 Original Article
 Open Access
 Published:
Prefilter that incorporates the noise model
Visual Computing for Industry, Biomedicine, and Art volume 3, Article number: 13 (2020)
Abstract
A linear denoising filter is usually of lowpass type, and the main parameter in a lowpass filter is the cutoff frequency. The lowpass filters are normally shift invariant and can be implemented as convolution in the spatial domain or as multiplication in the Fourier domain. This paper presents a linear filter that is not characterized by its cutoff frequency but is characterized by the noise model. An example of such a linear filter is presented for lowdose Xray computed tomography (CT).
Introduction
Nowadays, the iterative image reconstruction algorithms have become more popular than the analytic filtered backprojection (FBP) algorithm [1,2,3,4,5]. The main motivation of using an iterative algorithm is that it is much easier to incorporate the noise model into the iterative algorithm than into an analytic algorithm such as FBP [6, 7]. Many noise filtering techniques are available for the FBP algorithm [8,9,10,11]. In ref. [9], a nonlinear adaptive filter was proposed, in which a threshold value was assigned to determine whether the measurements were required to be filtered.
In refs. [11, 12], a linear FBP algorithm was proposed to mimic the iterative Landweber algorithm [13] and to incorporate the projection noise in a window function in the Fourier domain. The raybyray noiseweighting scheme is spatially variant. Its implementation was to decompose the spatially variant filter into 10 spatially invariant filters. The 10 versions of the filtered measurements were then combined into one and used in the backprojector.
Instead of mimicking an iterative algorithm to filter the noise, this paper takes a different approach and develops a convolution lookalike linear filter for the projection measurements. The filter being designed is shift variant and is based on the noise model. We treat each measurement as a random variable. By noise model, we mean the relationship between the measurement value and its standard deviation. For emission tomography measurements, the mean value is approximated by its measurement. The Poisson noise model is that the mean value and the variance are the same. For the transmission tomographic measurements, the logarithm has been taken. The variance of the lineintegral measurement is proportional to the exponential function of the lineintegral measurement.
Method
This section develops a nonstationary convolution lookalike filter for the lineintegral projection measurement data. Let p be the unfiltered lineintegral projection measurement and h be the spatialdomain filter kernel. In other words, p is the logarithm of the detected photon current from a computed tomography (CT) scanner bin at an angle. If h is a shiftinvariant convolution kernel, the conventional linear shiftinvariant filtering procedure can be expressed as a convolution integral below
where q(t,θ) is the filtered projection. In Eq. (1), the kernel h is shiftinvariant. In other words, the shiftinvariant filter blurs the projections p(t,θ) with the same kernel h everywhere. If the kernel h varies from location to location, Eq. (1) can be modified to
Equation (2) is no longer in the form of convolution. However, the computational complexity of Eq. (2) is almost the same as the complexity of Eq. (1), except that in Eq. (2) the kernel h must be evaluated differently for different locations (t,θ). In this paper, we assume the filter kernel h to be a two dimensional (2D) Gaussian function with a ‘standard deviation’ σ_{h}(t,θ). Our general strategy is to relate this σ_{h}(t,θ) with the noise standard deviation, σ_{p}(t,θ), of the noisy measurement p(t,θ).
For emission measurements, the projection p(t,θ) can be used to approximate the measurement noise variance according the Poisson distribution. Thus, \( {\sigma}_p\left(t,\theta \right)\approx \sqrt{p\left(t,\theta \right)} \).
For Transmission measurements, the noise variance of the postlog measurement p(t,θ) can be approximated as the exponential function of the lineintegral p(t,θ). Thus, \( {\sigma}_p\left(t,\theta \right)\approx \sqrt{\exp \left(p\left(t,\theta \right)\right)} \).
Our general strategy is to use a large kernel size of h if the corresponding σ_{p}(t,θ) value is large and to use a small kernel size if the corresponding σ_{p}(t,θ) value is small. The kernel size σ_{h}(t,θ) is thus a monotonic function of σ_{p}(t,θ). Here we propose this monotonic function to have the form of
For transmission tomography, Eq. (3) becomes
After combining the squareroot and parameter b into a single parameter b, Eq. (4) becomes
with user defined parameters a, b, and c.
As a special case of a = 0, h has a constant σ_{h}, the filter is then shiftinvariant, and the convolution lookalike expression Eq. (2) reduces to the true convolution Eq. (1).
Results
To illustrate the feasibility of the proposed linear nonstationary filter, a cadaver torso was scanned using an Xray CT scanner with a lowdose setting. The projections were first filtered by the proposed filter using Eq. (2) and then the conventional FBP algorithm was used to reconstruct the final image.
The cadaver data was collected with a diagnostic scanner (Aquilion ONE™, Toshiba America Medical Systems, Tustin, CA, USA; raw data courtesy of Leiden University Medical Center). The imaging geometry was conebeam, the Xray source trajectory was a circle of radius 600 mm. The detector had 320 rows, the rowheight was 0.5 mm, each row had 896 channels, and the fan angle was 49.2°. A lowdose noisy scan was carried out. The tube voltage was 120 kV and current was 60 mA. There were 1200 views uniformly sampled over 360°. Three patient slices were selected to test our proposed spatially variant filter.
Figure 1 shows the FBP reconstruction of three image slices from the lowdose CT data. It is observed that some horizontal streaking artifacts pass through both arms and the torso. The Xrays passing through both arms suffer from photon starvation and create large noise.
Figures 2, 3, 4, 5, 6 show the FBP reconstructions of three image slices from the prefiltered lowdose CT data. The prefilters were nonstationary, the associated kernel h was a Gaussian function, whose ‘standard deviation’ value σ_{h} was defined in Eq. (5). They shared the same values of a = 2.5 and c = − 2.49975. The b value was different, with b = 4, 3, 2, 1, and 0.5, respectively. It seems that a large b value helps removing the streaking artifacts.
Figures 7 and 8 show the FBP reconstructions from the prefiltered lowdose CT data. This time, the prefilters were stationary, that is, the kernel h was shiftinvariant by setting a = 0. The parameter c controls the degree of smoothing during filtering. In Figs. 7 and 8, c = 2 and 4, respectively. For a stationary filter, the entire image is smoothed at the same amount. It is observed that the stationary filters are not effective in removing streaking artifacts. The image is already blurry, while the streaking artifacts are still there.
Figure 9 shows 5 curves according to the relationship Eq. (5). The 5 curves correspond to the 5 cases with different b values as shown in Figs. 2, 3, 4, 5, 10, respectively. The horizontal axis represents the normalized projection value, and the vertical axis represents the size σ_{h} of the kernel h.
When the parameter b is small such as 0.5, the kernel size σ_{h} increases gently with the projection p(t,θ). When the parameter b is large such as 4, the kernel size σ_{h} increases dramatically with the projection p(t,θ). Intuitively speaking, the streaking artifacts are caused by a few projections with the largest values, whose rays pass through both arms. Using a large parameter b (e.g., b = 2) effectively separates the projections with large values from other projections. Only those large projection values are smoothed, while other smaller projection values are almost unchanged. The overall image resolution is unchanged.
The effectiveness of artifact removal of the proposed spatially variant filter can by assessed by visual inspection of the reconstructed images. It can also be evaluated by a numerical figureofmerit, called artifact index, which is empirically developed and explained as follows.
The artifact index compares a lowdose image against its corresponding regulardose image using the following steps. Step 1: Obtain a pair the reconstructed images, one with regulardose data and the other one with lowdose data. Step 2: Extract the edges (i.e., features) of these images using the Canny method. We used Matlab’s builtin function edge (Image, ‘Canny’) to accomplish this step. Step 3: Evaluate the difference image of these two edge images. Step 4: Evaluate the elementbyelement square of the error image. Step 5: Small isolated dotlike features are removed by using Matlab’s builtin function bwareaopen (Image, 70). Step 6: Find the artifact index as the total sum of the image values in processed image from Step 5. The results of the artifact index values for reconstructed images are summarized in Table 1. A smaller artifact index may imply fewer artifacts being detected from the lowdose image.
Conclusions
The proposed prefilter is designed for the FBP algorithm, aiming to reduce the photon starvation streaking artifacts. This is a linear filter as suggested by Eq. (2). It has a convolution lookalike form, with similar computational complexity as convolution.
The most important property of the proposed filter is its nonstationary characteristic, which allows the filter to smooth the data with specified range of values. As a result, some streaking artifacts can be reduced. Stationary filters do not have this ability.
The proposed spatially variant filter is a lowpass filter with a small kernel span. The low frequency components are almost not affected. However, at the same time of suppressing high frequency artifacts at certain views, some high frequency contrasts are suppressed as well. The low frequency image quantitation is almost not affected, but some high frequency details may get lost.
A mathematical analysis would conclude that the optimal value of parameter b in Eq. (5) should be 0.5 because it matches the noise model of the transmission tomography. However, computer simulations reveal that this is not the case and the b value should be larger than 0.5 to be effective. We still do not understand why it is more effective than the ‘correct’ value of b = 0.5. A previous paper published in 2016 [14] presented many examples to illustrate this unexplained phenomenon and concluded that the optimal parameter depends on the object and the image contrast. Therefore, there is no universal optimal parameter. The socalled ‘correct’ weighting function is suboptimal.
A lowdose CT study is used to test the feasibility of the proposed prefilter. The relationship between the kernel size and the noise variance is not uniquely determined. The model (5) is only one functional form that can be utilized. The user is encouraged to adopt other forms to label the undesired projection values to be smoothed out.
Availability of data and materials
Not applicable.
Abbreviations
 CT:

Computed tomography
 FBP:

Filtered backprojection
 2D:

Two dimensional
References
 1.
Geman S, McClure DE (1987) Statistical methods for tomographic image reconstruction. Bull Internat Statist Inst 52(Bk.4):5–21
 2.
Dempster AP, Laird NM, Rubin DB (1977) Maximum likelihood from incomplete data via the EM algorithm. J Roy Statist Soc Ser B (Methodol) 39(1):1–38. https://doi.org/https://doi.org/10.1111/j.25176161.1977.tb01600.x
 3.
Shepp LA, Vardi Y (1982) Maximum likelihood reconstruction for emission tomography. IEEE Trans Med Imag 1(2):113–122. https://doi.org/https://doi.org/10.1109/TMI.1982.4307558
 4.
Hudson HM, Larkin RS (1994) Accelerated image reconstruction using ordered subsets of projection data. IEEE Trans Med Imag 13(4):601–609. https://doi.org/https://doi.org/10.1109/42.363108
 5.
Lange K, Carson R (1984) EM reconstruction algorithms for emission and transmission tomography. J Comput Assist Tomogr 8(2):302–316
 6.
Zeng GL (2018) Technical note: emission expectation maximization lookalike algorithms for xray CT and other applications. Med Phys 45(8):3721–3727. https://doi.org/https://doi.org/10.1002/mp.13077
 7.
Zeng GL, Li Y (2019) Extension of emission expectation maximization lookalike algorithms to Bayesian algorithms. Vis Comput Ind, Biomed, Art 2(1):14. https://doi.org/https://doi.org/10.1186/s4249201900274
 8.
Hsieh J (1998) Adaptive streak artifact reduction in computed tomography resulting from excessive Xray photon noise. Med Phys 25(11):2139–2147. https://doi.org/https://doi.org/10.1118/1.598410
 9.
Kachelrieß M (2009)Branchless vectorized median filtering. Paper presented at 2009 IEEE nuclear science symposium conference record, IEEE, Orlando, pp 4099–4105. https://doi.org/https://doi.org/10.1109/NSSMIC.2009.5402362
 10.
Kachelrieß M, Watzke O, Kalender WA (2001) Generalized multidimensional adaptive filtering for conventional and spiral singleslice, multislice, and conebeam CT. Med Phys 28(4):475–490. https://doi.org/https://doi.org/10.1118/1.1358303
 11.
Zeng GL, Zamyatin A (2013) A filtered backprojection algorithm with raybyray noise weighting. Med Phys 40(3):031113. https://doi.org/https://doi.org/10.1118/1.4790696
 12.
Zeng GL (2017) Image reconstruction: applications in medical sciences. Walter de Gruyter GmbH, Berlin/Boston
 13.
Landweber L (1951) An iteration formula for Fredholm integral equations of the first kind. Am J Math 73(3):615–624. https://doi.org/https://doi.org/10.2307/2372313
 14.
Zeng GL, Wang WW (2016) Noise weighting with an exponent for transmission CT. Biomed Phys Eng Express 2(4):045004. https://doi.org/https://doi.org/10.1088/20571976/2/4/045004
Acknowledgments
The author thanks Raoul M.S. Joemai of Leiden University Medical Center for collecting and providing us raw data of the cadaver CT scan.
Funding
Not applicable.
Author information
Affiliations
Contributions
The author(s) read and approved the final manuscript.
Corresponding author
Ethics declarations
Competing interests
None of the authors have any competing interests in the manuscript.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Zeng, G.L. Prefilter that incorporates the noise model. Vis. Comput. Ind. Biomed. Art 3, 13 (2020). https://doi.org/10.1186/s4249202000051z
Received:
Accepted:
Published:
Keywords
 Lowdose computed tomography
 Denoising
 Linear filter
 Nonstationary filter