[IEEE 2011 18th IEEE International Conference on Image Processing (ICIP 2011) - Brussels, Belgium (2011.09.11-2011.09.14)] 2011 18th IEEE International Conference on Image Processing - A comparative evaluation of ring artifacts reduction filters for X-ray computed microtomography images

A COMPARATIVE EVALUATION OF RING ARTIFACTS REDUCTION FILTERS FOR X-RAY COMPUTED MICROTOMOGRAPHY IMAGES F. Brun 1,2, G. Kourousias 2, D. Dreossi 2, L. Mancini 2, G. Tromba 2 1 Department of Industrial Engineering and Information Technology, University of Trieste, Italy 2 Sincrotrone Trieste S.C.p.A., Basovizza (Trieste), Italy ABSTRACT Concentric rings are among the most recurrent artifacts in X-ray computed microtomography imaging. Their presence have a negative impact on the quantitative and qualitative ana- lyses of the data. An interesting approach for ring artifacts re- moval is based on direct filtering of the reconstructed slices. Slice filtering does not require additional imaging data (i.e. flat fielding images), it does not exclude methods applied du- ring the reconstruction process (i.e. sinogram processing) and it can be applied also to past archived tomographic data. In this work four different approaches for the removal of com- plete ring artifacts are presented and preliminary compared using a suitably created artificial image as well as an experi- mental image. After a critical analysis of the considered fil- ters, the results support that an approach based on a modified Sijbers and Postnov [1] method effectively compensates ring artifacts in reconstructed slices, thus resulting in a powerful tool for established microtomography workflows. Index Terms— computed tomography, ring artifacts, stripe artifacts, filtering, image quality 1. INTRODUCTION X-Ray computed microtomography (µ-CT) images are often affected by a series of artifacts that seriously degrade image quality [2]. The most recurrent artifact is the presence of con- centric rings that arise from inhomogeneities in the individual pixel response of detector elements. Ring artifacts complicate the segmentation of the image and further analysis. There- fore, a significant reduction of these artifacts is essential. Ring artifacts in µ-CT images result from the back- projection of stripe artifacts in the sinogram image, therefore de-striping the sinogram is a classical way to face the pro- blem. However, if a reconstructed slice is transformed into polar coordinates assuming the center of rings as the center of the cartesian to polar conversion, the problem of ring arti- facts compensation can be still brought back to a de-striping issue (figure 1). Doing so, de-striping filters may be applied directly to transformed reconstructed slices avoiding the need of additional imaging data (i.e. flat fielding and sinogram images). The assessment of the efficacy of a filter is usually done through visual inspection of the output image. Such an ap- proach inherits the danger of sensory limitations and percep- tual errors [3] and therefore it is limited for scientific data, like those of a µ-CT scan, where the images are often processed as sources for the extraction of quantitative information [4]. Evaluation approaches that provide numerical quantitative as- sessments may improve comparison studies [5]. In the present study four of the most common de-striping approaches are revised and considered for de-ringing µ-CT images collected over a scan angle of 360◦. In addition to a qualitative comparison of an experimental µ-CT slice, a suita- ble artificial image was created and compared using common image quality indices as an attempt to perform a preliminary quantitative comparison. 2. METHODOLOGY The four de-striping filters considered in the present study are hereafter briefly described. Implementations with no more than two “tuning” parameters were taken into account. Du- ring the evaluation, subsequent filtering through iteration was examined. The iterations were performed in the polar do- main avoiding multiple conversions. All the filters include the same cartesian to polar and viceversa conversion with a dense bicubic interpolation scheme [6] in order to milder arti- facts far from image center [1]. The methodology adopted for the preparation of the artificial test image as well as a brief description of the quantitative measurement process for the comparison are also described. 2.1. Oimoen approach The filter proposed in [7] tries to isolate the artifacts from the useful signal by means of a one-dimensional high-pass filter orthogonal to the stripe artifacts. Then a low-pass filter paral- lel to the stripes is applied to the high-pass filtered image as an attempt to separate the actual stripes from common high frequency noise. The difference between the input image and this filtered image is the output of the whole process. In this study, an implementation in which median filtering of size NHP and NLP , respectively, was used for the intermediate 2011 18th IEEE International Conference on Image Processing 978-1-4577-1303-3/11/$26.00 ©2011 IEEE 405 steps. Variations of this approach have been proposed in [8] and [9] including an additional parameter for better isolating the stripes from the structure edges. A preliminary segmen- tation process is also suggested in [8]. However, with these further variations too many input parameters have to be spe- cified. 2.2. Boin and Haibel approach This approach is similar to the one first proposed by Rivers (cars9.uchicago.edu/software/idl/tomography.html). Initially, the sum of the gray values of the original image f is computed for each column i resulting in the y(i) signal. Then the low- pass filtered signal ys(i) is computed and used to correct each row j according to f ′j(i) = fj(i) × ys(i)/y(i). While a mo- ving average filter is proposed in [10], a median filter or even a more refined weighted median filter may be adopted [11]. Median filtering with size N was used in the implementation considered for this study. 2.3. Sijbers and Postnov approach Within a sliding window of user defined size W , a set of ho- mogeneous rows is detected. The homogeneity criterion is based on an user specified threshold value T . Working on this set an artifact template is generated and used for the correc- tion of each column of the image. A preliminary segmenta- tion process is proposed in the original version [12], however it has been shown in [1] that this should be considered as an optional step. 2.4. Münch et al. filter The idea to use transform domain processing such as the Fou- rier domain was explored for instance in [13]. The approach presented in [5] exploits both the wavelet and Fourier domain. At first, the original image is wavelet decomposed into L le- vels in order to separate the structural information into hori- zontal, vertical and diagonal bands at different resolution sca- les. Subsequently, the bands containing the stripe information are FFT transformed to further tighten the stripe information into narrow bands. This stripe information is then removed by using a Gaussian function with damping factor σ. Finally, the de-striped image is reconstructed from the filtered coeffi- cients. The implementation reported by the authors in the ori- ginal work [5] was used imposing Daubechies wavelet with order 32. 2.5. Test images and quantitative comparison The ground-truth image is composed by various shapes ap- pearing in an ISO 12233 pattern as well as other suitable pat- terns. Ring artifacts are produced from the tomographic re- construction of a so-called flat image (i.e. X-ray beam image without the sample). The resulting artificial test image is a combination of the ground-truth plus the reconstructed noise (figure 2). The experimental µ-CT test image (figure 1) is a human skull bone fragment acquired using a cone-beam de- vice (http://www.elettra.trieste.it/Labs/TOMOLAB) 1. The proposed quantitative comparison consists in the computation of three common indices used for assessing differences between a filtered image and the ground truth: Normalized Cross-Correlation (NCC), Peak Signal-to-Noise Ratio (PSNR) and Root Mean Squared Error (RMSE) [14]. Getting higher values for PSNR and NCC means that the fil- tered image is closer to the ground truth, while low values for RMSE imply better results. An open source implementation of the proposed indices was used (www.imagemagick.org). (a) (b) Fig. 1. X-ray µ-CT slice considered for the qualitative com- parison: a) original image; b) polar transformed image. (The scale bar is 5 mm). (a) (b) Fig. 2. Artificial test image considered for the quantitative comparison: a) ground truth image; b) rings corrupted image. 3. RESULTS AND DISCUSSION Figures 3 and 4 report the filtering effects for the test images. In all the cases the filter parameters were “tuned” according to a “trial and error” approach after visual supervision. 1High resolution test images as well as the evolution of the project can be found at: http://ulisse.elettra.trieste.it/uos/pore3d/ringrem 2011 18th IEEE International Conference on Image Processing 406 (a) (b) (c) (d) Fig. 3. Filter effects on the artificial test image: a) Oimoen’s approach (NHP = 21,NLP = 121) after 5 iterations; b) Boin and Haibel approach (N = 121) after 10 iterations; c) Sijbers and Postnov approach (W = 51, T = 130.0) after 20 iterations; d) Munch et al. algorithm (L = 8, σ = 2.5). It can be noticed that the modified Sijbers and Postnov approach outperforms the other filters, in particular for the artificial image. Impressive results were obtained and this is confirmed by the quantitative computed measures (table 1). The other filters are not able to totally remove the artifacts, though a slight compensation is visible. They present simi- lar values for NCC, RMSE and PSNR, making not worthy to perform a ranking based on the quantitative measures. Quali- tatively, it can be noticed that the Oimoen’s approach slightly damages some of the bar patterns of the artificial test image, while the other filters seem to be “safe” also for an unsuper- vised application. It is worthy to notice also that the Sijbers and Postnov approach as well as the Boin and Haibel were ap- plied with several iterations meaning that a stable behaviour may be supposed for these filters. When considering the experimental µ-CT test image, the Oimoen’s approach apparently produces a valuable result. However, observing the difference image, too many high frequency content has been removed in the output and this smoothing effect could be a concern in practical applications. The other approaches do not present this effect. Neverthe- less the considered implementation of the Boin and Haibel approach introduces new ring artifacts in the middle left part of the object, though valuable effects are recorded in the center of the image. For the modified Sijbers and Postnov and the Munch et al. algorithm the difference between the original image and the filtered one presents almost only rings, meaning that most of the useful signal is preserved in the output. Taking into account the effects on both the artificial and the real µ-CT image, it is reasonable to consider the modified Sijbers and Postnov approach [1] as an effective technique for ring artifacts compensation. A more comprehensive test with also a computational comparison is in progress as well as future evaluations of the efficacy of these filters in the case of µ-CT images reconstructed with other modalities (e.g 180◦ acquisition and extended field of view) that might be affected NCC PSNR RMSE Oimoen [7] 0.893 15.339 0.171 Boin and Haibel [10] 0.896 15.491 0.168 Sijbers and Postnov [1] 0.968 20.604 0.093 Munch et al. [5] 0.882 14.994 0.178 Table 1. Quantitative measures of the filtered image against the ground truth for each considered filter. by imperfect rings [5]. 4. CONCLUSION Ring artifacts degrade the quality of X-ray computed microto- mography images. While many filtering processes take place prior or during the tomographic reconstruction phase, an inte- resting class of filters that operate directly on the reconstruc- ted data exhibits impressive efficacy. In this study four diffe- rent de-ringing filters were presented and preliminary compa- red. The preliminary results support that an approach based on a modified Sijbers and Postnov [1] method can effectively filter the reconstructed slices. Since slice processing does not require additional imaging data (i.e. flat fielding and sino- gram images), this approach may result in a powerful tool for established microtomography workflows and for past archi- ved tomographic data. 5. REFERENCES [1] F. Brun, G. Kourousias, D. Dreossi, and L. Mancini, “An improved method for ring artifacts removing in recon- structed tomographic images,” IFMBE Proc., vol. 25, no. 4, pp. 926–929, 2009. [2] J.F. Barrett and N. Keat, “Artifacts in CT: Recogni- 2011 18th IEEE International Conference on Image Processing 407 (a) (b) (c) (d) Fig. 4. Filter effects on the real µ-CT test image: a) Oimoen’s approach (NHP = 51, NLP = 121) after 3 iterations; b) Boin and Haibel approach (N = 121) after 3 iterations; c) Sijbers and Postnov approach (W = 51, T = 300.0); d) Munch et al. algorithm (L = 4, σ = 1.5). (Rescaled difference between the original and the filtered image in the bottom row). tion and avoidance,” Radiographics, vol. 24, no. 6, pp. 1679–1691, 2004. [3] D. Purves, M.S. Williams, S. Nundy, and R.B. Lotto, “Perceiving the intensity of light,” Psychological Rev, vol. 111, no. 1, pp. 142–158, 2004. [4] F. Brun, L. Mancini, K. Parnian, S. Favretto, D. Dreossi, and G. Tromba, “Pore3D: A software library for quan- titative analysis of porous media,” Nucl. Instrum. Meth in Phys Res A, vol. 615, no. 3, pp. 326–332, 2010. [5] B. Münch, P. Trtik, F. Marone, and M. Stampa- noni, “Stripe and ring artifact removal with combined wavelet-Fourier filtering,” Opt. Express, vol. 17, no. 10, pp. 8567–8591, 2009. [6] M. Axelsson, S. Svensson, and G. Borgefors, “Reduc- tion of ring artifacts in high-resolution X-ray microto- mography images,” Lect. Notes Comput. Sci., vol. 4174, pp. 61–70, 2006. [7] M.J. Oimoen, “An effective filter for removal of produc- tion artifacts in U.S. geological survey 7.5-minute digi- tal elevation models,” in Proc. of the 14th Int. Conf. on Applied Geologic Remote Sensing, Las Vegas, Nevada, 6-8 November, 2000, pp. 311–319. [8] D. Prell, Y. Kyriakou, and W.A. Kalender, “Comparison of ring artifact correction methods for flat-detector CT,” Phys. Med. Biol., vol. 54, no. 12, pp. 3881–3895, 2009. [9] M.A. Yousuf and M. Asaduzzaman, “An efficient ring artifact reduction method based on projection data for micro-CT images,” J. Sci. Res., vol. 2, no. 1, pp. 37–45, 2010. [10] M. Boin and A. Haibel, “Compensation of ring artefacts in synchrotron tomographic images,” Opt. Express, vol. 14, no. 25, pp. 12071–12075, 2006. [11] F. Sadi, S.Y. Lee, and K. Hasan, “Removal of ring ar- tifacts in computed tomographic imaging using iterative center weighted median filter,” Comput. Biol. Med., vol. 40, no. 1, pp. 109–118, 2010. [12] J. Sijbers and A. Postnov, “Reduction of ring artifacts in high resolution micro-CT reconstructions,” Phys. Med. Biol., vol. 49, no. 14, pp. 247–253, 2004. [13] C. Raven, “Numerical removal of ring artifacts in mi- crotomography,” Rev. Sci. Instrum., vol. 69, no. 8, pp. 2978–2980, 2010. [14] A.M. Eskicioglu and P.S. Fisher, “Image quality measu- res and their performance,” IEEE T. Commun., vol. 43, no. 12, pp. 2959–2965, 1995. 2011 18th IEEE International Conference on Image Processing 408