Counter examples for unmatched projector/backprojector in an iterative algorithm
 29 Downloads
Abstract
It is rather controversial whether it is justified to use an unmatched projector/backprojector pair in an iterative image reconstruction algorithm. One common concern of using an unmatched projector/backprojector pair is that the optimal solution cannot be reached. This concern is misleading and must be clarified. We define a figureofmerit in the image domain as the distance between the reconstructed image and the true image, as the normalized meansquarederror (NMSE). The NMSE is used to determine whether an unmatched matched projector/backprojector pair can provide a better image than a matched projector/backprojector pair. Hot and cold lesion’s contrasttonoise ratio is also used as an alternative secondary figureofmerit for algorithm comparison. Computergenerated counterexamples are used to test the performance for matched and unmatched projection/backprojection pairs for different reconstruction algorithms. The projectors are raydriven, and the backprojectors are raydriven and pixeldriven. For the attenuationfree data examples, the unmatched pixeldriven backprojector outperforms the matched raydriven backprojector. For the attenuated data example, the matched raydriven backprojector performs better. The raydriven backprojector can be slightly improved by using an attenuation coefficient that is larger than the true one; in this case the backprojector becomes unmatched. Unmatched projector/backprojector pairs are fairly flexible. If the backprojector is properly chosen, good results can be obtained. However, we have not found a general rule to select a good backprojector.
Keywords
Tomographic image reconstruction Iterative algorithms Projector/backprojector pairIntroduction
Every iterative medical image reconstruction algorithm consists of a projector/backprojector pair. A projector can be represented as a matrix A. A matched backprojector is the transpose matrix A^{T}. In other words, the matched backprojector is the adjoint operator of the projector. An example of matched projector/backprojector pair is the distancedriven pair [1, 2]. Almost all iterative algorithms are derived by assuming a matched projector/backprojector pair.
A projector should be chosen to mimic the physical data generation procedure as close as possible. In many applications, researchers may use a backprojector that is somewhat different from the matrix A^{T}, referring to it as an unmatched backprojector. People have various reasons to use an unmatched backprojector [3, 4, 5, 6, 7, 8, 9, 10]. An unmatched backprojector may simplify computer/GPU implementation, may reduce computational cost, may reduce image artifacts, may increase the algorithm convergence rate, and so on.
Good results can be obtained by using valid unmatched backprojectors. For example, if the ramp filter is used in the backprojector, the backprojector is valid, the algorithm is fast, but the resultant images are quite noisy [11, 12].

An invalid backprojector is sometimes useful in the early iterations.

Choosing a valid backprojector may not be a very critical factor in a practical image reconstruction problem.

A “converged” solution usually is very noisy and not desirable.

In practice, we should pay attention to the initial “convergent” trend and choose a rapid projector/backprojector pair, which may be an invalid pair, and then use regularization methods to guide or stop the iteration process.
The main goal of our 2000 paper was to establish the conditions, under which an unmatched projector/backprojector pair was valid. A valid projector/backprojector pair does not change the convergence property of the reconstruction algorithm. A valid pair may make the algorithm converge faster or slower. A naïve conception is that a valid pair with an unmatched backprojector does not give as good reconstruction as the matched backprojector can provide. The main goal of the current paper uses counter examples to point out that this naïve conception is not true. This point was not discussed in our 2000 paper.
In the paper by Kamphuis et al. an unmatched projector/backprojector pair was used to speed up the iterative image reconstruction, obtaining almost the same image quality [13]. In fact, the most common motivation of using an unmatched projector/backprojector pair in medical imaging is to significantly reduce the computation burden, even though the image quality is slightly compromised. Our paper is mainly concerned about the image quality.
Of course, if not carefully chosen, an unmatched projector/backprojector pair is most likely performs poorly and may introduce artifacts, as observed by Rahmim et al. in a motion tracking application [14].
Some researchers believe that the optimal solution can only be reached by matched projector/backprojector pairs [15, 16, 17, 18]. They also believe that the unmatched projector/backprojector pairs may give a faster algorithm, but the price to pay is the higher noise and image artifacts.
This paper argues that the common beliefs can be misleading, because the objective function measures the discrepancy between the forward projections and the measured data in the projection domain with some image domain constraints. The “optimal solution” should be defined in the image domain. The matched/unmatched projector/backprojector issues are still not well understood. We make no attempts to explain why a certain unmatched pair is preferable for a particular problem/algorithm choice. The investigation carried out in this paper is focused on single photon emission computed tomography (SPECT), even though the results can be applied to other modalities as well. In this paper, we use some counterexamples to disprove the statement that optimal solutions can only be achieved by matched projector/backprojector pairs.
Methods
Theoretical background
Iterative algorithms are used to reconstruct an image by minimizing an objective function (or, equivalently, maximizing a likelihood function). The iterative algorithm minimizes the objective function step by step. When the objective function is minimized, the conventional optimal solution is reached. The objective function is set up in the projection domain. Due to noise, this conventional optimal solution is noisy.
For example, the (MLEM maximum likelihood expectation maximization) algorithm is proven to increase the likelihood function at each iteration [19, 20]. The conventional optimal solution of the MLEM algorithm is the maximum likelihood solution. In practice, all projection data contain noise, and the optimal maximum likelihood solution is too noisy to be useful.
It is observed from Fig. 1 that there is an initial converging trend, a minimum discrepancy point, and a later diverging trend. The minimum discrepancy solution is not the maximum likelihood solution due to noise. It is beneficial to stop the MLEM iteration at the minimum discrepancy point and capture a less noisy image.
We do not yet have a relationship between the maximum likelihood solution and the minimum discrepancy solution. We argue that a projection/backprojection pair that leads to the maximum likelihood solution may not lead to the optimal minimum discrepancy solution. By optimal minimum discrepancy solution we mean that any other projector/backprojector pair cannot reach a smaller discrepancy. In this paper, the algorithm is fixed; the projector is fixed; only the backprojector is allowed to vary. Since the matched projector/backprojector pair is able to reach the maximum likelihood solution, our argument can be rephrased as follows: a matched backprojector may not be able to provide the minimum discrepancy solution. An unmatched backprojector may outperform the matched backprojector in the sense that an unmatched backprojector may give a smaller discrepancy at its minimum discrepancy point.
Computer simulations
A uniform twodimensional (2D) circular phantom with a diameter of 120.32 pixels was used in the computer simulations. The phantom, based on SPECT imaging, contained two smaller cold discs and two smaller hot discs, all with a diameter of 25.6 pixels. Let the image intensity of the larger circular disc be 1 unit. The cold discs had an intensity value of 0.5, and the hot discs had an intensity value of 1.5. Two situations were considered. In the first situation, the projections were attenuationfree. In the second situation, a uniform circular attenuator with a diameter of 120.32 pixels and attenuation coefficient of 0.05 per pixel was used in data generation and in image reconstruction. The projections were generated analytically without using pixels, to avoid committing the inverse crime. The inverse crime occurs when the same (or almost the same) theoretical ingredients are employed to synthesize as well as to invert data in an inverse problem.
There were 180 projection views over 360°. The detection array had 128 bins. The images were reconstructed in a 128 × 128 array. Poisson noise was incorporated in the projection data and three levels of noise were generated. Noise level 1 had 1,025,000 projection total photon counts, level 2 had 2,050,000 counts, and level 3 had 4,100,000 counts.
For the attenuationfree situation, the projector was the raydriven, linelength weighted. We tested two backprojectors: the matched raydriven, linelengthweighted backprojector and the unmatched pixeldriven backprojector that is normally used in the FBP (filtered backprojection) algorithm. Both the EM algorithm and the Green’s MAP–EM–TV algorithm were used for image reconstruction, with 100 iterations.
For the situation where the projections were attenuated, the projector was the raydriven, linelength and attenuation weighted. We tested three backprojectors: the matched raydriven, linelength and attenuationweighted backprojector, the unmatched raydriven, linelength and attenuationweighted backprojector with attenuation coefficient changed from 0.05 per pixel to 0.064 per pixel, and the unmatched pixeldriven backprojector without attenuation modeling. The EM algorithm was used for image reconstruction, with 100 iterations.
Evaluation
Results
Attenuationfree MLEM
Attenuationfree MAP–EM–TV
Attenuated MLEM
Lesion contrasttonoise ratio (CNR) curves
Discussion
According to the MLEM algorithm derivation, the backprojector matrix is the transpose of the projection matrix. Thus the maximum likelihood solution is achieved by a matched projector/backprojector pair. If an unmatched projector/backprojector pair is used, the solution of the MLEM algorithm may not result in the maximum likelihood solution.
When we compare with the true image, the maximum likelihood solution is noisy and is far away from the true image. The solution closest to the true image (we refer to it as the optimal solution or minimum discrepancy solution) is achieved with a low iteration number. This optimal solution may or may not be obtained by a matched projector/backprojector pair.
This paper uses some counterexamples to disprove the hypothesis that an optimal solution can only be achieved by matched projector/backproject pairs. The projectors in this paper are raydriven linelength weighted. For the attenuationless projection data, the unmatched pixeldriven backprojector gives better results than the matched raydriven backprojector. The contrasttonoise ratio (CNR) curves also show the better performance given by the pixeldriven backprojector. The results are consistent for all three noise levels.
However, we cannot conclude that the unmatched pixeldriven backprojector is always better than the matched raydriven backprojector. This point is illustrated by the examples with attenuated projection data. The matched raydriven backprojector models the attenuation effects, while the unmatched pixeldriven backprojector does not model the attenuation effects. As a result, the matched backprojector gives better results. This example does not imply that the matched backprojector is the best.
When projections are attenuated, we observe that the raydriven backprojector outperforms the pixeldriven projector that does not model the attenuation effects. However, if we use an unmatched raydriven backprojector by slightly increasing the value of attenuation coefficient (from 0.05 to 0.064), the performance is slightly improved.
The role of the projector is much more important than the backprojector. The projector ought to emulate the measurement physics accurately. The backprojector ought to be similar to the adjoint transform of the projector (i.e., the matched backprojector). Thus the projector is more relevant for image quantity. Slightly deviation of the backprojector from the matched backprojector may slightly alter the algorithm’s performance. We do not know yet how to find a better backprojector in general, because a good backprojector may be application dependent, for example, depending on the object to be imaged. This paper does not suggest that a pixeldriven backprojector may be better, or using a larger attenuation coefficient in the backprojector is better. From the attenuated data simulation results, we see some inconsistencies between the NMSE criterion and the CNR criterion. In clinical studies, there are no true images to compare with. One can use lesion contrasttonoise or detectability to evaluate the reconstruction. In situations where the true images are given (e.g., in computer simulations), mutual information can be used to compare the similarities between the reconstructed images with the true images. If the true image and the reconstructed image are perfectly registered, the most popular way to measure their distance is to use the Euclidean norm, which is essentially the NMSE. The NMSE criterion is our primary criterion.
Conclusion
Many theoretical results/properties are available for iterative algorithms at their convergence. For example, the noise sensitivity of the solution can be studied by the singular value decomposition and its associated condition number [24]. If the iterative algorithm stops at its minimum discrepancy point, we know almost nothing about the properties of the minimum discrepancy solution. In this paper, we use counterexamples to disprove the hypothesis that an optimal solution can only be achieved by matched projector/backprojector pairs. Unmatched projector/backprojector pairs are fairly flexible. If the backprojector is properly chosen, good results can be obtained. However, it is an open problem to find this minimum discrepancy point when the true solution is not known beforehand.
Once the reconstruction algorithm and the projector are chosen, our computer simulations suggest that the matched backprojector may not give the optimal minimum discrepancy solution and an unmatched backprojector may achieve the optimal minimum discrepancy solution. It is almost impossible to judge the performance when comparing two algorithms using real data, because the optimal stopping point for each algorithm is unknown. The value of this paper is the theoretical insight. At least we know that better results exist than the result that is provided by the matched projector/backprojector pair. How to find those better results is still an open problem.
Notes
Acknowledgements
Research reported in this publication was supported by the National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health under Award Number R15EB024283. The content is solely the responsibility of the author and does not necessarily represent the official views of the National Institutes of Health.
Compliance with ethical standards
Conflict of interest
Dr. Zeng reports a grant from the National Institute of Biomedical Imaging and Bioengineering (NIBIB) of the National Institutes of Health. This research was supported by the National Institute of Biomedical Imaging and Bioengineering (NIBIB) of the National Institutes of Health under award number R15EB024283.
References
 1.De Man B and S Basu S. Distance–driven projection and backprojection Conference. In: Record of the 2002 IEEE nuclear science symposium and medical imaging conference 2002;3:1477–1480. http://ieeexplore.ieee.org/document/1239600/.
 2.De Man B, Basu S. Distancedriven projection and backprojection in three dimensions. Phys Med Biol. 2004;49:2463–75.CrossRefGoogle Scholar
 3.Min BJ, Lee NY, Jung JH, Hong KJ, Hu W, Lee K, Ahn YB, Joung J. Unmatched projector/backprojector pair for demultiplexing in multipinhole emission computed tomography. Opt Eng. 2010;49:127004.CrossRefGoogle Scholar
 4.Xie L, Hu Y, Yan B, Wang L, Yang B, Liu W, et al. An effective CUDA parallelization of projection in iterative tomography reconstruction. PLos One. 2015;10:e0142184. https://doi.org/10.1371/journal.pone.0142184.CrossRefGoogle Scholar
 5.Zhang H, Wang L, Li L, Cai A, Hu G, Yan B. Iterative metal artifact reduction for xray computed tomography using unmatched projector/backprojector pairs. Med Phys. 2016;43:3019–33.CrossRefGoogle Scholar
 6.Harasse S, Yashiro W, Momose A. Iterative reconstruction in xray computed laminography from differential phase measurements. Opt Exp. 2011;19:16560–73.CrossRefGoogle Scholar
 7.Xu F and Mueller K A comparative study of popular interpolation and integration methods for use in computed tomography. In: The 3rd IEEE international symposium on biomedical imaging: Nano to Macro, Arlington, VA, 2006;1252–1255. https://doi.org/10.1109/isbi.2006.1625152.
 8.Peterson M, Gustafsson J, Ljungberg M. Monte Carlobased quantitative pinhole SPECT reconstruction using a raytracing backprojector. EJNMMI Phys. 2017;4:32. https://doi.org/10.1186/s406580170198z.CrossRefGoogle Scholar
 9.Glick SJ, Soares EJ. Noise characteristics of SPECT iterative reconstruction with a mismatched projector–backprojector pair. IEEE Trans Nucl Sci. 1998;45:2183–8.CrossRefGoogle Scholar
 10.Guedouar R, Zarrad BA. Comparative study between matched and mismatched projection/back projection pairs used with ASIRT reconstruction method. Nucl Instrum Methods Phys Res A. 2010;619:225–9.CrossRefGoogle Scholar
 11.Lee NY, Choi Y. Theoretical investigation on an unmatched backprojector for iterative reconstruction in emission computed tomography. J Korean Phys Soc. 2011;59:367–75.CrossRefGoogle Scholar
 12.Zeng GL, Gullberg GT. Unmatched projector/backprojector pairs in an iterative reconstruction algorithm. IEEE Trans Med Imaging. 2000;19:548–55.CrossRefGoogle Scholar
 13.Kamphuis C, Beekman FJ, Rijk PP, Viergever M. Dual matrix ordered subsets reconstruction for accelerated 3D scatter compensation in singlephoton emission tomography. Eur J Nucl Med. 1998;25:8–18.CrossRefGoogle Scholar
 14.Rahmim A, Cheng JC, Dinelle K, Shilov M, Segars WP, Rousset OG, Tsui BMW, Wong DF, Sossi V. System matrix modeling of externally tracked motion. Nucl Med Commun. 2008;29:574–81.CrossRefGoogle Scholar
 15.Dahmen T, Kohr H, Jonge N, Slusallek P. Matched backprojection operator for combined scanning transmission electron microscopy tilt and focal series. Microsc Microannal. 2015;21:725–38. https://doi.org/10.1017/s1431927615000525.CrossRefGoogle Scholar
 16.Chen JL, Li L, Wang LY, Cai AL, Xi XQ, Zhang HM, Li JX, Yan B. Fast parallel algorithm for threedimensional distancedriven model in iterative computed tomography reconstruction. Chin Phys B. 2015;24:028703. https://doi.org/10.1088/16741056/24/2/028703.CrossRefGoogle Scholar
 17.Qiao Z, Tai H, Song W, Li X. Investigation of unmatched projector/backprojector pairs in adaptivesteepestdescentprojectionontoconvexsets algorithm. J Med Imaging Health Inf. 2017;7:630–5.CrossRefGoogle Scholar
 18.Fessler JA. Statistical methods for tomographic image reconstruction. https://web.eecs.umich.edu/~fessler/papers/files/talk/00/00,01,06,seminar.pdf;2000.
 19.Shepp LA, Vardi Y. 1982 Maximum likelihood reconstruction for emission tomography. IEEE Trans Med Image. 1982;1:113–22. https://doi.org/10.1109/TMI.1982.4307558.CrossRefGoogle Scholar
 20.Lange K, Carson R. 1984 EM reconstruction algorithms for emission and transmission tomography. J Comp Ass Tomogr. 1984;8:302–16.Google Scholar
 21.On Zeng G L. Fewview tomography and staircase artifacts. IEEE Trans Nucl Sci. 2015;62:851–8.CrossRefGoogle Scholar
 22.Panin VY, Zeng GL, Gullberg GT. Total variation regulated EM algorithm. IEEE Trans Nucl Sci. 1999;46:2202–10.CrossRefGoogle Scholar
 23.Bayesian Green P J. reconstruction from emission tomography data using a modified EM algorithm. IEEE Trans Med Image. 1990;9:84–93.CrossRefGoogle Scholar
 24.Cheney E, Kincaid D. Numerical mathematics and computing. Boston: Cengage Learning; 2007.Google Scholar