Sunteți pe pagina 1din 4

INTERPOLATION COMPARISON IN THE DIRECT FOURIER METHOD FOR TOMOGRAPHIC RECONSTRUCTION D. F. G. de Azevedo1, P. C. C. Godoy2, S. Helegda1, A. A. D. de Mattos2, D. L. R.

Vidor2
1

Department of Electrical Engineering, Pontifical Catholic University, Porto Alegre, Brazil 2 Department of Electrical Engineering, Lutheran University of Brazil, Canoas, Brazil
ABSTRACT
g(i,) y
1D Fourier Transform

It seems to be general consensus seems that the Direct Fourier Method for tomographic reconstruction produces an image quality inferior to the other well-know methods. The crucial point is usually the interpolation step in the frequency domain. This paper compares different interpolation techniques to improve the image quality of the Direct Fourier Method. In order to analyze the efficiency of the results we evaluate the quality of the images with respect to Relative Error and the time of reconstruction for the same data. We analyze the data using a tomography platform consisting of a set of programs written in ANSI C++.

F(u,v) i x i G(i,) u

f(x,y)

Spatial Domain

Frequency Domain

Figure 1 The Fourier Slice Theorem.

1. INTRODUCTION Tomographic techniques are used in radiology and in many branches of science and technology to image 2D cross sections of 3D objects [1]. There are many methods to perform data acquisition and reconstruction of an object. In this paper we are especially interested in the investigation of the Direct Fourier Method (DF Method) for reconstruction, because of its low computation complexity advantage [1][2][3]. This method is a direct application of the Fourier Slice Theorem. This theorem is illustrated in Figure 1 and is stated as follows: The Fourier transform G(i,) of a parallel projection of an image f(x,y) taken at angle i is found in the twodimensional Fourier transform F(u,v) on a line subtending the angle i with the u-axis. [2][3]. The Fourier Slice Theorem indicates that we can determine the values of F(u, v) radially by taking the Fourier transform of each the projections of an object function at angles 1, 2, ..., n. If an infinite number of projections is taken, then F(u,v) will be known at all points in the uv-plane, and the object function f(x,y) can be recovered by using the inverse 2D Fourier transform [2].

In practice only a finite number of projections of an object can be acquired, and each projection consists of a finite number of points. Therefore, primarily, F(u,v) is only known at discrete points along the radial lines. To compute the inverse Fourier Transform (using a FFT algorithm) we must determine the values on a square grid using an interpolation method. A summary of Direct Fourier Method is shown in Figure 2.
y v

f(x,y)

F(u,v)

Pa ra lle l

Pr oj e cti on
y y

1D Fourier Transform

Interpolation
v

f(x,y)

2D Inverse Fourier Transform

F(u,v)

x x

Figure 2 - Overview of Direct Fourier Method.

In spite of the potential speed advantage of the DF Method as compared to other well-know methods, it has not been used in practice because of its inferior image quality. There are two main sources of artifacts: the interpolation procedure in frequency domain and the inherent ramp filtering (circular convolution) [4]. 2. METHODOLOGY We often use synthetic representations of objects in the development of new imaging techniques and tools. The simulation of the human body, in projection radiology, is called a phantom. In our simulation the phantom is described in two dimensions as a collection of ellipses with different density functions. We have used the SheppLogan Head Phantom, one standard phantom in the literature (Figure 3.a) [2]. This phantom shown is a standard in the field, and is meant to represent a crosssection of the human brain, containing both large and small objects at both high and low intensities.

A possible remedy for circular convolution in the signal domain is oversampling the F(u,v) function. The resampled grid density factor (ng) increases the number of data points in the Cartesian Fourier Domain from N2 to (ngN)2 [4]. We compared different algorithms to solve the interpolation problem in the frequency domain: Nearest Neighborhood [1], Linear Radial [2], Bilinear [2], Lagrange polynomial (with degree 4 and 8) [6], Sinc filter (with 4 and 8 point) [5] and Cos3 filter (with 8 and 16 point) [4]. We varied the following parameters during each reconstruction: the zero-padding factor (nz), the grid density factor (ng) and the type of interpolation algorithm. Measurements of the reconstruction accuracy were obtained by point-wise comparisons between the digitized picture of the object (the phantom) and the reconstructed image. To measure output image quality we used the error figure Relative Error (RE), defined as follows: The Relative Error:

RE =
Where: frec(x, y) fobj(x, y)

y x

frec( x, y ) fobj ( x, y )

y x

fobj ( x, y )

(1)

= pixel of the reconstructed image. = pixel of the digitized phantom.

a) the Phantom

b) The Sinogram

In tomography reconstruction, another important measurement is the time to reconstruct. The reconstruction time was measured in seconds, considering only the starting point and the ending point of each reconstruction procedure. In all experiments we used a 1.8 GHz AthlonXP PC based microcomputer. 3. RESULTS Figure 4 shows the Relative Error (RE) variation for different interpolation algorithms using a fixed zero-padding factor and a variable grid density factor of 1, 2 and 4. Figure 5 shows the reconstruction time versus the type of interpolation, for different ng and nz factors. The increase of ng led to a reduction of the error in the less sophisticated interpolation methods (nearest neighborhood and linear radial). For the other interpolation methods we noticed error reduction only for ng 4. If we increase the ng factor, the number of points of interpolation increases by a factor of ng2. This causes a considerable increase in the reconstruction time. However, in the visual analysis the ng factor improved the image quality (less artifacts), despite the RE increase. The increase of nz diminishes the RE in all interpolation algorithms, except for the Sinc 4-point interpolation. As shown in [7], when nz is higher than 8, no significant improvement is apparent. The time to

Figure 3 - The Shepp-Logan Head Phantom

The output data of the simulation tool is pilled up in a sinogram file (Figure 3.b). A sinogram is the collection of parallel projections of the object taken at equidistant angles, and forming a map of the projection data. The simulation tools we developed can vary several parameters that directly affect the measurements such as the local geometry of the tomograph, choice of the phantom, the number of sensors, the number of projections and the relative position of all system components (object, source and starting angles). In all the experiments described in this paper we used the same sinogram with 512 sensors and 300 parallel projections. To reduce the imperfect interpolation artifacts in the reconstruction, we can either use a more sophisticated interpolation procedure and/or we can append zeros in the signal domain (zero padding). The zero-padding factor (nz) increases the number of radial data points in the Fourier domain from N to nzN, leading to a more accurate interpolation process [4].

Figure 4 Relative Error x Interpolation Algorithm

Figure 5 Time to Reconstruct x Interpolation Algorithm

reconstruct does not degrade much with this parameter because it affects only the 1D FFT of the projections. The good results with respect to the RE were obtained with the Lagranges polynomial interpolation (with degree 8). With a factor nz=2 we obtained an error value close to the value obtained with nz=8. In this case, the reconstruction time did not increase excessively. However, in the visual analysis is perceived the effects of the circular convolution. Is necessary that ng=4 for a good visualization of the structures. For the Cos3 interpolation method, the reconstruction time was the critical parameter. The good results with respect to RE were overcome by the complexity of the algorithm, and its consequent long reconstruction time. The same values of the RE were obtained with simpler interpolation functions. The Sinc function interpolation method presented RE compatible with simpler interpolation functions, producing worse results with larger values of the nz factor. For the Nearest Neighborhood interpolation method, we obtained the same results found in [7]. The Linear

Radial interpolation function is the good choice when it is used with higher nz and ng factors (typically nz=8 and ng=4). The reconstruction time in this case is relatively lower. The Bilinear interpolation function produces RE values compatible with the Linear Radial interpolation function, suggesting the use of the last function due to its relative simplicity. 4. CONCLUSIONS We have successfully designed and implemented a tomography platform, consisting of a set of (1) tools to simulate the data generation in a tomograph and (2) a set of tools to reconstruct a 3D object based on its 2D projections. These tools, written in C++, should help investigators to develop and to evaluate tomographic techniques and parameters. We compared different interpolation algorithms to improve the image quality in DF Method for tomographic reconstruction. We showed that good results can be obtained with Lagranges Polynomials (with degree 4 and 8). In this case, the reconstruction time is not excessively long, keeping a good compromise with respect to the computational complexity of the DF Method. More complex interpolation methods were tested but the resulting image quality with respect to the relative error did not improve significantly, while excessively increasing the reconstruction time. The software and the images used in this paper can be sent by email upon request. The requests can be sent for pcgodoy@portoweb.com.br. 5. ACKNOWLEDGMENT We would like to the Lutheran University of Brazil and the Pontifical Catholic University of Rio Grande do Sul for the support. 6. REFERENCES [1] Natterer F. Numerical Methods in Tomography. Institut fr Numerische und instrumentelle Mathematik, Universitt Mnster, Germany. [2] Magnusson M. The Linogram Method for Image Reconstruction from Projections. Thesis no. 192, Department of Eletrical Engineering, Linkping University, Sweden. [3] Macovsky A. Medical Imaging Systems, Prentice-Hall Inc., USA, data . [4] Magnusson M., Danielsson P., Edholm P. Artefacts and Remedies in Direct Fourier Tomographic Reconstruction. Image Processing Laboratory, Dept.

of Eletrical Engineering , Linkping University, Sweden. [5] Stark H., Woods J. W., Indraneel P., Rajesh H. Direct Fourier Reconstruction in Computer Tomography, IEEE Trans. Acoust. Speech Signal Processing, vol ASSP-29, pp. 237-245, 1981. [6] Cludio, D. M., Marins, J. M., Clculo Numrico Computacional Teoria e Prtica, Atlas, Brazil, 1989. [7] Azevedo, D. F. G., Helegda S., Godoy, P. C. C., Castro, F.,Castro, M. C., Tomography Simulation and Reconstruction Tools Applied in the Evaluation of 23rd Annual Parameters and Techniques, International Conference of the IEEE Engineering in Medicine and Biology Society, Istambul, 2001.

S-ar putea să vă placă și