Cone-beam and fan-beam image reconstruction algorithms based on spherical and circular harmonics

INSTITUTE OF PHYSICS PUBLISHING Phys. Med. Biol. 49 (24) 2239 2256 PHYSICS IN MEDICINE AND BIOLOGY PII: S31-9155(4)7992-6 Cone-beam and fan-beam image reconstruction algorithms based on spherical and circular harmonics Gengsheng L Zeng 1 and Grant T Gullberg 2 1 Utah Center for Advanced Imaging Research, University of Utah, 729 Arapeen Drive, Salt Lake City, Utah 8418, USA 2 E O Lawrence Berkeley National Laboratory, One Cyclotron Road, Mail Stop 55R121, Berkeley, CA 9472, USA E-mail: larry@ucair.med.utah.edu and gtgullberg@lbl.gov Received 27 October 23 Published 19 May 24 Online at stacks.iop.org/pmb/49/2239 DOI: 1.188/31-9155/49/11/9 Abstract A cone-beam image reconstruction algorithm using spherical harmonic expansions is proposed. The reconstruction algorithm is in the form of a summation of inner products of two discrete arrays of spherical harmonic expansion coefficients at each cone-beam point of acquisition. This form is different from the common filtered backprojection algorithm and the direct Fourier reconstruction algorithm. There is no re-sampling of the data, and spherical harmonic expansions are used instead of Fourier expansions. As a special case, a new fan-beam image reconstruction algorithm is also derived in terms of a circular harmonic expansion. Computer simulation results for both cone-beam and fan-beam algorithms are presented for circular planar orbit acquisitions. The algorithms give accurate reconstructions; however, the implementation of the cone-beam reconstruction algorithm is computationally intensive. A relatively efficient algorithm is proposed for reconstructing the central slice of the image when a circular scanning orbit is used. 1. Introduction Over the last ten years there have been remarkable advancements in cone-beam image reconstruction algorithms. Concurrent to the algorithm developments there have been significant advancements in x-ray CT helical detector hardware. This technology has been, and remains, a fundamental component of the medical industry. Continuing advancements in conebeam reconstruction theory are important for the development of better and faster algorithms. The goal of this paper is to develop a cone-beam data-interpolation-free algorithm, which may have a potential to outperform the current cone-beam image reconstruction algorithms in terms of image resolution. 31-9155/4/112239+18$3. 24 IOP Publishing Ltd Printed in the UK 2239

224 G L Zeng and G T Gullberg This paper presents a cone-beam image reconstruction algorithm that utilizes spherical harmonic expansions. Cormack and others first applied a circular harmonic expansion to the reconstruction of 2D images from parallel projections (Cormack 1963, 1964, Marr 1974, Hansen 1981). Hawkins et al (1988) successfully used circular harmonic expansions to develop an algorithm for the reconstruction of exponential Radon projection data and applied it to single photon emission computed tomography (SPECT). Later, 2D image reconstruction of data acquired using converging geometries was performed using harmonic expansions (You et al 1998). Our paper is the first to propose the use of spherical harmonic expansions to reconstruct 3D images from cone-beam projection data. In previous work (Cormack 1963, 1964, Hawkins et al 1988, You et al 1998) the expansion of the projection data were first found; then a relationship between the expansion coefficients of the data and the expansion coefficients of the image was derived. Using this relationship, the expansion coefficients of the image are determined and the final image is reconstructed by synthesizing these expansion coefficients. Only for the parallel geometry (Cormack 1963, 1964, Marr 1974, Hansen 1981) does an orthogonal polynomial expansion in the image space correspond to an orthogonal polynomial expansion in the projection space. In our work we take a different approach from those mentioned previously. We first followed a similar approach wherein we developed a relationship between the expansion coefficients of the data and the expansion coefficients of the image. However, we found implementation of this approach to be numerically unstable. Instead we developed a method that does not require a spherical expansion of the final image. We simply keep the spherical expansion coefficients of the data until the final step of the image reconstruction procedure. Our final image is not the synthesis of the image expansion but rather it is an inner product of the expansion coefficients of the cone-beam data with the expansion coefficients of the filter, which is summed over all sampled cone-beam vertex points. In our earlier publications (Basko et al 1999, Taguchi et al 21), we used the spherical harmonic expansion to convert the cone-beam line integrals into the first derivative of the Radon planar integral; second half of Grangeat s algorithm (Grangeat 1991) was used to reconstruct the image. In this paper, we use the spherical expansion to convert the data, and to transform the tomographic filter kernel as well, in order to reconstruct the 3D image. The approach is shown to lead to a new formulation of the fan-beam reconstruction problem. For the case of fan beam, a closely related method is the direct Fourier reconstruction method. In that method the Fourier transform of the projection data is first calculated via rebinned parallel beam data. Then, by using the well-known central slice theorem the Fourier transform of the image is determined. The final image is obtained by taking the inverse Fourier transform of the result (Bracewell 1956). Our method is similar except that the fan beam data is not rebinned into the parallel geometry, the central slice theorem is not used, and there is no re-sampling of the data. This paper first presents the development of the algorithm based on spherical harmonics for 3D cone-beam reconstruction. It is then shown how similar results based on circular harmonics can be derived for 2D fan-beam reconstruction. Subsequent to that, computer simulations are presented showing results of both the reconstruction of cone-beam data acquired from a circular orbit and the reconstruction of fan-beam data. Our cone-beam result is then compared with that obtained from Grangeat s algorithm. Finally a discussion is presented. 2. Cone-beam data reconstruction In our algorithm, we first find the spherical harmonic expansion of the cone-beam projections at each cone-beam focal point location. We then transform the expansion coefficients into those

Cone-beam and fan-beam image reconstruction algorithms based on spherical and circular harmonics 2241 z Cone-beam detector θ = Cone-beam ray g ψ ( n) Vertex orbit ψ n Ray with θ = π/2 and ϕ = Figure 1. In a cone-beam imaging geometry g ( n)is a line integral of an object, and n is defined in a local coordinate system with latitude and longitude angles θ and ϕ. of the first-order derivative of Radon integrals. We finally reconstruct the image based on the Radon inversion formula. This method is different from the existing cone-beam reconstruction algorithms, e.g., Grangeat s and Katsevich s cone-beam reconstruction algorithms (Grangeat 1991, Katsevich 23) in the sense that the proposed algorithm does not require forming line-integrals on the projection plane, rebinning data or evaluating derivatives at some certain directions. 2.1. Step 1: the data acquisition Cone-beam projection g ( n) are acquired at vertex location, where n is a unit vector indicating the direction of each cone-beam projection ray (see figure 1). Here it is assumed that the detector is large enough so that the projection data are not truncated. 2.2. Step 2: the harmonic expansion of the data It is known that any periodic function (or any function defined on a circle) can be expanded as a Fourier series (circular harmonics). Similarly, any function defined on the unit sphere can be expanded in spherical harmonics. At each fixed vertex location, a spherical harmonic expansion of g ( n) is g ( n) = N l l= m= l l g lm Y lm ( n) (1) where N l is determined by the bandwidth of the projection data and Y lm ( n) can be any orthonormal basis on the surface of the sphere. Here we use Y lm ( n) = Y lm (θ, ϕ) = 1 2π lm(cos θ)e imϕ (2) with θ<π, π ϕ<π, and lm (t) = ( 1) m 2l +1(l m)! 2 (l + m)! P l m (t) for m 2l +1(l m )! 2 (l + m )! P m l (t) for m< (3)

2242 G L Zeng and G T Gullberg where Pl m (t) is the associated Legendre function and lm (t) is the normalized associated Legendre polynomial. Both Pl m (t) and lm (t) can be readily evaluated with commercially available software such as Matlab (The MathWorks 22). Taking the advantage of orthogonality of Y lm ( n), i.e. Y lm ( n)yl m ( n) d n = δ ll δ mm (4) S the expansion coefficients can be calculated as g lm = g ( n)ylm ( n) d n (5) S or g lm = 1 π π g (θ, ϕ) lm (cos θ)e imϕ sin θ dθ dϕ (6) 2π π where S represents a unit sphere. If the half-cone-angle of the cone-beam detector is α, (6) can be evaluated as g lm = 1 α π/2+α g (θ, ϕ) lm (cos θ)e imϕ sin θ dθ dϕ. (7) 2π α π/2 α 2.3. Step 3: the derivative of the Radon planar integral In order to reconstruct (see equation (11)), we need to obtain the derivatives of Radon planar integrals for all planes intersecting the point. Let us denote the derivatives of Radon planar integrals for all planes intersecting the point as p ( n), with n being the normal vector of the associated plane (see figure 3). The derivatives are taken in the direction of n. The function p ( n) is defined on the unit sphere and has a spherical harmonic expansion: p ( n) = N l l= m= l l p lm Y lm ( n) (8) where n is the normal vector of the plane. Previously, we have shown a relationship between coefficients for the expansions (1) and (8)(Baskoet al 1999, Taguchi et al 21) p lm = 2πlP l 1 ()g lm (9) where P l 1 (.) is the Legendre function and P l 1 () = ifl is even. Therefore, p ( n) = [ 2πlPl 1 ()g ] lm Ylm ( n). (1) l odd m In (1), the n is the direction of a cone-beam ray, while the same n in (1) is the normal direction of a plane. This plane is orthogonal to the ray with the direction n. 2.4. Step 4: the image reconstruction According to Radon inversion formula, a 3D image f( x)can be reconstructed from its Radon projection by f( x)= 1 8π 2 S t p(t, k) d k (11) t= x k

Cone-beam and fan-beam image reconstruction algorithms based on spherical and circular harmonics 2243 y Radon plane with n as its normal vector Cone-beam ray n ϕ t Focal point ψ λ + π ϕ λ R x k Orbit Figure 2. Top view of the xyz coordinate system with a special case where n is in the xy-plane. where p(t, k)is the first-order derivative of the Radon projections with respect to t(t ), and S represents a unit sphere. The derivative operation in the Fourier domain is a multiplication by iν where ν is the frequency, and can be implemented as a convolution with a kernel h h(t) = For band-limited signals we can use h(t) = 2πiν e 2πitν dν. (12) 2πiν e 2π itν dν = 2πν sin(2πtν)dν if t = = 2 cos(2πt) t sin(2πt) πt 2 if t. Here is the signal bandwidth and =.5 corresponds to the highest frequency that is determined by the sampling interval. Thus (11) can be rewritten as f( x)= 1 p(t, k)h( x k t)dt d k. (14) 8π 2 S In the above equations, p(t, k) can be related to p ( n). Here n is defined in the local coordinate system as shown in figure 1, and n = points to the centre of the cone-beam detector. On the other hand, k is defined in the global coordinate system where the vertexorbit vector is defined. If we consider a circular orbit (see figure 2) (13) = R (cos λ, sin λ, ) (15)

2244 G L Zeng and G T Gullberg z n p ψ ( n) Vertex orbit ψ t k θg y ϕ g x Figure 3. A Radon plane that contains a cone-beam vertex. Both n and k are normal to the plane. and let the unit vectors k be k = (sin θg cos ϕ g, sin θ g sin ϕ g, cos θ g ) (16) then the unit vector n points at the same direction as k (see figure 3) except that n is defined in a local coordinate system as shown in figure 1. The unit vector n is parametrized by its latitude and longitude angles θ and ϕ. When θ =, n points to the z-axis direction, and when ϕ =, n points to the centre of the detector. We have (see figure 2) where p(t, k)= p ( n) (17) θ g = θ (18) ϕ g = λ + π ϕ (19) t = k = R sin θ cos ϕ. (2) If we change the variables (t, θ g,ϕ g ) to variables (λ,θ,ϕ), the transformation Jacobian is given by J = R sin θ sin ϕ. (21) By changing the variables, (14) becomes f( x)= 1 2π p ( n)h(g)j dλ d n (22) 4π 2 S where g is defined as x sin θ cos(ϕ λ) g = x k t = y z sin θ sin (ϕ λ) cos θ + R sin θ cos ϕ. (23)

Cone-beam and fan-beam image reconstruction algorithms based on spherical and circular harmonics 2245 Changing the order of integrations, we have f( x)= 1 2π p ( n)h(g)j d n dλ. (24) 8π 2 S In the following, we substitute the expansion (8) into(24) and use the fact that p lm = for even l, and we have f( x)= 1 2π N l l p 8π 2 lm Y lm ( n)h(g)j d n dλ (25) or f( x)= 1 2π 8π 2 S N l l=1 l odd l=1 l odd m= l Equation (26) can be further written as f( x)= 1 2π N l l 8π 2 l=1 l odd m= l l p lm h(g)j Y lm ( n) d n dλ. (26) m= l S p lm h lm dλ (27) where h lm is the complex conjugate of the spherical harmonic expansion coefficient of h(g)j h lm = h(g)j Ylm ( n) d n. (28) S Equation (27) is the spherical harmonic cone-beam image reconstruction formula in the form of an inner product. 3. Special case: central slice reconstruction If a circular scanning orbit is used, the central slice of the image volume can be reconstructed more efficiently. The central slice is the 2D image in the plane that contains the circular cone-beam orbit. In this special case, the function g in (23) can be further simplified because z = for the central plane: x sin θ cos(ϕ λ) g = x k t = y sin θ sin(ϕ λ) cos θ + R sin θ cos ϕ (29) = [( R + x cos λ) cos ϕ + (x sin λ y cos λ) sin ϕ]sinθ (3) = [A cos ϕ + B sin ϕ]sinθ (31) where A = R + x cos λ and B = x sin λ y cos λ. (32) Let g =, we have two solutions for ϕ ( ) ( ) A A ϕ 1 = tan 1 and ϕ 2 = tan 1 (33) B B and [ 2 ϕ g = [ A sin ϕ 1 + B cos ϕ 1 ] ϕ=ϕ1] 2 sin 2 θ = (A 2 + B 2 ) sin 2 θ (34)

2246 G L Zeng and G T Gullberg [ 2 ϕ g = (A ϕ=ϕ2] 2 + B 2 ) sin 2 θ. (35) Sine A < in (32), we always have sin ϕ 1 = A/ A 2 + B 2 > and sin ϕ 2 = A/ A 2 + B 2 <. Recall from (12) that h(t) = δ (t), a delta function property gives h(g) = δ (g) = δ (ϕ ϕ 1 ) + δ (ϕ ϕ 2 ) = δ (ϕ ϕ 1 ) + δ (ϕ ϕ 2 ). (36) [g (ϕ 1 )] 2 [g (ϕ 2 )] 2 (A 2 + B 2 ) sin 2 θ Therefore, (28) can be further evaluated as follows, using (21) and (2): h lm = = = = = where S π 2π h(g)j Y lm ( n) d n (37) δ (ϕ ϕ 1 ) + δ [ ] (ϕ ϕ 2 ) 1 (R (A 2 + B 2 ) sin \ 2 sin \ θ sin ϕ ) θ 2π lm(cos θ)e imϕ sin \ θ dϕ dθ R π 2π(A 2 + B 2 lm (cos θ)dθ ) R 2π(A 2 + B 2 ) q lm R 2π(A 2 + B 2 ) q lm 2π 2π 2π (38) [δ (ϕ ϕ 1 ) + δ (ϕ ϕ 2 )] sin ϕ e imϕ dϕ (39) [δ (ϕ ϕ 1 ) + δ (ϕ ϕ 2 )] sin ϕ e imϕ dϕ (4) [δ(ϕ ϕ 1 ) + δ(ϕ ϕ 2 )][ sin ϕ e imϕ ] dϕ (41) R = 2π(A 2 + B 2 ) q B +ima lm A2 + B 2 [e imϕ 1 ( 1) m e imϕ 1 ] (42) R = π(a 2 + B 2 ) q B +ima lm A2 + B 2 e imϕ 1 if m is odd (43) if m is even q lm = π lm (cos θ)dθ (44) is independent of the projection measurements and can be pre-calculated. Finally, the reconstruction algorithm (27) for the central slice is reduced to f( x)= R 2π 1 l N l (B ima) 8π 3 (A 2 + B 2 ) 3/2 p lm q lm dλ. (45) m= l m odd Algorithm (45) is more efficient than algorithm (27). In (27), a different spherical harmonic expansion of h is required for every reconstruction point x and every focal point position (λ), while the expansion of h is not required in (45). The coefficients q lm of (44) only needs to be evaluated once. l=1 l odd

Cone-beam and fan-beam image reconstruction algorithms based on spherical and circular harmonics 2247 y Projection ray φ g = t ( x φ) x φ t λ ϕ ψ R x Orbit Figure 4. A fan-beam imaging coordinate system. Here ϕ is locally defined and φ is globally defined. 4. 2D case: fan-beam reconstruction Following the same procedure as in section 2, a similar 2D fan-beam reconstruction algorithm can be obtained. We point out to readers that we use the same notation as in section 2. For example, h in this section represents a ramp filter, while h represents a derivative filter in section 2. Let us start with the 2D Radon inversion formula for the parallel-beam data p(t, φ), and the reconstructed image f( x)is given by π f( x)= 1 p(t, φ)h( x φ t)dt dφ (46) 2 π where h is the ramp-filter kernel. For a circular fan-beam vertex orbit = R (cos λ, sin λ) (47) we can change the parallel-beam variables (t, φ) to fan-beam variables (λ, ϕ) (see figure 4): where p(t, φ) = p (ϕ) (48) t = R sin ϕ (49) φ = π 2 + λ ϕ. (5) The transformation Jacobian is J = R cos ϕ. (51) Thus (46) becomes f( x)= 1 2 2π π π p (ϕ)h(g)r cos ϕ dϕ dλ (52)

2248 G L Zeng and G T Gullberg Fan-beam projection ray y x τ = g ϕ ξ α ( x ψ) Focal point ψ λ R x Orbit Figure 5. The fan-beam ray is parametrized with a locally defined angle ξ. with g = x φ t = [ ] x y ( π ) cos 2 + λ ϕ ( π ) R sin ϕ. (53) sin 2 + λ ϕ with Let the local circular harmonic expansion of p (ϕ)r cos ϕ be p (ϕ)r cos ϕ = p m e imϕ (54) m p m = 2π p (ϕ)r cos ϕ e imϕ dϕ (55) then (52) becomes f( x)= 1 2π π p 2 m e imϕ h(g) dϕ dλ (56) π m or f( x)= 1 2π p 2 m h m dλ (57) m where h m is the complex conjugate of the circular harmonic expansion coefficient h m: h m = π π h(g) e imϕ dϕ. (58) Equation (57) is the fan-beam image reconstruction algorithm in the Fourier coefficient domain in the form of an inner product. In 2D, the circular harmonic expansion is the same as the Fourier expansion. Equation (57) is not efficient for computer implementation, because the function g depends on both x and. In the following we derive an algorithm that is more efficient than (57). For afixed x and, we introduce two angles ξ and α as defined in figure 5 with

Cone-beam and fan-beam image reconstruction algorithms based on spherical and circular harmonics 2249 and ϕ = ξ + α (59) α = λ ± π ( x ) (6) τ = x sin ξ. (61) In figure 4, g is the distance from the point x to the fan-beam projection ray. With the local coordinate system defined in figure 5, τ is the distance from the point x to the fan-beam projection ray. Thus and g = τ = x sin ξ (62) dt = d( x φ t) = dg = dτ = x cos ξ dξ. (63) It is known that the ramp-filter kernel satisfies (Horn 1979) h( τ) = h(τ) and h(aτ) = 1 h(τ) (64) a2 which give h( x φ t) = h(g) = h( τ) = h(τ) = h( x 1 sin ξ) = h(sin ξ). (65) x 2 Using (48), (59), (63) and (65), (46) becomes 2π π f( x)= 1 p 1 (ξ + α) 2 π x 2 h(sin ξ) x cos ξ dξ dλ. (66) Let the local Fourier expansion of p (ϕ)r be p (ϕ)r = m ˆp m e imϕ = m ˆp m e im(ξ+α) (67) with ˆp m = p (ϕ)r e imϕ dϕ (68) then (66) becomes f( x)= 1 2 or where 2π π π m f( x)= 1 2 ĥ m = π π ˆp m e imα e imξ 1 x h(sin ξ)cos ξ dξ dλ (69) 2π 1 x p m e imα ĥ m dλ (7) m h(sin ξ)cos ξ e imξ dξ. (71) Algorithm (7) is more efficient for computer implementation than (57), because ĥ m of (71) only needs to be evaluated once, while h m of (58) must be evaluated for each x and.

225 G L Zeng and G T Gullberg Table 1. 2D Shepp Logan phantom (scaling factor = 128). Ellipse index Half-axis ( a) Half-axis (b) Centre (x) Centre (y) Tilt angle Density 1.69.92 1.5 2.6624.874.184.98 3.11.31.22 18.2 4.16.41.22 18.2 5.21.25.35.1 6.46.46.1.1 7.46.46.1.1 8.46.23.8.65.1 9.23.23.65.1 1.23.46.6.65.1 11.333.26.5538.3858 18.3 Table 2. 3D Shepp Logan phantom (scaling factor = 2). Ellipsoid Half-axis Half-axis Half-axis Centre Centre Centre Tilt Tilt index (a) (b) (c) (x) (y) (z) angle (α) angle (β) Density 1 15.43 2.574 27.93 2 2 14.95 19.725 26.114.393.393.98 3 2.79 2.79 2.79 5.51 16.73 1 4 2.79 2.79 2.79 5.51 16.73 1 5 9.76 13.11 1.837 16.256 1 6.981.491.491 1.77 12.97 8.128.48 7.491.491.981 12.97 8.128.48 8.491.981.491 1.28 12.97 8.128.48 9.981.981.981 2.133 8.128.48 1 5.56 5.56 5.56 2.133 2.79.48 11 4.48 5.53 4.97 7.467 8.128.48 12 2.347 6.613 5.419 4.693 8.128 18.52 13 3.413 8.747 8.128 4.693 8.128 18.52 14.64 4.267 4.267 11.947 8.533 8.128 18.48 5. Computer simulations Computer simulations were performed to verify the feasibility of the proposed image reconstruction algorithms for both fan-beam and cone-beam geometries. For both imaging geometries the fan-beam or cone-beam vertex orbit was a circle with a radius of 1 units. The fan-beam and cone-beam focal length was 1 units. A slice of the image volume was 1 1 units 2. In both imaging geometries, equiangular detectors were used. The angular sampling interval on the detector was.35 which was approximately the resolution of a SPECT measurement. The fan-beam and cone-beam detectors rotated around the object 36 with 12 stops. Computer-generated Shepp Logan head phantoms were used in computer simulations. The phantom parameters are given in tables 1 and 2 for the fan-beam and cone-beam simulations, respectively. The reconstruction code was written in Matlab. The fan-beam code involved the fast Fourier transform. A 124-point FFT was used. The cone-beam code involved both the fast Fourier transform and the Legendre transform. The spherical coefficients were evaluated first by performing the Legendre transformation in

Cone-beam and fan-beam image reconstruction algorithms based on spherical and circular harmonics 2251 Figure 6. Fan-beam reconstruction of a computer-generated head phantom using algorithm (7). the θ direction and by performing the FFT in the ϕ direction. A 512-point FFT was used for the ϕ angle transformation and associated Legendre polynomials of the highest degree of 512 were used for the θ angle transformation. The proposed algorithms are numerically stable. Figure 6 shows the fan-beam reconstruction of the 2D Shepp Logan head phantom; the reconstruction algorithm was (7). Figure 7(a) shows the central slice of the cone-beam reconstruction of the 3D Shepp Logan head phantom; the reconstruction algorithm was (45). The display window was linear and was from the minimal value to maximal value in the image slice. For the cone-beam data, an image (see figure 7(b)) was also reconstructed using Grangeat s algorithm for comparison purpose. The projection data was almost the same as that described above, except that an equispaced (instead of an equiangular) cone-bean detector was used. The sampling interval was chosen such that the detector solid angles for both detectors were the same for the central ray. For other cone-beam projection rays, the equispace-detector had a finer angular sampling than the angular sampling in the equiangular-detector. Our implementation details of Grangeat s algorithm can be found in Weng et al (1993). In order to compare these two cone-beam images, line profiles are shown in figure 7(c), which indicate that the proposed algorithm gives more accurate reconstruction than Grangeat s algorithm. This can be easily seen at the thin skull regions. 6. Discussion In the cone-beam algorithm derivation, we assumed for simplicity that the cone-beam focal point orbit is a circle in a plane. In fact, this planar circular orbit does not satisfy Tuy s cone-beam data sufficiency condition (Tuy 1983). A non-planar orbit such as a helix or one consisting of a circle and lines should be used in practice. The orbit must be parametrized using a single parameter similar to (15). The remainder of the derivation in section 2 is relatively unchanged for this case. However, when a set of complete data is used, the multiply measured data must be handled properly. One way is to modify (28) by introducing a function M( x,, n) so that 1 h lm = S M( x,, n) h(g)j Y lm ( n) d n (72) where M( x,, n) is the number of times the planar integral is measured for the plane with normal direction n containing the points x and. One point of caution is that the function p(t, k) in (11) is different from the function p ( n) in (8). The function p ( n) represents the derivative of the Radon planar integral with

2252 G L Zeng and G T Gullberg (a) (b) (c) Ideal (a) (b) Figure 7. Cone-beam reconstructions (central slice) of a computer-generated Shepp Logan head phantom: (a) with algorithm (45) and (b) with Grangeat s algorithm. Line profiles along the broken line in (a) and (b) are shown in (c). respect to the variable normal to the plane for the plane passing the point with the normal direction n, while p(t, k)in (11) represents the derivative of the Radon planar integral for the plane with respect to the variable t for the plane with the normal direction k and a distance t from the (global) origin. The expression in (19) and (2) gives the transformation between the global coordinates (t, k)and the local coordinates n. There are several methods that can be used to reconstruct a 3D image from the Radon projection data. The method presented in the appendix is one that relates the spherical harmonic expansion of the image to a spherical harmonic expansion of the Radon planar projections. However, the integrals in (A.2) and (A.3) are numerically unstable, because the evaluation of these integrals requires the calculation of Legendre polynomials outside the interval [ 1, 1]. Outside this interval the Legendre polynomials have large values, which can result

Cone-beam and fan-beam image reconstruction algorithms based on spherical and circular harmonics 2253 in numerical overflow. Our numerical simulations verified this, which caused us to direct our attention towards the development of the algorithm in (27). The main differences between our proposed cone-beam method and other existing analytical cone-beam reconstruction methods (e.g., Grangeat s (1991) or Katsevich s (23) methods) are (i) the method does not involve performing a backprojection, (ii) the method does not require forming line integrals on the projection plane or evaluating derivatives along some certain directions and (iii) the method does not require any interpolation. Due to (iii), a potential advantage of the algorithm is that it may provide higher resolution cone-beam reconstructions than other approaches. If one would apply the direct Fourier method to fan-beam data, first the fan-beam data must be rebinned (re-sorted) into parallel-beam data. This rebinning step involves data interpolation. A 1D Fourier transform is taken for the rebinned parallel-beam data at each projection angle. According to the central section theorem, the Fourier transformed data is mapped on the 2D Fourier space. The 2D Fourier domain data are then re-sampled (i.e. interpolated) into uniformly spaced grid points. Finally, an inverse 2D Fourier transform is used to obtain the reconstructed image. On the other hand, our fan-beam algorithm (7) does not require rebinning the fan-beam projection data into parallel-beam data before performing the 1D Fourier transform. The transformed data are not mapped into the 2D Fourier domain. Instead, an inner product is calculated between the 1D Fourier data and a location ( x) dependent function. Finally, a weighted sum is calculated over all projection angles. No re-sampling of the data and no inverse 2D Fourier transform are utilized during image reconstruction. The fan-beam algorithm (7) does not involve any rebinning or interpolation; however, it has a higher computational cost than the direct Fourier method. The calculation complexity of a direct Fourier method is at the order of O(N 2 log N), where N is the size in an N N 2D image array and is proportional to the number of projection angles. The computational complexity for the fan-beam algorithm (7) is at the order of O(N 4 ). In general the computation burden in a reconstruction algorithm is dominated by the backprojector. A direct implementation of a backprojector for the 3D Radon data has a computational complexity of O(N 5 ), where N is the size in an N N N 3D image array and is proportional to the number of projection angles. Marr s method implements a 3D parallel planar backprojector as two sequential 2D parallel linear backprojectors (Marr 1974). As a result, the computational complexity is reduced to O(N 4 ). If linogram method (Axelsson and Danielsson 1994) is adapted the computational complexity can be further reduced to O(N 3 log N). The proposed cone-beam algorithm (45), on the other hand, has a poor computational efficiency, and the computational complexity is at the order of O(N 6 ). The proposed fan-beam algorithm (7) is equivalent to the conventional fan-beam filtered backprojection (FBP) algorithm if the data sampling is continuous. In (7) if one replaces m ˆp m e imα ĥ m by p (ϕ) h(ϕ), where is a convolution operator, (7) becomes the conventional fan-beam FBP algorithm. In fact, substituting the Fourier expansion (67) (ignoring the constant R )intop (ϕ) gives p (ϕ) h(ϕ) = m ˆp m [e imϕ h(ϕ)] = m ˆp m e imα ĥ m. (73) In discrete implementation these two methods are not equivalent. In the FBP algorithm the discrete samples of the convolution p (ϕ) h(ϕ) need some interpolation in the backprojection step. On the other hand in the Fourier expansion, a shift from a sampling point to any location is an exact phase shift in the Fourier expansion. Thus, data interpolation can be avoided using the expansion method.

2254 G L Zeng and G T Gullberg Fan-beam FBP algorithms for equiangular and equispaced measurements are almost the same except for some weighting factors. However, for the circular expansion method (7), the measurements are assumed to be equiangular sampled. If the measurements are equispace sampled, one cannot use the circular expansion directly. Instead, the angular variable (such as ϕ) must first be changed into the linear variable (say, t) along the detector surface, the Jacobian factor is then evaluated to relate these two variables, and the expansion is finally carried out in terms of t. In other words, the expansion method (7) can still be used for equispaced fan-beam measurements, except that an extra Jacobian factor needs to be introduced. The aim of this paper was to investigate a different approach to reconstructing conebeam and fan-beam images using orthogonal polynomials. The algorithms still need to be carefully evaluated and compared with existing algorithms in terms of propagation of noise and reconstruction accuracy. The idea of using orthogonal polynomials in image reconstruction is not new and can be extended to other complex cone-beam geometries and potentially other more general imaging geometries as well. The goal is to some day be able to specify sets of orthogonal polynomials that represent singular value decompositions of complex cone-beam geometries. Unlike Grangeat s algorithm, the proposed algorithm processes data one cone-beam projection at a time, and the algorithm could be executed while the data acquisition is in progress. Our preliminary comparison study shows that the proposed spherical harmonics expansion cone-beam image reconstruction method provides better spatial resolution than the image reconstructed with Grangeat s algorithm. More careful and systematic comparison studies will be carried out in future work. The main drawback of our proposed algorithms is the poor computational efficiency. Our future work will focus on improving the computational efficiency of the reconstruction algorithms that use spherical and circular harmonic expansions. Acknowledgments This work was supported by the National Institute of Biomedical Imaging and Bioengineering, and National Cancer Institute of the National Institutes of Health under grants R21-CA1181, R1-EB121 and by the Director, Office of Science, Office of Biological and Environmental Research, Medical Sciences Division of the US Department of Energy under contract DE-AC3-76SF98. We thank Dr Katsuyuki Taguchi of Medical Systems Company, Toshiba Corporation for assistance in this project. Appendix In this appendix, an alternative approach of a spherical harmonic cone-beam image reconstruction algorithm is presented. This approach is in line with the traditional circular harmonic image reconstruction methods. Step 1. Convert the cone-beam line integral to the first derivative of the Radon planar integral via spherical harmonic expansion. See (1) and (1). The first derivative of the Radon planar integrals is p ( n) which has a spherical harmonic expansion (8) with coefficients p lm. Step 2. Rebin the convergent data format p ( n) to parallel data format p(t, k).

Cone-beam and fan-beam image reconstruction algorithms based on spherical and circular harmonics 2255 The rebinning equation is given by (17). In the Radon space we have the first derivative of the planar integral p(t, k)using (18) (2). Step 3. Evaluate the spherical harmonic expansion of p(t, k): p(t, k)= N l l= m= l l p lm (t)y lm ( k). (A.1) It is possible to evaluate the coefficients p lm (t) directly from the coefficients p lm using (18) (2). Step 4. Evaluate the Legendre integrals q lm (r) = 1 ( ) t p lm 2πr (t)p l dt. (A.2) r r This integral can also be evaluated via using integration by parts q lm (r) = 1 2πr p lm(r) 1 ( ) t p 2πr 2 lm (t)p l dt. (A.3) r r Thus( calculating the derivative of p lm (t) numerically can be avoided, while an exact expression of P l t ) r is available. Step 5. Sum the harmonic expansion to obtain the reconstruction. The 3D image is finally reconstructed by evaluating the following summation: f(r, ω) = N l l= m= l l q lm (r)y lm ( ω) (A.4) where ω is a unit vector in the 3D image space, and the point, (r ω), is reconstructed. Steps 4 and 5 are based on relations given by Deans (1983). If a 3D object f(r, ω) can be expressed as a spherical harmonic expansion f(r, ω) = q lm (r)y lm ( ω) (A.5) l m then the Radon transform (i.e. the planar integrals) of f(r, ω) is where f(t, ˆ k) = ˆq lm (t)y lm ( k). l m Here q lm (r) and ˆq lm (t) are related by the following Legendre transform pair: ˆq lm (t) = 2π q lm (r) = 1 2πr t ˆq lm (t) = d2 dt 2 ˆq lm(t). r ) dr ( t rq lm (r)p l r ( ) t ˆq lm (t)p l dt r (A.6) (A.7) (A.8)

2256 G L Zeng and G T Gullberg References Axelsson C and Danielsson P E 1994 Three-dimensional reconstruction from cone-beam data in O(N 3 logn ) time Phys. Med. Biol. 39 477 91 Basko R, Zeng G L and Gullberg G T 1999 Application of spherical harmonics to cone-beam image reconstruction Conf. Rec. 1998 IEEE Nuclear Science Symp. and Medical Imaging Conf. (Toronto, Canada) (New York: IEEE) pp 1649 5 Bracewell R N 1956 Strip integration in radio astronomy Aust. J. Phys. 9 198 217 Cormack A M 1963 Representation of a function by its line integrals, with some radiological applications J. Appl. Phys. 34 2722 7 Cormack A M 1964 Representation of a function by its line integrals, with some radiological applications: II J. Appl. Phys. 35 195 27 Deans S R 1983 The Radon Transform and Some of its Applications (New York: Wiley) Grangeat P 1991 Mathematical framework of cone-beam 3D reconstruction via the first derivative of the Radon transform Mathematical Methods in Tomography (Lecture Notes in Mathematics vol 1497) ed G T Herman, A K Louis and F Natterer (Berlin: Springer) pp 66 97 Hansen E W 1981 Theory of circular image reconstruction J. Opt. Soc. Am. 71 34 8 Hawkins W G, Leichner P K and Yang N-C 1988 The circular harmonic transform for SPECT reconstruction and boundary conditions on the Fourier transform of the sinogram IEEE Trans. Med. Imaging 7 135 48 Horn B K P 1979 Fan-beam reconstruction methods Proc. IEEE 67 1616 23 Katsevich A 23 A general scheme for constructing inversion algorithm for cone beam CT Int. J. Math. 21 135 21 Marr R B 1974 On the reconstruction of a function on a circular domain from a sampling of its line integrals J. Math. Anal. Appl. 45 357 74 Taguchi K, Zeng G L and Gullberg G T 21 Cone-beam image reconstruction using spherical harmonics Phys. Med. Biol. 46 N127 38 The MathWorks 22 Learning Matlab 6.5 Tuy H 1983 An inversion formula for cone-beam reconstruction SIAM J. Appl. Math. 43 546 52 Weng Y, Zeng G L and Gullberg G T 1993 A reconstruction algorithm for helical cone-beam SPECT IEEE Trans. Nucl. Sci. 4 192 11 You J, Liang Z and Bao S 1998 A harmonic decomposition reconstruction algorithm for spatially varying focal length collimators IEEE Trans. Med. Imaging 17 995 12