Advancements of Medical Image Enhancement in Healthcare Applications
View this Special IssueResearch Article  Open Access
FrameBased CT Image Reconstruction via the Balanced Approach
Abstract
Framebased regularization method as one kind of sparsity representation method has been developed in recent years and has been proved to be an efficient method for CT image reconstruction. However, most of the developed CT image reconstruction methods are analysisbased frame methods. This paper proposes a novel framebased balanced hybrid model with two sparse regularization terms for CT image reconstruction. We generalize the fast alternating direction method to solve the proposed model so that every subproblem can be easily solved. The numerical experiments suggest that the proposed hybrid balancedbased wavelet regularization scheme is efficient in terms of reducing the defined reconstruction root mean squared error and improving the signal to noise ratio in CT image reconstruction.
1. Introduction
Xray computed tomography (CT) image reconstruction is an indispensable tool for diagnosing diseases and research requirements. However, Xray radiation is harmful, and highdose Xray radiation may induce genetic mutation, cell canceration, and so on [1]. Therefore, more and more attentions are paid on the lowdose Xray CT image reconstruction. Since Xray CT imaging quality depends on the Xray dose, reducing the Xray dose will result to poor image reconstruction quality. Consequently, how to decrease the Xray dose not to affect the diagnosis is a hot topic in recent years. Mathematically, CT image reconstruction often can be formulated as a linear inverse problem. For the detected measurements data , the objective is to find the targeted image from the following equation: where is the discrete radon linear transform operator and denotes the noise with variance .
One strategy for Xray radiationdose reduction is to reduce the projection data, but this fewview method will result in insufficient data. As a result of the undersampling of this strategy and the system errors, the abovementioned problem (1) is usually illposed from the mathematical point of view. Therefore, traditional filtered back projection (FBP) method [2] cannot yield desirable imaging quality. Sparse regularization methods are developed in recent years to overcome the illposedness of these problems, and moreover, these methods can acquire higher quality images in fewview circumstance. The sparse representation methods assume that images are sparse in some transformed domains. Discrete gradient [3], that is, the socalled total variation (TV), is such a sparse transform domain to solve the fewview image reconstruction problems. Although the TVbased regularization sparse domain is useful in reducing radiation dose [4, 5], there are also some shortcomings, for example, the staircase effect [6], and the power of it is still limited [7]. Therefore, many improved methods are proposed, such as PWLSTGV proposed by Niu et al. [8], TVSPOCS proposed by Liu et al. [9], and the method proposed by Ritschl et al. [10]. Besides, some other sparse transform domains are developed, such as the gamma regularizationbased method [11], some nonlocal domains [12, 13], different kinds of wavelet frame domains [14–16], and some dictionary learning sparse methods [17, 18].
This paper mainly considers the sparse representation by wavelet tight frame. Wavelet tight frame can ensure the given signal be perfectly represented as a linear combination of the sparse wavelet coefficients which is also called the perfect reconstruction property [19]. Due to the flexibility of decomposition and reconstruction and the well performance, wavelet tight framebased methods have been widely used in almost every branch of image processing [14, 20–22]. In recent years, most of the developed wavelet tight frame methods for CT image reconstruction are analysisbased wavelet frame methods. This paper proposes a novel balancedbased method. The remainder of this paper is organized as follows. In Section 2, we present some necessary preliminaries and theories about the wavelet tight frame. Section 3 presents the proposed balanced model for CT image reconstruction and develops the efficient computational algorithm for solving the proposed strategy. Numerical simulations to demonstrate the improvement of our proposed method in terms of RMSE and PSNR are given in Section 4. In the end, we make the conclusion in Section 5.
2. Preliminaries
For convenience, we first present some basic definitions and some results of the wavelet tight frame used in the proposed model. More details can be seen in [20, 21].
In the 1D discrete circumstance, a set of vectors is called a wavelet tight frame if for each , where denotes the inner product. The corresponding analysis operator denoted as is written as
Then, is called the wavelet coefficients. Another operator which is usually called the synthesis operator is the synthesis of the wavelet coefficients, that is, if denotes the wavelet coefficients, then
Then which can be derived by the identity (2). Here, is the identity operator. This property is often called the “perfect reconstruction property,” which can reduce the calculation amount in some applications.
Based on the Unitary Extension Principle (UEP) condition [20], the wavelet tight frame often can be generated by some filters that satisfy
Here, when ; otherwise it is zero. Piecewise linear Bspline framelet is such a wavelet tight frame whose associated filters are
The waveletbased method has been used in almost every image processing branch [22–24]. Twodimensional wavelet tight frame filters can be obtained by the tensor product of the corresponding onedimensional filters.
Given the observed data ; the sparse regularization methods for image processing based on the wavelet tight frame can be summarized as where denotes a linear transform operator, which is a discrete radon transform in CT image reconstruction, a Fourier transform in MR image reconstruction, and a convolution operator in image deblurring. denotes the norm so as to finally obtain a sparse solution. as the approximation of is often used to realize the sparse regularization. Compared with , is convex so that the wellposed property can be guaranteed. Then, the target image . Here, the middle term in (8), that is, , is used to balance the distance between the target image and the coefficient . Then, in terms of different values of , three approaches are distinguished, that is, the analysis approach , the balanced approach , and the synthesis approach . Obviously, the three approaches are the same when is orthogonal. Generally speaking, it is difficult to make a conclusion that which approach among the three approaches described in (8) is better. Every approach has its own favorite image sets [25].
For the balancedbased approach, we have the following result [19]. Lemma 1. Let and , respectively, denote the analysis operator and the synthesis operator of a wavelet tight frame; then holds.
3. The Proposed Model and Method
In recent years, most of the developed methods for CT image reconstruction are analysisbased frame methods [16, 21, 26], and many stateoftheart methods such as Split Bregman method [27], alternative direction [28], and augmented Lagrangian method [29] are implemented to solve these problems. The researches of image reconstruction modeled based on balanced method [25] and synthesis method [30] are relatively few. Since every approach has its own favorite image sets, it is difficult to make a conclusion which one is better [25].
This paper proposes a novelconstrained balancedbased model CT image reconstruction as follows: where denotes a radon transform operator, is the obtained data from the scanner, and is the wavelet tight frame. Then, the reconstructed image , where denotes the solution of (9). Compared with the model in (8), the added term is used to regularize the solution further and avoid the Gibbs defects bring from the wavelet tight framelet. In general, the parameter is chosen to be smaller than the parameter . Actually, bigger will result in the overall smoothness of the results’ image. In recent years, some efficient algorithms are developed for the balancedbased models, for example, proximal linearized alternating direction method in which the linearization of quadratic term of the augmented function for the model with one regularization term was used [31].
Next, let us investigate the corresponding flexible iteration algorithm for our proposed balancedbased model. We also can use the fast alternating direction method for solving our proposed balancedbased model. The corresponding convergence analysis of the alternating direction method can resort to the [32, 33]. By introducing , the constrained minimization problem (9) can be changed into the following unconstrained one:
Actually, (10) is equivalent to the following: where and are parameters. Here, we omit the relationship between these parameters and the parameters in (10).
Here, we choose for the convenience of calculation. Then, based on Lemma 1, (11) is equivalent to the following:
Then by the alternative direction method, the minimization problem (12) can be decomposed into the following four subproblems:
As for the first subproblem, by the KKT condition, can be easily solved by the wellknown conjugate gradient (CG) method. On one hand, the added term in (12) can regularize the solution further; on the other hand, we can see from the solution of the usubproblem (13) that parameter can further overcome the illconditioning of the operator in CT scanning. Then, we summarize our proposed algorithm based on balanced wavelet tight frame approach for CT image reconstruction as in Algorithm 1.

4. Numerical Simulations
In this section, two numerical studies are presented to illustrate the well performance of our proposed scheme. We compare our proposed hybrid regularization scheme in (12) with the traditional FBP algorithm and the one with only one regularization term by our proposed balancedbased frame algorithm. In the following two simulations, we adapt the piecewise linear Bspline tight frame as the transform operator. For evaluating the quality of the reconstructed images, the following root mean squared error (RMSE) and the peak value signal to noise ratio (PSNR) are used:
Here, and denote the reconstructed image and the original image, respectively, and and denote the size of the image. Generally speaking, smaller RMSE means better reconstructed quality and higher PSNR means more closer to the original images. Some other appraisal criteria can also be used, such as the SSIM [34]. The iteration is stopped when . In these experiments, we use GPU (graphics processing units) to accelerate the computation of the and for more fast reconstruction [35]. In the following simulations, the projection views were equally distributed over and detector offset was not to be considered. All the experiments are performed by Matlab 2009 on the PC with 64bit operating system and 2.90 GHz processor. In addition, after a great deal of our simulations, we found that the parameter should be much smaller than the parameter , since larger will result in the overall smoothness of the reconstruction image. The parameter can be set as 5 or 6 when the noise and the projection views are not changed greatly. The stronger the noise, the greater the parameter should be used. Next, we will present two concrete examples. Example 4.1. We use “NCAT” shown in Figure 1(a) to evaluate the proposed model and algorithm. In recent years, “NCAT” has been widely used to evaluate medical imaging technology. The reconstructions and by FBP method (Figure 2(a)) with 50 projection views are 0.2701 and 69.0921, respectively. Figures 2(b) and 2(c) are, respectively, the recovered results with 50 projection views by models with one regularization and two regularizations in (12) based on our proposed balanced approach with algorithm parameters . These parameters are optimized. 23 iterations are carried out with about 3.6 s. The values of and are 0.0271 and 89.0138, respectively, for two regularization term balanced algorithm and 0.0305 and 87.9635, respectively, for the corresponding one regularization term balanced algorithm. We can see that in terms of RMSE, 12.5% error was reduced with the added regularization terms . The reconstruction quality is more pleasing for the balanced algorithm with two regularization term than the corresponding balanced algorithm with one regularization term.
(a) Original “NCAT” image for simulation
(b) Original “SL” image for simulation
(a) Result by FBP method without noise
(b) Result by one regularization term without noise
(c) Result by two regularization terms without noise
(d) Result by FBP method with noise
(e) Result by one regularization term with noise
(f) Result by two regularization terms with noise
(a) Result by FBP method with 40 views
(b) Result by one regularization term with 40 views
(c) Result by two regularization terms with 40 views
(d) Result by FBP method with 50 views
(e) Result by one regularization term with 50 views
(f) Result by two regularization terms with 50 views
Due to the device defect or the lowexposing dose, noise is often inevitable. So, we add 1% noise to the projection data of “NCAT.” The corresponding reconstructions and by FBP method (Figure 2(d)) are 0.2751 and 68.8931, respectively. The corresponding reconstructions and are 0.0338 and 87.1018, respectively, for two regularization term balancedbased algorithm (Figure 2(f)) and 0.0362 and 86.2801, respectively, for the one regularization term balancedbased algorithm (Figure 2(e)). Consequently, in the environments both with noise and without noise, the balanced algorithm with two regularization terms can obtain more better reconstruction quality in terms of and . Example 4.2. SheppLogan image “SL” as our second test image can be seen in Figure 1(b). We evaluate algorithms under different views. The corresponding reconstructions and by FBP method with 40 projection views (Figure 3(a)) are 0.5906 and 64.8555, respectively. The corresponding reconstructions and by FBP method with 50 projection views (Figure 3(d)) are 0.5726 and 65.1283, respectively. Obviously, the streak artifacts are introduced by FBP algorithm with fewview projections. Figures 3(c) and 3(d) are, respectively, the recovered results with 40 projection views by models with one regularization in (12) and two regularizations in (12) based on our proposed balanced approach with algorithm parameters . Every parameter has been optimized. The values of are 0.0780 and 0.0759. The values of are 82.1841 and 82.6393. The reconstruction results with 50 projection views are displayed in Figures 3(e) and 3(f). The values of are 0.0581 and 0.0557. The values of are 84.2616 and 85.9924. The RMSE and PSNR of our proposed hybrid regularization method are improved under different views which means more pleasing reconstruction results. Both the two kinds of the balancedbased image reconstruction schemes yield more desiring results than the FBP method.
5. Conclusion
In this paper, we propose a balancedbased wavelet CT reconstruction model with two regularization term. We investigate the fast algorithm for the proposed hybrid model based on the alternative direction method. Simulation results evidently demonstrate the superiority of our proposed scheme in reducing the values of RMSE and promoting the PSNR. This method is very flexible and can also be easily generalized to some other image processing problems.
Conflicts of Interest
The authors declare that there is no conflict of interest regarding the publication of this paper.
Authors’ Contributions
Weifeng Zhou and Hua Xiang are cofirst authors with equal contribution.
Acknowledgments
This work was supported by the Natural Science Foundation of China (NSFC) (11526118 and 51605238), Natural Science Foundation of Shandong Province (ZR2015PF001), Science and Technology Planning Project of Shandong Province (2015GGX101020), the Foundation of China Postdoctoral (2015M570594), and the Talent Startup Fund of Qingdao Science and Technology University (010022670).
References
 D. J. Brenner and E. J. Hall, “Computed tomographyan increasing source of radiation exposure,” New England Journal of Medicine, vol. 357, pp. 2277–2284, 2007. View at: Publisher Site  Google Scholar
 I. Elbakri and J. Fessler, “Statistical image reconstruction for polyenergetic Xray computed tomography,” IEEE transactions on medical imaging, vol. 21, no. 2, pp. 89–99, 2002. View at: Publisher Site  Google Scholar
 L. Rudin, S. Osher, and E. Fatemi, “Nonlinear total variation based noise removal algorithms,” Physica D, vol. 60, no. 1–4, pp. 259–268, 1992. View at: Publisher Site  Google Scholar
 H. Yu and G. Wang, “Compressed sensing based interior tomography,” Physics in Medicine and Biology, vol. 54, no. 9, pp. 2791–2805, 2009. View at: Publisher Site  Google Scholar
 H. Yu, J. S. Yang, M. Jiang, and G. Wang, “Supplemental analysis on compressed sensing based interior tomography,” Physics in Medicine and Biology, vol. 54, no. 18, pp. 425–432, 2009. View at: Publisher Site  Google Scholar
 J. Tang, B. Nett, and G. Chen, “Performance comparison between total variation (TV)based compressed sensing and statistical iterative reconstruction algorithms,” Physics in Medicine and Biology, vol. 54, no. 19, pp. 5781–5804, 2009. View at: Publisher Site  Google Scholar
 Y. Liu, J. Ma, Y. Fan, and Z. Liang, “Adaptiveweighted total variation minimization for sparse data toward lowdose Xray computed tomography image reconstruction,” Physics in Medicine and Biology, vol. 57, no. 23, pp. 7923–7956, 2012. View at: Publisher Site  Google Scholar
 S. Niu, Y. Gao, Z. Bian et al., “Sparseview xray CT reconstruction via total generalized variation regularization,” Physics in medicine and biology, vol. 59, no. 12, pp. 2997–3017, 2014. View at: Publisher Site  Google Scholar
 Y. Liu, Z. Liang, J. Ma et al., “Total variationstokes strategy for sparseview Xray CT image reconstruction,” IEEE transactions on medical imaging, vol. 33, no. 3, pp. 749–763, 2014. View at: Publisher Site  Google Scholar
 L. Ritschl, F. Bergner, C. Fleischmann, and M. Kachelrie, “Improved total variationbased CT image reconstruction applied to clinical data,” Physics in Medicine and Biology, vol. 56, no. 6, pp. 1545–1561, 2011. View at: Publisher Site  Google Scholar
 J. Zhang, Y. Hu, J. Yang, Y. Chen, J. L. Coatrieux, and L. Luo, “Sparseview Xray CT reconstruction with gamma regularization,” Neurocomputing, vol. 230, pp. 251–269, 2017. View at: Publisher Site  Google Scholar
 Y. Wang, F. Shu, W. Li, and C. Zhang, “An adaptive nonlocal filtering for lowdose CT in both image and projection domains,” Journal of computational design and engineering, vol. 2, no. 2, pp. 113–118, 2015. View at: Publisher Site  Google Scholar
 X. Jia, Y. Lou, B. Dong, Z. Tian, and S. Jiang, “4D computed tomography reconstruction from fewprojection data via temporal nonlocal regularization,” Medical Image Computing and ComputerAssisted Intervention–MICCAI, vol. 13, Part 1, pp. 143–150, 2010. View at: Google Scholar
 B. Zhao, H. Gao, H. Ding, and S. Molloi, “Tightframe based iterative image reconstruction for spectral breast CT,” Medical Physics, vol. 40, no. 3, article 031905, 2013. View at: Publisher Site  Google Scholar
 H. Gao, R. Li, Y. Lin, and L. Xing, “4D cone beam CT via spatiotemporal tensor framelet,” Medical Physics, vol. 39, no. 11, pp. 6943–6946, 2012. View at: Publisher Site  Google Scholar
 W. Zhou, J. Cai, and H. Gao, “Adaptive tight frame based medical image reconstruction: a proofofconcept study for computed tomography,” Inverse Problems, vol. 29, no. 12, article 125006, 2013. View at: Google Scholar
 Q. Xu, H. Yu, X. Mou, L. Zhang, J. Hsieh, and G. Wang, “Lowdose Xray CT reconstruction via dictionary learning,” IEEE Transactions on Medical Imaging, vol. 31, no. 9, pp. 1682–1697, 2012. View at: Publisher Site  Google Scholar
 Y. Lu, J. Zhao, and G. Wang, “Fewview image reconstruction with dual dictionaries,” Physics in Medicine and Biology, vol. 57, no. 1, pp. 173–189, 2012. View at: Publisher Site  Google Scholar
 J.F. Cai, S. Huang, H. Ji, Z. Shen, and G. Ye, “Datadriven tight frame construction and image denoising,” Applied and Computational Harmonic Analysis, vol. 37, no. 1, pp. 89–105, 2014. View at: Publisher Site  Google Scholar
 I. Daubechies, B. Han, A. Ron, and Z. Shen, “Framelets: MRAbased constructions of wavelet frames,” Applied and Computational Harmonic Analysis, vol. 14, pp. 1–6, 2003. View at: Publisher Site  Google Scholar
 Y. Wang, G. Wang, S. Mao et al., “A spectral interior CT by a frameletbased reconstruction algorithm,” Journal of Xray Science and Technology, vol. 24, no. 6, pp. 771–785, 2016. View at: Publisher Site  Google Scholar
 C. Bao, B. Dong, L. Hou, Z. Shen, X. Zhang, and X. Zhang, “Image restoration by minimizing zero norm of wavelet frame coefficients,” Inverse problems, vol. 32, no. 1, pp. 1–28, 2016. View at: Publisher Site  Google Scholar
 X. Zhang, W. Zhou, X. Zhang, and H. Gao, “Forwardbackward splitting method for quantitative photoacoustic tomography,” Inverse problems, vol. 30, article 125012, pp. 1–19, 2014. View at: Google Scholar
 Y. Liu, J.F. Cai, Z. Zhan et al., “Balanced sparse model for tight frames in compressed sensing magnetic resonance imaging,” PLoS One, vol. 10, no. 4, article e0119584, 2015. View at: Publisher Site  Google Scholar
 Z. Shen, K. Toh, and S. Yun, “An accelerated proximal gradient algorithm for framebased image restoration via the balanced approach,” SIAM Journal on Imaging Sciences, vol. 4, no. 2, pp. 573–596, 2011. View at: Google Scholar
 H. Gao, H. Yu, S. Osher, and G. Wang, “Multienergy CT based on a prior rank, intensity and sparsity model (PRISM),” Inverse problems, vol. 27, no. 11, article 115012, 2011. View at: Publisher Site  Google Scholar
 J. F. Cai, S. Osher, and Z. Shen, “Split Bregman methods and frame based image restoration,” Multiscale Modeling and Simulation, vol. 8, no. 2, pp. 337–369, 2010. View at: Google Scholar
 M. Figueiredo and J. BiouscasDias, “Restoration of poissonian image using alternating direction optimization,” IEEE transactions on Image Processing, vol. 19, no. 12, pp. 3133–3145, 2010. View at: Publisher Site  Google Scholar
 E. Esser, “Applications of Lagrangianbased alternating direction methods and connections to split Bregman method,” Tech. Rep., Computational Applied Mathematics Tech. Rep, University of California, Los Angeles, 2009. View at: Google Scholar
 N. Pustelnik, C. Chaux, and J.C. Pesquet, “Parallel algorithm and hybrid regularization for image restoration using hybrid regularization,” IEEE transactions on Image Processing, vol. 20, no. 9, pp. 2450–2462, 2011. View at: Publisher Site  Google Scholar
 T. Jeong, H. Woo, and S. Yun, “Framebased poisson image restoration using a proximal linearized alternating direction method,” Inverse problems, vol. 29, no. 7, article 075007, 2013. View at: Google Scholar
 C. Wu and X. Tai, Augmented Lagrangian Method, Dual Methods, and Split Bregman Iteration for ROF, Vectorial TV, and High Order Models, UCLA, CAM Report, 2009.
 J. Eckstein and D. Bertsekas, “On the DouglasRachford splitting method and the proximal point algorithm for maximal monotone operators,” Mathematical Programming, vol. 55, no. 1–3, pp. 293–318, 1992. View at: Publisher Site  Google Scholar
 Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004. View at: Publisher Site  Google Scholar
 H. Gao, “Fast parallel algorithms for the xray transform and its adjoint,” Medical Physics, vol. 39, no. 11, pp. 7110–7120, 2012. View at: Publisher Site  Google Scholar
Copyright
Copyright © 2017 Weifeng Zhou and Hua Xiang. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.