6)

the integral limit in Eq.(1) [-γm, γm] can be extended to (-∞,∞).

3
 http://www.paper.edu.cn

γ object fuction.
t
D β
u
(r,φ )
φ θ
O

Fig 1. Fan beam geometry.

Define a modified fan beam projection

Wβ(t)=cos(γ) Rβ(γ)|γ=t/D (7-a)
so
1
Rβ(γ)|γ= W (t)|
cos(γ) β t=Dγ (7-b)

where t is the arc which has the radius of D

Considering the Fourier transform
∞ ∞
1
η t= ⌠⌡exp(-jηt)dt; t 1η= ⌠
2π ⌡
exp(jηt)dη (8)
-∞ -∞
Define the one dimensional Fourier transform of Wβ(t) as
Gβ(η)=η t(Wβ(t))
∞
=⌡⌠Wβ(t)exp(-jηt)dt
-∞
∞
=⌠⌡cos(γ) Rβ(γ)exp(-jηDγ)Ddγ (9)
-∞
So
Wβ(t)
=cos(γ) Rβ(γ)

4
 http://www.paper.edu.cn

∞
1
= ⌠
2π ⌡
G(η)exp(jηDγ)dη (10)
-∞
where t=Dγ. We also have
t
g^(η)=η t(g(D))
∞
⌠ t
=  g(D)exp(-jηt)dt
⌡
-∞
∞
=⌠⌡ g(γ) exp(-jηDγ) D dγ (11)
-∞
and
g(γ) =t η(g^(η))|t=D γ
∞
1
=2π ⌡⌠g^(η) exp(jηt)dη
-∞
∞
1
= ⌠ (12)
2π ⌡
g^(η) exp(jηDγ)dη
-∞
Combining (1) - (12) and assuming that all functions are
integralable, so that the order of integration can be changed, we have
2π

⌠ γm
1 ⌠
∞

f(r,φ)= L2 cos(γ)Rβ(γ) ⌠
1
2π ⌡
g^(η)exp(jηD(γ'-γ))dηDdγdβ
 ⌡ -∞
⌡ -γm
0

5
 http://www.paper.edu.cn

2π

⌠ ⌠∞ γm
= L2 g^(η) ⌠
1 1
⌡cos(γ)Rβ(γ)exp(-jηDγ)Ddγexp(jηDγ')dηdβ
2π
 ⌡ -γm
⌡ -∞
0

2π
∞
1 ⌠1
= L2 ⌠⌡g^(η)Gβ(η)exp(jηDγ')dηdβ (13)
2π ⌡
0 -∞
Eq.(7), (9) and (13) can be used to calculate f(r,φ) from Rβ(γ). In
order to obtain the general Hankel transform pair, we need another
equation by which Rβ(γ) can be calculated from f(r,φ).
The parallel beam projection Pθ(u) is[12]
2π a
Pθ(u)= ⌠
⌡ ⌠
⌡ f(r,φ)δ(u-r cos(θ−φ))rdrdφ (14)
φ=0 r=0
where u is the distance from the origin to the line of x-ray, see Fig 1, a is
limited such that if r>a, f(r,φ)=0.
Considering the fan beam geometry Fig. 1
u=D sinγ
θ=β+γ (15)
as in ref. [10] we have
1
δ(u-r cos(θ−φ))= δ(γ-γ') (16)
cos(γ)(D+r sin(β−φ))
where γ' are defined in equation (5). Considering Ref.[11] and Eq.(14) -
(16) the fan beam projection Rβ(γ) is
Rβ(γ)=Pθ(u)|u=D sin(γ);θ=β+γ
2π a 1
= ⌠
⌡ ⌠ ⌡ f(r,φ) cos(γ)[D+r sin(β−φ)]δ(γ-γ')rdrdφ (17)
φ=0 r=0

6
 http://www.paper.edu.cn

Using equations (17) and assuming that all function are

integralable, so that the order of integration can be changed , equation (9)
becomes
∞
Gβ(η)= ⌠
⌡cos(γ) Rβ(γ)exp(-jηDγ)Ddγ=
-∞

∞
⌠ 2π a 1
cos(γ) ⌡⌠ ⌠⌡ f(r,φ)cos(γ)[D+r sin(β−φ)]δ(γ-γ') rdrdφexp[−jηDγ]Ddγ
⌡ φ=0 r=0
-∞
2π a D ∞
= ⌠
⌡ ⌠ ⌡ [D+r sin(β−φ)]f(r,φ) ⌠
⌡ δ(γ-γ') exp[−jηDγ]dγ rdrdφ
φ=0 r=0 -∞
2π
⌠ a1
=  Uf(r,φ)exp[-jηDγ']rdrdφ
⌠
(18)
 ⌡
⌡ r=0
φ=0
where
U=1+ρ sin(β−φ) (19)
Equation (13) and (18) are two important integral formulas. (13)
can be used for image reconstruction. (18) can be used to calculate the
projection data from the image function. These data are needed for
simulation. These two integral formulas can be expanded using Fourier
series,
∞
f(r,φ)= ∑fm(r)exp(j m φ) (20−a)
m=-∞

2π
1
fm(r,φ)= ⌠ f(r,φ)exp(-j m φ)dφ
2π ⌡
(20-b)
0

7
 http://www.paper.edu.cn

and
∞
Wβ(Dγ)= ∑Wm(γ)exp(j m β) (21)
m=-∞
2π
1
Wm(γ)= ⌡ ⌠Wβ(Dγ)exp(-j m β)dβ (22)
2π
0
∞
Gβ(η)= ∑Gm(η)exp(j m β) (23)
m=-∞
combining definition (9) and equations (21) ,(22) and (23) gives
∞
Gm(η)= ⌠ ⌡Wm(γ)exp(-j Dγ η)D dγ (24)
-∞
These leads to the new integral transformation pair
∞
π
fm(r)=exp(j m2) ⌡ ⌠Gm(η)Hm(r,η)g^(η)dη (25-a)
η=−∞
where
2π
1 ⌠1
Hm(r,η)=  2exp[j(m ε+D γ' η)] dε (25−b)
2π ⌡L
ε=0
and
a
π
Gm(η)=2π exp(-j m )⌠
2⌡m m
f (r)I (r,η)rdr (26-a)
0
where
2π
1 ⌠1
Im(r,η)= Uexp[-j(m ε+D γ' η)]dε (26−b)
2π ⌡
0
In these two formulas the substitution β−φ=ε+π/2 has been used and
U=1+ρcos(ε)
2
L= 1+ρ +2 ρ cos(ε)

8
 http://www.paper.edu.cn

-ρ sin(ε)
γ'=tan-1( ) (27)
1+ρ cos(ε)
Equations
(7)--(22)--(24)--(25)--(20-a)
can be used as an algorithm for the fan beam image reconstruction.
Equations
(20-b)--(26)--(23)--(10)--(7-b)
can be used to produce fan beam simulation data Rβ(γ) from a known
object function f(r,φ).

3. The connection between the general Hankel

transform and the Hankel transform

In this section the connection between the new integral transform

and the Hankel transform is derived.
One of important algorithms for the parallel-beam image
reconstruction is the Hankel transform method [8]
∞
π
fm(r)= exp(j m ) ⌠ F (η)Jm(r η)ηdη
2 ⌡ m
(28)
0
a
π
Fm(η)=2π exp(-j m )⌠ f (r)J (r η)rdr
2⌡m m
(29)
0
where Fm(η) is defined by
Fθ(η)=η u{Pθ(u)} (30)
∞
Fθ(η)= ∑Fm(η)exp(j m θ) (31)
m=-∞
Jm is the first class Bessel function
2π
1
Jm(r η)= ⌡ ⌠exp[+j(m ε−r η sin(ε))]dε (32)
2π
ε=0

9
 http://www.paper.edu.cn

r
we note that if ρ=D −−−>0 then β−−−>θ,

Hm(r,η) −−−> Jm(r η)

Im(r,η) −−−> Jm(r η)
Eq.(25)−−−>Eq.(28)
Eq.(26)−−−>Eq.(29)
Gm(η)−−−>Fm(η) (33)
Τhis means that the Hankel transform is a special case ( where ρ −−−>0)
of the general Hankel transform.

4. Implementation of this algorithm

1) The implementation of (7) calculates the weighted

projections:
W(βi , tk)=cos(γ) R(βi , γ)|γ=tk/D (34)
where R(βi,γ) --data of fan-beam projections, βi=i
∆β, ∆β=2π/(Νp+1), (Np+1) --number of projections, i in [0,Np]; tk=k∆t,
∆t=(D1/(D1+D2))∆v, D1--distance from the source Φ to the centre O,
D2-- distance from the centre to detectors, ∆v detector distance, k in [-
Ns2,Ns2], Ns2= Ns div 2, div --integer division, Ns--detector number.
See fig 2.

10
 http://www.paper.edu.cn

∆v

t
∆t
detectors
umax
Ο
Φ
D1 D2
Object

Fig. 2. Object, projections and detector distance.

2) The implementation of (15-b) calculates the components of

the Fourier series. First we define
N
MODm-->n (35)
N
IMODm-->n (36)
as operators, which change the integer value m to n according to
n=m mod N (37)
and
m if m in [0,N2-1]
n={ (38)
m-N if m in [N2,N-1]
respectively, where N2=N div 2. Define the FFT and IFFT operators for
any series Xm;
N-1
N
FFTm-->n{Xm}=
∑ 2π
Xmexp[-j m n ]
Ν
(39)

m=0

11
 http://www.paper.edu.cn

N-1
N
IFFTm-->n{Xm}=
∑ 2π
Xmexp[j m n ]
Ν
(40)

m=0
Since Eq.(22). We have
1 Np+1 Np+1
Wm(tk)=N +1IMODn-->m FFTi-->n {W(βi;tk)} (41)
p
where i,n in[0,Np] m in [-Np2,Np2-1], Np2=(Np+1) div 2.

3) The implementation of (24) calculates the Fourier

components Gm.
Nv+1 Nv+1 Nv+1
Gm(ηl)=∆t IMODi-->l FFTj-->i {MODk-->j W'm(tk)} (42)

where
W'm(tk)=Wm(tk) if k in [-Ns2,Ns2]
0 if k in[-Nv2,Nv2-1]-[Ns2,Ns2]; (43)
and ηl=l ∆η,∆η=2π/tT; tT=(Nv+1)∆t, l, k in [-Nv2,Nv2-1], Nv2=Nv div
2, (Nv+1) is size of FFT algorithm; j, i in [0,Nv], and Nv>Ns .

4) The implementation of (25) calculates the elements of the

Matrix Hm.
I-1
1
Hm(rn,ηl)= I
∑ 1
L2 cos[m εi -tan -1(ρn sin(εi)) ηl] (44)
U
i=0
where U=1+ρncos(εi), L= 1+ρn2+ρncos(εi) ρn=rn/D1, rn=n ∆r,
∆r=umax/Nf2, Nf2 = Nf div 2; Nf--this number can be chosen , for
example, Nf=Ns; εi=i∆ε, ∆ε=2π/Ι, Ι−−defined by a large integer,
umax=D1sin(γmax), γmax= (Ns2+0.5)∆γ. ∆γ=∆t/D1=∆v/(D1+D2), Note
only m in [0,Np2-1] needs to be calculated.

5) The implementation of (25-a) calculates fm(rn).

12
 http://www.paper.edu.cn

(Nv2-1)
π
fm(rn)=exp(j m )
2 ∑Hm(rn,ηl)g^(ηl)Gm(ηl)∆η (45)
l=-(Nv2-1)
where
1 tk/D
g^(ηl)=2 sin(t h(t ) (46)
k/D) k
where
∞
1
h(tk)= ⌡ ⌠|η|exp(jηtk)dη
2π
-∞
1
= n=0
4 ∆t2
0 n=even
1
- n=odd (47)
(nπ∆t)2
in which the dc value of the object function has been considered
according to the Eq.(3.61) of Ref[11].

6)The implementation of (14) reconstructs the object

function.
Np2-1
f(xi,yj)=f0(r)+2 Re{ ∑fm(r)exp[j m φ]} (48)
m=0
where fm(r) is a linear interpolation of fm(rn).
fm(rn) fm(r1)
r0 + r1 fm(rn)r1+fm(rn+1) r0
fm(r)= = (49)
1 1 ∆r
+
r0 r1

rn=trunc(r/∆r)∆r ; rn+1=rn+∆r
r0=r-rn ; r1=rn+1-r (50)
trunc takes the integer value and
r= xi2+yj2
φ=tan-1(yj/xi) (51)

13
 http://www.paper.edu.cn

xi=i ∆x , yj=j ∆y , ∆x=∆y=umax/(Nv2+0.5); i,j in [-Nv2, Nv2-1].

It is remarkable that if the matrix Hm(rn,ηl) has been
precalculated, most of the computation time is spent in Step 6. In order to
reduce the computation time, the implementation of Step 6. can be
changed as following

6'. Alternative of 6.
f1 f2 f3 f4
+
r 1 +r 2 +r 3 r 4
f(xi,yj)= 1 1 1 1 (52)
+
r 1 +r 2 +r 3 r 4
where
f1=f(φk;rn); f2=f(φk+1;rn)
f3=f(φk;rn+1); f4=f(φk+1;rn+1) (53)

Np+1 Np+1 Np+1

f(φk,rn)=IMODk'-->k FFTm'-->k' {MODm-->m'fm(rn)} (54)

and
r1= r2(φ-φk)2+(r-rn)2; r2= r2(φk+1-φ)2+(r-rn)2
r3= r2(φ-φk)2+(rn+1-r)2; r4= r2(φk+1-φ)2+(rn+1-r)2 (55)
and
rn=trunc(r/∆r) ∆r ; rn+1=rn+∆r
φk=trunc(φ/∆φ) ∆φ ; φk+1=φk+∆φ (56)
and r,φ is defined in (51).
If only few projections are available, Step 6. is the best choice and
the quality of the reconstruction is better than with the convolution
backprojection method. If more projections are available, Step 6'. is the
best choice and the same quality of reconstruction as with the
convolution backprojection method is achieved with only a quarter of

14
 http://www.paper.edu.cn

computation time or even less (supposed that the matrix Hm(rn,ηl) has
been precalculated ).

5. Conclusion

We derived an new integral transform pair. We explored an

algorithm which modified the Hankel algorithm used for the
reconstruction from the parallel-beam projections to suit the case of
reconstruction from the fan beam projections.

Appendix

1. Derivation of Eq.(15). First we show

1
δ[Ψ(ξ)(Φ(ξ)−α)]= δ[ξ−Φ−1(α)] 1−1
Ψ(Φ (α))Φ'(Φ−1(α))
−1
if
I Φ(ξ)−α=0 has only one root ξ=Φ−1(α)
II Ψ(Φ−1(α))≠0
III Φ'(Φ−1(α))≠0 1−2
Let
τ=Ψ(ξ)(Φ(ξ)−α) 1−3
dτ=[Ψ'(ξ)(Φ(ξ)−α)+Ψ(ξ)Φ'(ξ)]dξ 1−4
so

⌠f(ξ)δ[Ψ(ξ)(Φ(ξ)−α)]dξ
⌡

⌠ 1
=f(ξ) δ(τ)dτ 1−5
⌡ Ψ'(ξ)(Φ(ξ)−α)+Ψ(ξ)Φ'(ξ)

because we have ξ=Φ−1(α) when τ=0, so above equation is

15
 http://www.paper.edu.cn

1
f(Φ−1(α))
Ψ(Φ (α))Φ'(Φ−1(α))
−1

⌠ 1
=f(ξ) δ(ξ−Φ−1(α))dξ 1−6
⌡ Ψ(Φ (α))Φ'(Φ−1(α))
−1

Comparing Eq.1-5 Eq.1-6 we obtain Eq.1-1 ,1-2. Second we cosider

u-r cos(θ−φ)
= sin(γ)-r cos((β+γ)-φ)
=D sin(γ)-r cos((β-φ)+γ)
=D sin(γ)-r cos(β−φ)cos(γ)+r sin(β−φ)sin(γ)
=(D+r sin(β−φ)) sin(γ)-r cos(β−φ)cos(γ)
r cos(β−φ)
=(D+r sin(β−φ))cos(γ)[tan(γ)- ] 1-7
D+r sin(β−φ)
so
δ(u-r cos(θ−φ))
r cos(β−φ)
=δ((D+r sin(β−φ))cos(γ)[tan(γ)- ]) 1-8
D+r sin(β−φ)
From Eq.(1-1) and (1-2), we obtain
1
δ(u-r cos(θ−φ))= δ(γ-γ') 1-9
cos(γ)(D+r sin(β−φ))
where
r cos(β−φ)
γ'=tan-1( ) 1-10
D+sin(β−φ)
These are Eq.(15)

16
 http://www.paper.edu.cn

References
[1]. G.N. Ramachandran and A.V.Lakshmin-arayanan, "Three-
dimensional reconstruction from radiographs and electron micrograph:
application of convolutions instead of Fourier transform," Proc. Not
Acad. Sci. USA, Vol.68, PP.2236-2234, Sept. 1971.
[2]. L.A. Shepp and B.F. Logan," The Fourier reconstruction of a
head section ," IEEE trans, Nucl, Sci., Vol. NS-21, PP 21-43, June 1974.
[3]. R.M.lewitt, R.H.T.Bates and T.M.Peters, " Image
reconstruction from projections, II Modified back-projection methods,"
Optik Vol, 50, PP.85-109, Mar. 1978.
[4]. G.T. Herman, Ed., "Image Reconstruction from Projections:
The Fundamentals of Computer Tomography, New York:Academic,
1980.
[5]. R.N. Mersereau and A.V. Oppenheim, "Digital reconstruction
of multidimensional signals from their projections." Proc. IEEE, Vol 62,
PP 1319 1338, Oct. 1974.
[6]. H.J. Scudder," Introduction to computer aided tomgraphy,"
Proc IEEE, Vol,66, PP 628-637, June 1978.
[7]. Hui Peng, Hener Stark, "Direct Fourier Reconstruction in fan
beam tomography," IEEE trans. on Medical Imaging, Vol. MI-6 No.3
September 1987.
[8]. William. E.Higgins,DAvid C.Munson, "A Hankel Transform
Approach to Tomographic Image Reconstruction," IEEE Trans. On
Medical Imaging. Vol. 7, No.1,March 1988.
[9] F.Natterer."The mathematics of computerized Tomography."
Cpyright 1986 by John Wiley & Sons Ltd and BG Teubner, Stuttgart. pp
119.

17
 http://www.paper.edu.cn

[10]. Zhao S.R., H. Halling, " Image Reconstruction for Fan Beam
X-ray Tomography Using a New Integral Transform pair" February
1992.
[11]. Avinash C.Kak and Malcom Slaney, " Computerized
tomographic Imaging" Printed in the United States of America. IEEE
order Number: Pco2071, PP.90.
[12]. Paul A.Rattey, and Allen G.Lindgren, "Sampling the 2-D
Random Transform IEEE Transactions on Acoustics Speech, and Signal
Processing. Vol Ascp-29, No.5 October, 1981.
[13]. Zhao. Shuang-Ren, H.Halling G.Fuhrmann Sampling the Fan
beam Projection Using fan function. " KFA internal publication.
[14]. G.Fuhrmann "Testphantom-Projektionsdatensaetze fuer die
2D-Computer-Tomographie", KFA internal publication.
[15]. P.A. Rattey," Tomographic Image Reconstruction from a
Limited Number of Projections." Signal Processing 7 (1984) 371-387
North-Holland.

18