- Email: [email protected]
An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study
Physica Medica xxx (2015) 1e9
Contents lists available at ScienceDirect
Physica Medica journal homepage: http://www.physicamedica.com
Original paper
An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study Nikolai V. Slavine a, *, Jeffrey Guild b, Roderick W. McColl b, Jon A. Anderson b, Orhan K. Oz c, Robert E. Lenkinski a a b c
Translational Research, Department of Radiology, UT Southwestern Medical Center, 5323 Harry Hines Boulevard, Dallas, TX 75390-9061, USA Clinical Medical Physics, Department of Radiology, UT Southwestern Medical Center, 5323 Harry Hines Boulevard, Dallas, TX 75390-9061, USA Nuclear Medicine, Department of Radiology, UT Southwestern Medical Center, 5323 Harry Hines Boulevard, Dallas, TX 75390-9061, USA
a r t i c l e i n f o
a b s t r a c t
Article history: Received 3 February 2015 Received in revised form 18 May 2015 Accepted 13 June 2015 Available online xxx
Purpose: To study the feasibility of using an iterative reconstruction algorithm to improve previously reconstructed CT images which are judged to be non-diagnostic on clinical review. A novel rapidly converging, iterative algorithm (RSEMD) to reduce noise as compared with standard filtered backprojection algorithm has been developed. Materials and methods: The RSEMD method was tested on in-silico, Catphan®500, and anthropomorphic 4D XCAT phantoms. The method was applied to noisy CT images previously reconstructed with FBP to determine improvements in SNR and CNR. To test the potential improvement in clinically relevant CT images, 4D XCAT phantom images were used to simulate a small, low contrast lesion placed in the liver. Results: In all of the phantom studies the images proved to have higher resolution and lower noise as compared with images reconstructed by conventional FBP. In general, the values of SNR and CNR reached a plateau at around 20 iterations with an improvement factor of about 1.5 for in noisy CT images. Improvements in lesion conspicuity after the application of RSEMD have also been demonstrated. The results obtained with the RSEMD method are in agreement with other iterative algorithms employed either in image space or with hybrid reconstruction algorithms. Conclusions: In this proof of concept work, a rapidly converging, iterative deconvolution algorithm with a novel resolution subsets-based approach that operates on DICOM CT images has been demonstrated. The RSEMD method can be applied to sub-optimal routine-dose clinical CT images to improve image quality to potentially diagnostically acceptable levels. © 2015 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.
Keywords: Resolution subsets-based iterative method Rapid converging CNR and SNR improvement Noisy CT image recovery
Introduction The most common method employed for image reconstruction is filtered back projection (FBP), even though early commercial CT scanners [1] used a variant of an iterative reconstruction (IR) algorithm. The FBP approach was adopted because of its vastly faster computational speed as compared to IR methods. This is a very fast and relatively robust algorithm for high-dose reconstruction, but there is a compromise between speed and image noise level for low
* Corresponding author. Division of Translational Research, Department of Radiology, UT Southwestern Medical Center, 5323 Harry Hines Boulevard, Dallas, Texas 75390-9071, USA. Tel.: þ1 214 648 9196; fax: þ1 214 648 7783. E-mail address: [email protected] (N.V. Slavine).
dose (<50 mAs) CT scans [2]. The practical limitation of FBP has become more apparent as dose levels drop for patient screening. During the last few years several research groups have been developing IR algorithms [3e13] as alternatives to FBP. These modern IR algorithms have become integral methods for image recovery because they can provide reduced image noise and lower artifacts as compared with FBP images acquired with lower radiation dose CT. These IR methods can be optimized for specific clinical needs [5,8]. Technological improvements in computational speed (high level of parallelism, fast processors) have reduced the image reconstruction times for IR to the point where the time required for image reconstruction approaches a clinically acceptable level of about 2 min per series (for example see Ref. [1]). Because these more sophisticated reconstruction methods can lead to both “noise reduction” as well as preservation of the critical features of the CT images, all of the major manufacturers [3,6,7,10] of modern CT
http://dx.doi.org/10.1016/j.ejmp.2015.06.009 1120-1797/© 2015 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
2
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
scanners offer IR algorithms on their platforms (for a recent review see Ref. [2]). IR algorithms can be classified in several ways [9e13] including whether they operate on projection data, images, or both, or whether they employ statistical or model based algorithms. The quality of images obtained using IR algorithms from four different CT commercial systems was recently classified and evaluated as provided in Ref. [3]. The image space IR algorithms start after an initial FBP (usually after a first-pass) image has been created [9,12,13], and improve image quality by better filtering noise than FBP does. The number of required iterations depends on both the desired image quality and the convergence speed of the chosen iterative algorithm. Iterative approaches that focus on image space have the added valuable benefits of availability and retrospective processing, and are still an active area of research. The advantages (reduced noise, repetitive reconstruction) and disadvantages (longer reconstruction time) of using IR methods to produce accurate high resolution CT images were recently described in the literature [1e11]. A number of these reports compare the performance of IR commercial algorithms both in phantoms [3,9], as well as clinical CT scans [8]. Different vendors propose and market their original versions of IR software with different functions and limitations [2,3]. In general, these methods all perform better than FBP in terms of noise reduction as well as their ability to preserve lesion detection. There have been several clinical evaluations of these IR methods with conventional FBP in studies using radiation dose reduction strategies (see examples in Ref. [2]). In general, these clinical evaluations have both an objective comparison of image noise and more subjective evaluations of image quality between FBP and the IR methods. As a rule these studies show the capacity of IR methods to reduce image noise and improve image quality at lower radiation doses when IR methods are compared with FBP. The immediate purpose of this study was to develop and evaluate a novel, fast converging iterative deconvolution method with the resolution subsets-based approach (RSEMD) for already binned data (image space). The RSEMD operates on Digital Imaging and Communications in Medicine (DICOM) images and can be employed retrospectively to de-noise and improve the quality of the sub-optimal clinical CT images previously reconstructed with the FBP method.
where s is the standard deviation parameter of Gaussian kernel rs ðrÞ ¼ rðr; sÞ and 5 - denotes convolution. The standard blind deblurring algorithm iterates the blurred clinical CT image value g CT ðrÞ with rsD ðrÞ for each step in the iteration, k
Materials and methods
f n1 þ1 ðrÞ ¼ f n1 ðrÞ
( f kþ1 ðrÞ ¼ f k ðrÞ f
nþ1
rsD ðrÞ5
where k ¼ 1,2 … n, f kþ1 ðrÞ and f k ðrÞ are the pixel values for the updated and current image estimates and f nþ1 ðrÞ is the desirable image estimate obtained after n iterations by post filtering f n ðrÞ with kernel rsP ðrÞ. Instead of performing the iterative procedure (2) with subsequent post filtering, we model the uncertainty caused in the system as an iterative deconvolution with resolution subsets (approximation of PSF- the point spread function) to de-noise and enhance image quality. This novel algorithm (RSEMD) iterates the blurred image with different resolution parameters s(to maximize SNR) and a corresponding number of iterations ns for each subset S(rs,ns) are taken in turn (Fig. 1). The resolution parameter s is changed but constant for each subset. In this case the total number of iterations after all of the resolution subsets are employed is much less than other methods (Fig. 2) which use one resolution parameter for multiple image updates passed through the data. The proposed deconvolution algorithm starts by iterating the CT ðrÞ with kernel r ðrÞ (initial subsetSðr ; n Þ) initial CT image gFBP 0 s0 s0 for each step in the iteration,n0
!) CT ðrÞ gFBP rs0 ðrÞ5 rs0 ðrÞ5f n0 ðrÞ
( f
n0 þ1
ðrÞ ¼ f
n0
ðrÞ
g CT ðrÞ ¼ rs ðrÞ5f ðrÞ þ nðrÞ þ aðrÞ
(1)
where n0
¼ 1; 2…N0 (3a) After N0 iterations when the SNR reaches a plateau, the algorithm continues to iterate the resulting image value associated with initial subset gsN00 ðrÞ with next kernel rs1 ðrÞ (current subset Sðrs1 ; n1 Þ) for each step in the iteration, n1
rs1 ðrÞ5
!) gsN00 ðrÞ rs1 ðrÞ5f n1 ðrÞ
where n1
¼ 1; 2…N1 (3b) and so on. Finally, to obtain the desired image quality (when the SNR Nfinal1 reaches a plateau), the algorithm iterates the image value gsfinal1 ðrÞ with the kernel rsfinal ðrÞ (final subset Sðrsfinal ; nfinal Þ) for each step in the iterations, nfinal
f
For the space-invariant case, restoration of a blurred CT image g CT ðrÞ with noise nðrÞ and artifacts aðrÞ can be approximated for each pixel position r as a convolution of a Gaussian function rs ðrÞ and the actual cross section (object function) in an image volumef ðrÞ
(2)
ðrÞ ¼ rsP ðrÞ5f ðrÞ
(
Image recovery method
rsD ðrÞ5f k ðrÞ n
(
The performance of RSEMD is described in in-silico phantoms, the Catphan®500 phantom [14], and an anthropomorphic 4D XCAT phantom [15]. Several conventional metrics widely used in the clinical community for describing the image quality of a CT scan are included for consideration. An additional metric related to the performance of the algorithm that was considered is the minimal number of iterations (convergence speed) to achieve optimal results. The algorithm was developed using MATLAB® and a LINUX platform with a multi-modality image handling environment.
!)
g CT ðrÞ
nfinal þ1
ðrÞ ¼ f
nfinal
N
final1 ðrÞ gsfinal1 ðrÞ rsfinal ðrÞ5 rsfinal ðrÞ5f nfinal ðrÞ
!) where
nfinal ¼ 1;2…Nfinal (3c) CT ðrÞ is the initial CT image after FBP reconstruction with where gFBP Nfinal1 all of the corrections included and gsN00 ðrÞ, gsN11 ðrÞ, …gsfinal1 ðrÞ, Nfinal gsfinal ðrÞ are current image updates after a blurring with different kernel parameters s0, s1, … sfinal1, sfinal and iteration numbers N0, N1, … Nfinal1,Nfinal for each subset S(rs,ns) respectively.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
3
Figure 1. Flowchart description of post-processing DICOM CT images with RSEMD iterative recovery method following standard clinical FBP reconstruction.
During the iteration procedure (3) the SNR is checked in each iterative step and this process can be repeated until the deblurring restoration reaches the highest SNR with D(SNR) x, where parameter x can be set as a small fraction of the initial SNRCT FBP . The second parameter is an initial deblurring parameter s0. The RSEMD algorithm converges more rapidly if the resolution width s0 is larger than estimated PSF while the average value of all deconvolution kernels width {s0, s1 …, sfinal1, sfinal} will approach to PSF
value. This hypothesis works well. For most practical cases, the total number of iterations (N0 þ N1 þ…þ Nfinal) for enhanced CT image quality is approximately 20 with a total number of resolution subsets around 5. To process CT data with an iterative procedure (3), a positive CT constant 2048 HU needs to be added to avoid computational problems (CT numbers in Hounsfield Units can be negative), and this value is subtracted after the final iteration from the deblurred image.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
4
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
210,105, 52 and 22 mAs. Slices were reconstructed with a reconstruction kernel FC64 and a pixel size of 0.468 mm (512 512 pixels). The contrast-to-noise ratio was calculated as: CNR ¼ mins mbg =sbg , where mins and mbg are the mean of the pixel values (HU) over a chosen ROI in an insert and in the background area; sbg is the standard deviation of pixel values in the same size ROI in the background (~100 mm2).
Results
Figure 2. A plot of the dependence of FWHM of the reconstructed 3D mathematical point-like phantom compared with the iteration number of conventional EM, EM with System Modeling (EMSM), Blind Deblurring (EMBD) and the RSEMD algorithm.
Application of the RSEMD algorithm reduced CT image noise and preserved important image features: no additional artifacts are observed. In general, values of both SNR and CNR reached a plateau after approximately 20 iterations with an improvement factor of about 1.5 in noisy CT images. The images generated after applying RSEMD showed both qualitative and quantitative improvements in lesion conspicuity compared to FBP alone. It was found that the accuracy of CT numbers, the linearity and uniformity of signal, high and low contrast resolution, lesion conspicuity as well as CNR and SNR, after the application of RSEMD are in good agreement with other IR algorithms employed in image space or with hybrid reconstruction algorithms.
The 4D XCAT phantom
Convergence speed
The extended 4D XCAT phantom was chosen to simulate abdominal CT scans because it's a whole body computer model of the complex shapes of real human organs based on Non-Uniform Rational B-Spline (NURBS) surfaces. Using this unique CT simulation tool, high resolution 3D/4D projection images can be simulated and reconstructed. The XCAT simulation tool was used to create realistic human CT images with the addition of different levels of signal dependent noise to simulate noisy CT data. In order to determine the metrics associated with lesion conspicuity, the XCAT phantom simulation tool was used to add an 18 mm low contrast (51 HU, mlesion ¼ 0.1616/cm) lesion in the liver of this phantom. The performance of RSEMD in improving lesion conspicuity was assessed at different levels of zero-mean Gaussian noise added with SNR ¼ 1.52, 2.18, 3.04, 4.02 and 8.05. The CNR of the lesion was calculated as: CNR ¼ mlesion mbg =sbg , where mlesion and mbg were the mean attenuation coefficients for same sized (~100 mm2) circular ROI's that were placed in the lesion and in the background area and sbg is the standard deviation of the background (noise). The FOV was subdivided into voxels of 0.3125 mm in size.
The convergence speed for different IR methods which operate on images by determining the variation of the image resolution with the number of iterations was determined. Figure 2 shows the image resolution parameter, FWHM (Full-Width at HalfMaximum), of the reconstructed 3D mathematical point-like phantom as a function of the iteration number for different iterative reconstruction methods [16,17] such as simple EM, EM with System Modeling, Blind Deblurring EM and RSEMD. As can be seen in Fig. 2, the value of the FWHM using RSEMD reaches a minimum after far fewer iterations with the best resolution result compared to the iterative reconstruction methods which use one resolution parameter for multiple image updates employed. The processing speed was improved by a factor of 4e5 as compared to other methods.
The Catphan®500 phantom This 20 cm diameter image quality phantom was employed because it includes multiple modules for the assessment of clinical CT image performance with respect to noise, spatial resolution, low contrast resolution (lesion conspicuity), CT numbers (HU), and their linearity. CT images of the Catphan®500 were acquired on a Toshiba Aquilion CT scanner (Toshiba, Tustin, CA) using parameters based on a clinical head protocol with FBP reconstructions using slice thicknesses of 0.5 and 5 mm. Images were acquired with a 240 mm field of view and a pitch of 0.828. Using a 1.5 s rotation time allowed all the images to be acquired with a small focal spot. The Volume Computed Tomography Dose Index (CTDIvol) values for the acquired image were 101.7, 52.2, 26.1, 13.1 and 5.6 mGy. The values of tube current-time, obtained by multiplying the exposure (rotating time) by the tube current, associated with these dose indices were 405,
Low and high contrast resolution Figure 3 quantitatively compares signal-to-noise (SNR) and contrast-to-noise (CNR) ratios for FBP and FBP þ RSEMD at different values of tube current-time for the low contrast resolution section of the Catphan®500 phantom. For very noisy images with low CNR, the application of RSEMD improves the CNR about 1.37 times, as compared to the FBP method. Figure 4c shows the improvement in spatial resolution and noise of the high contrast module for a noisy Toshiba CT scan at 52 mAs (Fig. 4a and b), when the RSEMD recovery method is applied after FBP. The visual inspection (arrows) of the bar patterns confirm the spatial resolution improvement using the RSEMD method up to 8-th bar pair. Figure 5 shows the profiles of the high contrast resolution section (arrows on Fig. 4) of the phantom after reconstruction with FBP (a) and with the RSEMD recovery method (b). The signal-to noise and spatial resolution improvement with the RSEMD method are quantitatively confirmed up to 8-th bar pair. No additional or significant artifacts are observed. Figure 4c looks brighter because of higher resolution and signal-to-noise (up to 60%) improvement (compare Fig. 5a and b) achieved as compared to FBP on Fig. 4b.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
5
Figure 3. A plot of the low contrast resolution section of the Catphan®500 phantom. Signal-to-noise (a) and contrast-to-noise (b) ratios compare standard FBP reconstruction with RSEMD recovery for deferent values of tube current-time (mAs).
CT number linearity
Noise reduction
The linearity of HU was initially estimated by determining the CT number for five different regions in the Catphan®500 phantom acquired with the standard dose (210 mAs and slice thickness of 0.5 mm) on the Toshiba CT scanner and reconstructed using the FBP algorithm. The metrics describing the accuracy of HU (average HU and SD) produced by the RSEMD algorithm, applied after FBP, are presented in Table 1 for comparison. After applying the RSEMD method, the CT numbers are very similar to those obtained with FBP and their standard deviation (SD) is lower than the results obtained with FBP. The standard deviation of the CT numbers between the central and peripheral ROI's (uniformity) was less than 4 HU in the results obtained after the application of RSEMD. The specific imaging characteristics for five different regions of the Catphan®500 phantom (see Table 1) are in excellent agreement with analogous results in Table 2 presented in Ref. [8] for the hybrid iDose algorithm and also with Iterative Reconstruction in Image Space algorithm (IRIS) [9].
In all of the phantom studies performed we found that the image noise was reduced after the application of RSEMD in all of the images where the zero-mean Gaussian noise was added. As an example, Fig. 6a shows the standard deviation (SD) of HU measured after FBP, and FBP þ RSEMD in the Teflon® (HU ¼ 990) region of the Catphan®500 uniformity section at different values of tube currenttime 22, 52, 105, 210 and 405 mAs. Figure 6b shows the ratio of this SD, demonstrating that the average improvement in SD of the HU after RSEMD is about 1.5. Figure 6c presents the contrast-to-noise ratio (CNR) for FBP and FBP þ RSEMD in the same region. For very noisy images with low CNR, the application of RSEMD improves the CNR about 1.48 times, as compared to the FBP method. Lesion conspicuity A qualitative improvement in lesion conspicuity is shown in Fig. 7. Zero-mean Gaussian noise (SNR is 3.04) was added to the
Figure 4. A plot of the high contrast resolution section of Catphan®500 phantom scanned on a Toshiba CT at 52 mAs: (a) and (b) with FBP reconstruction and (c) with RSEMD recovery method. The visual inspection (arrows) of the bar patterns (b) and (c) confirm the spatial resolution improvement using the RSEMD method up to 8-th bar pair (c).
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
6
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
Figure 5. Profiles of the high contrast resolution section (arrows on Figure 4) of the Catphan®500 phantom after reconstruction with FBP (a) and with the RSEMD recovery method (b). The signal-to noise and spatial resolution improvement with the RSEMD method quantitatively confirmed are up to 8-th bar pair (compare with Figure 4). No additional or significant artifacts are observed.
Table 1 Relative Hounsfield Units (HU) and standard deviation (SD) for five different regions of the Catphan®500 phantom. Material
FBP
Air Polyethylene Water Acrylic Teflon®
1008.3 95.3 4.9 123.6 995.2
RSEMD ± ± ± ± ±
10.2 12.2 3.7 9.9 9.3
1004.1 92.8 4.7 122.4 993.1
± ± ± ± ±
8.2 10.1 3.02 8.1 7.5
initial image generated by 4D XCAT. The image shown on Fig. 7a was obtained using FBP (SNR is 4.02 in the small low contrast lesion). The image shown in Fig. 7b was obtained after RSEMD was applied to the image shown in Fig. 7a (SNR is 5.82 in the liver lesion indicated by the arrow). Figure 8 quantitatively demonstrates the dependence of the signal-to-noise and contrast-to-noise ratios for the small liver lesion at different levels of SNR. The CNR and SNR values after RSEMD in the lesion reach a plateau after approximately 20 iterations with a total number of resolution subsets of around 5. Figure 9 shows the contrast-tonoise ratio versus number of iterations for a low contrast spherical liver lesion simulated by 4D XCAT with different levels of noise added (SNR ¼ 1.52, 2.18, 3.04, 4.02 and 8.05). The application of RSEMD after recovery by FBP improves the CNR about 1.45 times.
Figure 6. A plot of the linearity section of CT numbers in the Teflon® region of the Catphan®500 phantom. FBP reconstruction followed by the RSEMD recovery method versus current-time values at 22, 52, 105, 210 and 405 mAs: standard deviation (SD) associated with noise (a), ratio of SD before and after recovery with RSEMD (b) and contrast-to-noise ratio (CNR) for FBP and FBP þ RSEMD (c).
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
7
Figure 7. An XCAT phantom image obtained after FBP method (a) with SNR ¼ 4.02 and image obtained after 20 iterations with the RSEMD recovery method following FBP with improved SNR ¼ 5.82 (b). The visual inspection (arrow) confirms enhanced image quality in both SNR and CNR ratios.
Discussion In spite of the widespread availability of manufacturer provided IR methods [6,9e13], there still may be a need for additional fast IR methods that can be applied retrospectively to DICOM images which were previously reconstructed using the standard FBP methods. As an example, only about 20 percent of the CT scans performed in our county hospital are done on scanners equipped with IR algorithms. Also about 3e5% of the CT scans acquired at normal radiation dose are of marginal to non-diagnostic quality, usually, but not limited to patients with high body mass index (BMI). Desai et al. [6] compared the performance of ASIR (Adaptive Statistical Iterative Reconstruction, GE Healthcare) with FBP in 100 adults weighing more than 91 kg and found lower noise in the images at a reduced radiation dose in images reconstructed with ASIR. Kligerman et al. [7] compared FBP and iDose (Phillips Healthcare) [10] in terms of both image noise and image quality in obese patients (33 patients with an average BMI of 42.7) undergoing CT pulmonary angiography. They found an improvement in both image noise and contrast-to-noise with increasing levels of
iDose as compared to FBP. However, these added-on algorithms come with additional financial and time costs. The retrospective use of IR methods applied to projection data (which is usually stored in a proprietary format and not generally archived) to improve image quality becomes difficult, if not impossible. Several hours may pass between the acquisition of the CT scans and their clinical interpretation at which time the evaluation of their diagnostic quality is made. Artifacts (noise) are commonly encountered in clinical CT, can be problematic in medical imaging and may simulate pathology. Manufacturers use a variety of sophisticated techniques for artifact reduction (see for example existing commercial products iDose4 Premium Package [18] and Toshiba AIDR 3D [19]) in projection and imaging spaces to substitute the over range values in attenuation profiles. Noise in image space is greatly reduced using appropriate iterative reconstruction algorithms to rapidly post-process nondiagnostic CT scans. Currently, patients with non-diagnostic CT scans are often called back for a repeat CT examination, leading to increased radiation dose and delays in diagnosis and/or clinical management.
Figure 8. A plot of contrast-to-noise (a) and signal-to-noise (b) ratios compare the FBP method with FBP followed by RSEMD recovery on a low contrast, small, spherical liver lesion simulated by a 4D XCAT phantom.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
8
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
has shorter processing time than conventional algorithms. To reconstruct an image with higher resolution it is not necessary to spend time for extra iterations as is required for the deblurring algorithm (2) or any method which use one resolution parameter for multiple image updates passed through the data (conventional subsets). There is little systematic information in the literature about the dependence of the number of iterations in image quality [25]. An important next step will be to carry out a study on poor quality (noisy) clinical CT scans to determine what fraction of these scans can be “restored” by RSEMD to the point where they can be read clinically with confidence. It is also important to point out that the RSEMD code employed in this study was not optimized for computational speed (~300 msec per image update). It is expected that improvements will be made to reduce the processing time appreciably for typical clinical workflow. Conclusion
Figure 9. A plot of contrast-to-noise ratio versus the number of iterations for a low contrast spherical liver lesion simulated by 4D XCAT with different levels of noise added.
Note, it is not possible to determine the absolute differences in the existing spectrum of commercial reconstruction algorithms [3,7,10,18e22] due to the absence of detailed information about these algorithms in the literature. Thus, there is an unmet clinical need for rapid recovery algorithms in image space that are generalizable across multiple manufacturers' CT platforms, anatomic imaging sites, and patient populations to post-process CT scans that may be noisy and have low contrast. A mathematical approach to suppress both the noise and edge artifacts for the modified EM algorithm [23] was successfully adopted for CT image deblurring [12]. To estimate the parameters of the PSF, the blind deblurring algorithm similar to (2) maximizes the edge-to-noise ratio and deblurs the images with a relatively high number of iterations. According to Ref [12] this deblurring procedure (2) can produce a gain of up to 40% in image resolution. The permissible range of the blurring parameter sD can be estimated from the system modulation transfer function (MTF) and from data acquired in CT quality assurance tests. Unfortunately, the optimal number of iterations (about 100 in average) depends on image content and can only be estimated for a class of images through digital and experimental phantom studies. Incorporating a priori information about the actual characteristics of the blurred system can greatly improve some of the ill-posedness of this blind deblurring mechanism [24]. The use of the data subsets approach in the reconstruction procedure always offers an essential speed increase and better signal-to-noise characteristics. The conventional ordered subsets technique to partition the data into subsets can't be applied for clinical CT images since the system input is pixels not lines. To overcome all of these difficulties, the RSEMD method, incorporating an original deconvolution scheme in image space using resolution subsets (approximation of the PSF) to maximize SNR, was developed. The computations associated with each subset depend on the image update (intermediate image after one or more iterations) from the previous subset's iteration. When all of the resolution subsets are employed, an image recovery procedure has been performed. In this case an enhanced CT image recovery with RSEMD
A phantom study is a necessary first stage of the presented investigation for CT image enhancement. This is an important opportunity and one way to optimize diagnostic information for the older models of CT scanners where the modern versions of iterative reconstruction techniques may not be available. The application of this rapid technique to clinical imaging looks very promising for image resolution recovery, when applied to noisy clinical CT images, especially to images where noise is very problematic such as in ICU patients. Designing a study to evaluate the performance of the RSEMD as an image recovery method is the next step in determining the clinical utility of this approach. Conflicts of interest The authors have no conflicts of interest to declare. Acknowledgment The authors would like to thank Drs. Gary Arbique, Cecelia Brewington, Travis Browning, Julie Champine, Daniel Costa and the Clinically Appropriate and Accurate Radiation Exposure Committee of the Department of Radiology (UT Southwestern Medical Center at Dallas, TX) for their advice and collaboration. Dr. Robert E. Lenkinski acknowledges the support of the Missing Link Award from the Cancer Prevention Research Institute of Texas (CPRIT), 2011e2015. Dr. Nikolai Slavine would like to thank Dr. W. P. Segars (Duke University Medical Center, NC) for C-programming computer installation details of CT simulation tool 4D XCAT and Drs. Ananth Madhuranthakam and Zhongwei Zhang for valuable help with multiple DICOM CT files operation in MATLAB®. Special thanks to Jocelyn Chafouleas, Glenn Katz and Saleh Ramezani for technical assistance with the data presentation. References [1] Mehta D, Thompson R, Morton T, Dhanantwari A, Shefer ED. Iterative model reconstruction: simultaneously lowered computed tomography radiation dose and improved image quality. Med Phys Int J 2013;1(2):147e55. [2] Kordolaimi SD, Argentos S, Pantos I, Kelekis NL, Efstathopoulos EP. A new era in computed tomographic dose optimization: the impact of iterative reconstruction on image quality and radiation dose. J Comput Assist Tomogr 2013;37(6):924e31. €ve A, Olsson ML, Siemund R, Stålhammar F, Bjo €rkman-Burtscher IM, [3] Lo € derberg M. Six iterative reconstruction algorithms in brain CT: a phantom So study on image quality at different radiation dose levels. Br J Radiol 2013;86(1031):20130388. [4] Pan X, Sidky EY, Vannier M. Why do commercial CT scanners still employ traditional, filtered back projection for image reconstruction? Inverse Probl 2009;25:1e36.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9 €l BP, Fingerle AA, Renger B, Münzel D, Rummeny EJ, Dobritz M. Initial [5] Noe performance characterization of a clinical noise e suppressing reconstruction algorithm for MDCT. Am J Roentgenol 2011;197:1404e9. [6] Desai GS, Uppot RN, Yu EW, Kambadakone AR, Sahani DV. Impact of iterative reconstruction on image quality and radiation dose in multidetector CT of large body size adults. Eur Radiol 2012;22(8):1631e40. [7] Kligerman S, Mehta D, Farnadesh M, Jeudy J, Olsen K, White C. Use of a hybrid iterative reconstruction technique to reduce image noise and improve image quality in obese patients undergoing computed tomographic pulmonary angiography. J Thorac Imaging 2013;28(1):49e59. [8] Noel P, Fingerle A, Renger B, Rummeny EJ, Dobritz M. A clinical comparison study of a novel statistical iterative and filtered back-projection reconstruction. In: Pelc J, Ehsan S, Nishikawa R, editors. Medical imaging: physics of medical imaging. Bellingham: SPIE; 2011. [9] Ghetti C, Ortenzia O, Serreli G. CT iterative reconstruction in image space: a phantom study. Phys Med 2012;28:161e5. [10] Scibelli A. iDose4 iterative reconstruction technique (Philips Healthcare Web site), www.clinical.netforum.healthcare.philips.com/global/Explore/WhitePapers/CT/iDose4-iterative-reconstruction-technique [accessed 11.03.11]. [11] Pickhardt PJ, Lubner MG, Kim DH, Tang J, Ruma JA, del Rio AM, et al. Abdominal CT with Model-Based Iterative Reconstruction (MBIR): initial results of a prospective trial comparing ultra-low dose with standard-dose imaging. Am J Roentgenol 2012;299(6):1266e74. [12] Ge Wang, Vannier MW, Skinner MW, Cavalcanti M, Harding G. Spiral CT image deblurring for cochlear implantation. IEEE Trans Med Imag 1998;17:251e62. [13] Tipnis S, Ramachandra A, Huda W, Hardie A, Schoepf J, Costello P, et al. Iterative reconstruction in image space (IRIS) and lesion detection in abdominal CT. 2010. SPIE Press, Med, Imaging; Bellingham, WA, Proc SPIE 7622. [14] Catphan®500 and 600, manual. Cambridge, NY: The Phantom Laboratory; 2012. p. 1e40. [15] Segars WP, Mahesh M, Beck TJ, Frey EC, Tsui BM. Realistic CT simulation using the 4D XCAT phantom. Med Phys 2008;35(8):3800e8. http://dmip1.rad.jhmi. edu/xcat/.
9
[16] Antich P, Parkey R, Seliounine S, Slavine NV, Tsyganov E, Zinchenko A. Application of expectation maximization algorithms for image resolution improvement in a small animal PET system. IEEE Trans Nucl Sci 2005;52(3): 684e90. [17] E. Tsyganov, A. Zinchenko, N.V. Slavine, P. Antich, S. Seliounine, O. Oz, et al. Reconstruction algorithm with resolution deconvolution for 3D image in a small animal PET imager. Springer Science Media Inc, 163e176, [Chapter 8]. [18] IDose4 premium package, © Koninklijke Philips N.V. 2014. http://www. healthcare.philips.com/main/products/ct/products/idose4_premium_ package/. [19] AIDR 3D, Toshiba Medical System Corporation. 2015. http://www. toshibamedicalsystems.com/tmd/english/products/dose/lowdose/ reconstruction.html. [20] Silva AC, Lawder HJ, Hara A, Kujak J, Pavlicek W. Innovations in CT dose reduction strategy: application of the adaptive statistical iterative reconstruction algorithm. Am J Roentgenol 2010;194:191e9. [21] Matsuki M, Murakami T, Juri H, Yoshikawa S, Narumi Y. Impact of adaptive iterative dose reduction (AIDR) 3D on low-dose abdominal CT: comparison with routine-dose CT using filtered back projection. Acta Radiol 2013;54: 869e75. [22] Price RG, Vance S, Cattaneo R, Schulz L, Elsheikh MA, Chetty IJ, et al. Characterization of a commercial hybrid iterative and model-based reconstruction algorithm in radiation oncology. Med Phys 2014;41:081907. [23] Snyder DL, Miller M, Thomas L, Politte DG. Noise and edge artifacts in maximum-likelihood reconstructions for emission tomography. IEEE Trans Med Imaging 1987;6:228e38. [24] Kudo H, Courdurier M, Noo F, Defrise M. Tiny a priori knowledge solves the interior problem in computed tomography. Phys Med Biol 2008;53: 2207e31. [25] Pontana F, Duhamel A, Pagniez J, Flohr T, Faivre JB, Hachulla AL, et al. Chest computed tomography using iterative reconstruction vs FBP (part 2) image quality of low dose CT examinations in 80 patients. Eur Radiol 2010;21: 636e43.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
Contents lists available at ScienceDirect
Physica Medica journal homepage: http://www.physicamedica.com
Original paper
An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study Nikolai V. Slavine a, *, Jeffrey Guild b, Roderick W. McColl b, Jon A. Anderson b, Orhan K. Oz c, Robert E. Lenkinski a a b c
Translational Research, Department of Radiology, UT Southwestern Medical Center, 5323 Harry Hines Boulevard, Dallas, TX 75390-9061, USA Clinical Medical Physics, Department of Radiology, UT Southwestern Medical Center, 5323 Harry Hines Boulevard, Dallas, TX 75390-9061, USA Nuclear Medicine, Department of Radiology, UT Southwestern Medical Center, 5323 Harry Hines Boulevard, Dallas, TX 75390-9061, USA
a r t i c l e i n f o
a b s t r a c t
Article history: Received 3 February 2015 Received in revised form 18 May 2015 Accepted 13 June 2015 Available online xxx
Purpose: To study the feasibility of using an iterative reconstruction algorithm to improve previously reconstructed CT images which are judged to be non-diagnostic on clinical review. A novel rapidly converging, iterative algorithm (RSEMD) to reduce noise as compared with standard filtered backprojection algorithm has been developed. Materials and methods: The RSEMD method was tested on in-silico, Catphan®500, and anthropomorphic 4D XCAT phantoms. The method was applied to noisy CT images previously reconstructed with FBP to determine improvements in SNR and CNR. To test the potential improvement in clinically relevant CT images, 4D XCAT phantom images were used to simulate a small, low contrast lesion placed in the liver. Results: In all of the phantom studies the images proved to have higher resolution and lower noise as compared with images reconstructed by conventional FBP. In general, the values of SNR and CNR reached a plateau at around 20 iterations with an improvement factor of about 1.5 for in noisy CT images. Improvements in lesion conspicuity after the application of RSEMD have also been demonstrated. The results obtained with the RSEMD method are in agreement with other iterative algorithms employed either in image space or with hybrid reconstruction algorithms. Conclusions: In this proof of concept work, a rapidly converging, iterative deconvolution algorithm with a novel resolution subsets-based approach that operates on DICOM CT images has been demonstrated. The RSEMD method can be applied to sub-optimal routine-dose clinical CT images to improve image quality to potentially diagnostically acceptable levels. © 2015 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.
Keywords: Resolution subsets-based iterative method Rapid converging CNR and SNR improvement Noisy CT image recovery
Introduction The most common method employed for image reconstruction is filtered back projection (FBP), even though early commercial CT scanners [1] used a variant of an iterative reconstruction (IR) algorithm. The FBP approach was adopted because of its vastly faster computational speed as compared to IR methods. This is a very fast and relatively robust algorithm for high-dose reconstruction, but there is a compromise between speed and image noise level for low
* Corresponding author. Division of Translational Research, Department of Radiology, UT Southwestern Medical Center, 5323 Harry Hines Boulevard, Dallas, Texas 75390-9071, USA. Tel.: þ1 214 648 9196; fax: þ1 214 648 7783. E-mail address: [email protected] (N.V. Slavine).
dose (<50 mAs) CT scans [2]. The practical limitation of FBP has become more apparent as dose levels drop for patient screening. During the last few years several research groups have been developing IR algorithms [3e13] as alternatives to FBP. These modern IR algorithms have become integral methods for image recovery because they can provide reduced image noise and lower artifacts as compared with FBP images acquired with lower radiation dose CT. These IR methods can be optimized for specific clinical needs [5,8]. Technological improvements in computational speed (high level of parallelism, fast processors) have reduced the image reconstruction times for IR to the point where the time required for image reconstruction approaches a clinically acceptable level of about 2 min per series (for example see Ref. [1]). Because these more sophisticated reconstruction methods can lead to both “noise reduction” as well as preservation of the critical features of the CT images, all of the major manufacturers [3,6,7,10] of modern CT
http://dx.doi.org/10.1016/j.ejmp.2015.06.009 1120-1797/© 2015 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
2
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
scanners offer IR algorithms on their platforms (for a recent review see Ref. [2]). IR algorithms can be classified in several ways [9e13] including whether they operate on projection data, images, or both, or whether they employ statistical or model based algorithms. The quality of images obtained using IR algorithms from four different CT commercial systems was recently classified and evaluated as provided in Ref. [3]. The image space IR algorithms start after an initial FBP (usually after a first-pass) image has been created [9,12,13], and improve image quality by better filtering noise than FBP does. The number of required iterations depends on both the desired image quality and the convergence speed of the chosen iterative algorithm. Iterative approaches that focus on image space have the added valuable benefits of availability and retrospective processing, and are still an active area of research. The advantages (reduced noise, repetitive reconstruction) and disadvantages (longer reconstruction time) of using IR methods to produce accurate high resolution CT images were recently described in the literature [1e11]. A number of these reports compare the performance of IR commercial algorithms both in phantoms [3,9], as well as clinical CT scans [8]. Different vendors propose and market their original versions of IR software with different functions and limitations [2,3]. In general, these methods all perform better than FBP in terms of noise reduction as well as their ability to preserve lesion detection. There have been several clinical evaluations of these IR methods with conventional FBP in studies using radiation dose reduction strategies (see examples in Ref. [2]). In general, these clinical evaluations have both an objective comparison of image noise and more subjective evaluations of image quality between FBP and the IR methods. As a rule these studies show the capacity of IR methods to reduce image noise and improve image quality at lower radiation doses when IR methods are compared with FBP. The immediate purpose of this study was to develop and evaluate a novel, fast converging iterative deconvolution method with the resolution subsets-based approach (RSEMD) for already binned data (image space). The RSEMD operates on Digital Imaging and Communications in Medicine (DICOM) images and can be employed retrospectively to de-noise and improve the quality of the sub-optimal clinical CT images previously reconstructed with the FBP method.
where s is the standard deviation parameter of Gaussian kernel rs ðrÞ ¼ rðr; sÞ and 5 - denotes convolution. The standard blind deblurring algorithm iterates the blurred clinical CT image value g CT ðrÞ with rsD ðrÞ for each step in the iteration, k
Materials and methods
f n1 þ1 ðrÞ ¼ f n1 ðrÞ
( f kþ1 ðrÞ ¼ f k ðrÞ f
nþ1
rsD ðrÞ5
where k ¼ 1,2 … n, f kþ1 ðrÞ and f k ðrÞ are the pixel values for the updated and current image estimates and f nþ1 ðrÞ is the desirable image estimate obtained after n iterations by post filtering f n ðrÞ with kernel rsP ðrÞ. Instead of performing the iterative procedure (2) with subsequent post filtering, we model the uncertainty caused in the system as an iterative deconvolution with resolution subsets (approximation of PSF- the point spread function) to de-noise and enhance image quality. This novel algorithm (RSEMD) iterates the blurred image with different resolution parameters s(to maximize SNR) and a corresponding number of iterations ns for each subset S(rs,ns) are taken in turn (Fig. 1). The resolution parameter s is changed but constant for each subset. In this case the total number of iterations after all of the resolution subsets are employed is much less than other methods (Fig. 2) which use one resolution parameter for multiple image updates passed through the data. The proposed deconvolution algorithm starts by iterating the CT ðrÞ with kernel r ðrÞ (initial subsetSðr ; n Þ) initial CT image gFBP 0 s0 s0 for each step in the iteration,n0
!) CT ðrÞ gFBP rs0 ðrÞ5 rs0 ðrÞ5f n0 ðrÞ
( f
n0 þ1
ðrÞ ¼ f
n0
ðrÞ
g CT ðrÞ ¼ rs ðrÞ5f ðrÞ þ nðrÞ þ aðrÞ
(1)
where n0
¼ 1; 2…N0 (3a) After N0 iterations when the SNR reaches a plateau, the algorithm continues to iterate the resulting image value associated with initial subset gsN00 ðrÞ with next kernel rs1 ðrÞ (current subset Sðrs1 ; n1 Þ) for each step in the iteration, n1
rs1 ðrÞ5
!) gsN00 ðrÞ rs1 ðrÞ5f n1 ðrÞ
where n1
¼ 1; 2…N1 (3b) and so on. Finally, to obtain the desired image quality (when the SNR Nfinal1 reaches a plateau), the algorithm iterates the image value gsfinal1 ðrÞ with the kernel rsfinal ðrÞ (final subset Sðrsfinal ; nfinal Þ) for each step in the iterations, nfinal
f
For the space-invariant case, restoration of a blurred CT image g CT ðrÞ with noise nðrÞ and artifacts aðrÞ can be approximated for each pixel position r as a convolution of a Gaussian function rs ðrÞ and the actual cross section (object function) in an image volumef ðrÞ
(2)
ðrÞ ¼ rsP ðrÞ5f ðrÞ
(
Image recovery method
rsD ðrÞ5f k ðrÞ n
(
The performance of RSEMD is described in in-silico phantoms, the Catphan®500 phantom [14], and an anthropomorphic 4D XCAT phantom [15]. Several conventional metrics widely used in the clinical community for describing the image quality of a CT scan are included for consideration. An additional metric related to the performance of the algorithm that was considered is the minimal number of iterations (convergence speed) to achieve optimal results. The algorithm was developed using MATLAB® and a LINUX platform with a multi-modality image handling environment.
!)
g CT ðrÞ
nfinal þ1
ðrÞ ¼ f
nfinal
N
final1 ðrÞ gsfinal1 ðrÞ rsfinal ðrÞ5 rsfinal ðrÞ5f nfinal ðrÞ
!) where
nfinal ¼ 1;2…Nfinal (3c) CT ðrÞ is the initial CT image after FBP reconstruction with where gFBP Nfinal1 all of the corrections included and gsN00 ðrÞ, gsN11 ðrÞ, …gsfinal1 ðrÞ, Nfinal gsfinal ðrÞ are current image updates after a blurring with different kernel parameters s0, s1, … sfinal1, sfinal and iteration numbers N0, N1, … Nfinal1,Nfinal for each subset S(rs,ns) respectively.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
3
Figure 1. Flowchart description of post-processing DICOM CT images with RSEMD iterative recovery method following standard clinical FBP reconstruction.
During the iteration procedure (3) the SNR is checked in each iterative step and this process can be repeated until the deblurring restoration reaches the highest SNR with D(SNR) x, where parameter x can be set as a small fraction of the initial SNRCT FBP . The second parameter is an initial deblurring parameter s0. The RSEMD algorithm converges more rapidly if the resolution width s0 is larger than estimated PSF while the average value of all deconvolution kernels width {s0, s1 …, sfinal1, sfinal} will approach to PSF
value. This hypothesis works well. For most practical cases, the total number of iterations (N0 þ N1 þ…þ Nfinal) for enhanced CT image quality is approximately 20 with a total number of resolution subsets around 5. To process CT data with an iterative procedure (3), a positive CT constant 2048 HU needs to be added to avoid computational problems (CT numbers in Hounsfield Units can be negative), and this value is subtracted after the final iteration from the deblurred image.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
4
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
210,105, 52 and 22 mAs. Slices were reconstructed with a reconstruction kernel FC64 and a pixel size of 0.468 mm (512 512 pixels). The contrast-to-noise ratio was calculated as: CNR ¼ mins mbg =sbg , where mins and mbg are the mean of the pixel values (HU) over a chosen ROI in an insert and in the background area; sbg is the standard deviation of pixel values in the same size ROI in the background (~100 mm2).
Results
Figure 2. A plot of the dependence of FWHM of the reconstructed 3D mathematical point-like phantom compared with the iteration number of conventional EM, EM with System Modeling (EMSM), Blind Deblurring (EMBD) and the RSEMD algorithm.
Application of the RSEMD algorithm reduced CT image noise and preserved important image features: no additional artifacts are observed. In general, values of both SNR and CNR reached a plateau after approximately 20 iterations with an improvement factor of about 1.5 in noisy CT images. The images generated after applying RSEMD showed both qualitative and quantitative improvements in lesion conspicuity compared to FBP alone. It was found that the accuracy of CT numbers, the linearity and uniformity of signal, high and low contrast resolution, lesion conspicuity as well as CNR and SNR, after the application of RSEMD are in good agreement with other IR algorithms employed in image space or with hybrid reconstruction algorithms.
The 4D XCAT phantom
Convergence speed
The extended 4D XCAT phantom was chosen to simulate abdominal CT scans because it's a whole body computer model of the complex shapes of real human organs based on Non-Uniform Rational B-Spline (NURBS) surfaces. Using this unique CT simulation tool, high resolution 3D/4D projection images can be simulated and reconstructed. The XCAT simulation tool was used to create realistic human CT images with the addition of different levels of signal dependent noise to simulate noisy CT data. In order to determine the metrics associated with lesion conspicuity, the XCAT phantom simulation tool was used to add an 18 mm low contrast (51 HU, mlesion ¼ 0.1616/cm) lesion in the liver of this phantom. The performance of RSEMD in improving lesion conspicuity was assessed at different levels of zero-mean Gaussian noise added with SNR ¼ 1.52, 2.18, 3.04, 4.02 and 8.05. The CNR of the lesion was calculated as: CNR ¼ mlesion mbg =sbg , where mlesion and mbg were the mean attenuation coefficients for same sized (~100 mm2) circular ROI's that were placed in the lesion and in the background area and sbg is the standard deviation of the background (noise). The FOV was subdivided into voxels of 0.3125 mm in size.
The convergence speed for different IR methods which operate on images by determining the variation of the image resolution with the number of iterations was determined. Figure 2 shows the image resolution parameter, FWHM (Full-Width at HalfMaximum), of the reconstructed 3D mathematical point-like phantom as a function of the iteration number for different iterative reconstruction methods [16,17] such as simple EM, EM with System Modeling, Blind Deblurring EM and RSEMD. As can be seen in Fig. 2, the value of the FWHM using RSEMD reaches a minimum after far fewer iterations with the best resolution result compared to the iterative reconstruction methods which use one resolution parameter for multiple image updates employed. The processing speed was improved by a factor of 4e5 as compared to other methods.
The Catphan®500 phantom This 20 cm diameter image quality phantom was employed because it includes multiple modules for the assessment of clinical CT image performance with respect to noise, spatial resolution, low contrast resolution (lesion conspicuity), CT numbers (HU), and their linearity. CT images of the Catphan®500 were acquired on a Toshiba Aquilion CT scanner (Toshiba, Tustin, CA) using parameters based on a clinical head protocol with FBP reconstructions using slice thicknesses of 0.5 and 5 mm. Images were acquired with a 240 mm field of view and a pitch of 0.828. Using a 1.5 s rotation time allowed all the images to be acquired with a small focal spot. The Volume Computed Tomography Dose Index (CTDIvol) values for the acquired image were 101.7, 52.2, 26.1, 13.1 and 5.6 mGy. The values of tube current-time, obtained by multiplying the exposure (rotating time) by the tube current, associated with these dose indices were 405,
Low and high contrast resolution Figure 3 quantitatively compares signal-to-noise (SNR) and contrast-to-noise (CNR) ratios for FBP and FBP þ RSEMD at different values of tube current-time for the low contrast resolution section of the Catphan®500 phantom. For very noisy images with low CNR, the application of RSEMD improves the CNR about 1.37 times, as compared to the FBP method. Figure 4c shows the improvement in spatial resolution and noise of the high contrast module for a noisy Toshiba CT scan at 52 mAs (Fig. 4a and b), when the RSEMD recovery method is applied after FBP. The visual inspection (arrows) of the bar patterns confirm the spatial resolution improvement using the RSEMD method up to 8-th bar pair. Figure 5 shows the profiles of the high contrast resolution section (arrows on Fig. 4) of the phantom after reconstruction with FBP (a) and with the RSEMD recovery method (b). The signal-to noise and spatial resolution improvement with the RSEMD method are quantitatively confirmed up to 8-th bar pair. No additional or significant artifacts are observed. Figure 4c looks brighter because of higher resolution and signal-to-noise (up to 60%) improvement (compare Fig. 5a and b) achieved as compared to FBP on Fig. 4b.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
5
Figure 3. A plot of the low contrast resolution section of the Catphan®500 phantom. Signal-to-noise (a) and contrast-to-noise (b) ratios compare standard FBP reconstruction with RSEMD recovery for deferent values of tube current-time (mAs).
CT number linearity
Noise reduction
The linearity of HU was initially estimated by determining the CT number for five different regions in the Catphan®500 phantom acquired with the standard dose (210 mAs and slice thickness of 0.5 mm) on the Toshiba CT scanner and reconstructed using the FBP algorithm. The metrics describing the accuracy of HU (average HU and SD) produced by the RSEMD algorithm, applied after FBP, are presented in Table 1 for comparison. After applying the RSEMD method, the CT numbers are very similar to those obtained with FBP and their standard deviation (SD) is lower than the results obtained with FBP. The standard deviation of the CT numbers between the central and peripheral ROI's (uniformity) was less than 4 HU in the results obtained after the application of RSEMD. The specific imaging characteristics for five different regions of the Catphan®500 phantom (see Table 1) are in excellent agreement with analogous results in Table 2 presented in Ref. [8] for the hybrid iDose algorithm and also with Iterative Reconstruction in Image Space algorithm (IRIS) [9].
In all of the phantom studies performed we found that the image noise was reduced after the application of RSEMD in all of the images where the zero-mean Gaussian noise was added. As an example, Fig. 6a shows the standard deviation (SD) of HU measured after FBP, and FBP þ RSEMD in the Teflon® (HU ¼ 990) region of the Catphan®500 uniformity section at different values of tube currenttime 22, 52, 105, 210 and 405 mAs. Figure 6b shows the ratio of this SD, demonstrating that the average improvement in SD of the HU after RSEMD is about 1.5. Figure 6c presents the contrast-to-noise ratio (CNR) for FBP and FBP þ RSEMD in the same region. For very noisy images with low CNR, the application of RSEMD improves the CNR about 1.48 times, as compared to the FBP method. Lesion conspicuity A qualitative improvement in lesion conspicuity is shown in Fig. 7. Zero-mean Gaussian noise (SNR is 3.04) was added to the
Figure 4. A plot of the high contrast resolution section of Catphan®500 phantom scanned on a Toshiba CT at 52 mAs: (a) and (b) with FBP reconstruction and (c) with RSEMD recovery method. The visual inspection (arrows) of the bar patterns (b) and (c) confirm the spatial resolution improvement using the RSEMD method up to 8-th bar pair (c).
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
6
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
Figure 5. Profiles of the high contrast resolution section (arrows on Figure 4) of the Catphan®500 phantom after reconstruction with FBP (a) and with the RSEMD recovery method (b). The signal-to noise and spatial resolution improvement with the RSEMD method quantitatively confirmed are up to 8-th bar pair (compare with Figure 4). No additional or significant artifacts are observed.
Table 1 Relative Hounsfield Units (HU) and standard deviation (SD) for five different regions of the Catphan®500 phantom. Material
FBP
Air Polyethylene Water Acrylic Teflon®
1008.3 95.3 4.9 123.6 995.2
RSEMD ± ± ± ± ±
10.2 12.2 3.7 9.9 9.3
1004.1 92.8 4.7 122.4 993.1
± ± ± ± ±
8.2 10.1 3.02 8.1 7.5
initial image generated by 4D XCAT. The image shown on Fig. 7a was obtained using FBP (SNR is 4.02 in the small low contrast lesion). The image shown in Fig. 7b was obtained after RSEMD was applied to the image shown in Fig. 7a (SNR is 5.82 in the liver lesion indicated by the arrow). Figure 8 quantitatively demonstrates the dependence of the signal-to-noise and contrast-to-noise ratios for the small liver lesion at different levels of SNR. The CNR and SNR values after RSEMD in the lesion reach a plateau after approximately 20 iterations with a total number of resolution subsets of around 5. Figure 9 shows the contrast-tonoise ratio versus number of iterations for a low contrast spherical liver lesion simulated by 4D XCAT with different levels of noise added (SNR ¼ 1.52, 2.18, 3.04, 4.02 and 8.05). The application of RSEMD after recovery by FBP improves the CNR about 1.45 times.
Figure 6. A plot of the linearity section of CT numbers in the Teflon® region of the Catphan®500 phantom. FBP reconstruction followed by the RSEMD recovery method versus current-time values at 22, 52, 105, 210 and 405 mAs: standard deviation (SD) associated with noise (a), ratio of SD before and after recovery with RSEMD (b) and contrast-to-noise ratio (CNR) for FBP and FBP þ RSEMD (c).
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
7
Figure 7. An XCAT phantom image obtained after FBP method (a) with SNR ¼ 4.02 and image obtained after 20 iterations with the RSEMD recovery method following FBP with improved SNR ¼ 5.82 (b). The visual inspection (arrow) confirms enhanced image quality in both SNR and CNR ratios.
Discussion In spite of the widespread availability of manufacturer provided IR methods [6,9e13], there still may be a need for additional fast IR methods that can be applied retrospectively to DICOM images which were previously reconstructed using the standard FBP methods. As an example, only about 20 percent of the CT scans performed in our county hospital are done on scanners equipped with IR algorithms. Also about 3e5% of the CT scans acquired at normal radiation dose are of marginal to non-diagnostic quality, usually, but not limited to patients with high body mass index (BMI). Desai et al. [6] compared the performance of ASIR (Adaptive Statistical Iterative Reconstruction, GE Healthcare) with FBP in 100 adults weighing more than 91 kg and found lower noise in the images at a reduced radiation dose in images reconstructed with ASIR. Kligerman et al. [7] compared FBP and iDose (Phillips Healthcare) [10] in terms of both image noise and image quality in obese patients (33 patients with an average BMI of 42.7) undergoing CT pulmonary angiography. They found an improvement in both image noise and contrast-to-noise with increasing levels of
iDose as compared to FBP. However, these added-on algorithms come with additional financial and time costs. The retrospective use of IR methods applied to projection data (which is usually stored in a proprietary format and not generally archived) to improve image quality becomes difficult, if not impossible. Several hours may pass between the acquisition of the CT scans and their clinical interpretation at which time the evaluation of their diagnostic quality is made. Artifacts (noise) are commonly encountered in clinical CT, can be problematic in medical imaging and may simulate pathology. Manufacturers use a variety of sophisticated techniques for artifact reduction (see for example existing commercial products iDose4 Premium Package [18] and Toshiba AIDR 3D [19]) in projection and imaging spaces to substitute the over range values in attenuation profiles. Noise in image space is greatly reduced using appropriate iterative reconstruction algorithms to rapidly post-process nondiagnostic CT scans. Currently, patients with non-diagnostic CT scans are often called back for a repeat CT examination, leading to increased radiation dose and delays in diagnosis and/or clinical management.
Figure 8. A plot of contrast-to-noise (a) and signal-to-noise (b) ratios compare the FBP method with FBP followed by RSEMD recovery on a low contrast, small, spherical liver lesion simulated by a 4D XCAT phantom.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
8
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9
has shorter processing time than conventional algorithms. To reconstruct an image with higher resolution it is not necessary to spend time for extra iterations as is required for the deblurring algorithm (2) or any method which use one resolution parameter for multiple image updates passed through the data (conventional subsets). There is little systematic information in the literature about the dependence of the number of iterations in image quality [25]. An important next step will be to carry out a study on poor quality (noisy) clinical CT scans to determine what fraction of these scans can be “restored” by RSEMD to the point where they can be read clinically with confidence. It is also important to point out that the RSEMD code employed in this study was not optimized for computational speed (~300 msec per image update). It is expected that improvements will be made to reduce the processing time appreciably for typical clinical workflow. Conclusion
Figure 9. A plot of contrast-to-noise ratio versus the number of iterations for a low contrast spherical liver lesion simulated by 4D XCAT with different levels of noise added.
Note, it is not possible to determine the absolute differences in the existing spectrum of commercial reconstruction algorithms [3,7,10,18e22] due to the absence of detailed information about these algorithms in the literature. Thus, there is an unmet clinical need for rapid recovery algorithms in image space that are generalizable across multiple manufacturers' CT platforms, anatomic imaging sites, and patient populations to post-process CT scans that may be noisy and have low contrast. A mathematical approach to suppress both the noise and edge artifacts for the modified EM algorithm [23] was successfully adopted for CT image deblurring [12]. To estimate the parameters of the PSF, the blind deblurring algorithm similar to (2) maximizes the edge-to-noise ratio and deblurs the images with a relatively high number of iterations. According to Ref [12] this deblurring procedure (2) can produce a gain of up to 40% in image resolution. The permissible range of the blurring parameter sD can be estimated from the system modulation transfer function (MTF) and from data acquired in CT quality assurance tests. Unfortunately, the optimal number of iterations (about 100 in average) depends on image content and can only be estimated for a class of images through digital and experimental phantom studies. Incorporating a priori information about the actual characteristics of the blurred system can greatly improve some of the ill-posedness of this blind deblurring mechanism [24]. The use of the data subsets approach in the reconstruction procedure always offers an essential speed increase and better signal-to-noise characteristics. The conventional ordered subsets technique to partition the data into subsets can't be applied for clinical CT images since the system input is pixels not lines. To overcome all of these difficulties, the RSEMD method, incorporating an original deconvolution scheme in image space using resolution subsets (approximation of the PSF) to maximize SNR, was developed. The computations associated with each subset depend on the image update (intermediate image after one or more iterations) from the previous subset's iteration. When all of the resolution subsets are employed, an image recovery procedure has been performed. In this case an enhanced CT image recovery with RSEMD
A phantom study is a necessary first stage of the presented investigation for CT image enhancement. This is an important opportunity and one way to optimize diagnostic information for the older models of CT scanners where the modern versions of iterative reconstruction techniques may not be available. The application of this rapid technique to clinical imaging looks very promising for image resolution recovery, when applied to noisy clinical CT images, especially to images where noise is very problematic such as in ICU patients. Designing a study to evaluate the performance of the RSEMD as an image recovery method is the next step in determining the clinical utility of this approach. Conflicts of interest The authors have no conflicts of interest to declare. Acknowledgment The authors would like to thank Drs. Gary Arbique, Cecelia Brewington, Travis Browning, Julie Champine, Daniel Costa and the Clinically Appropriate and Accurate Radiation Exposure Committee of the Department of Radiology (UT Southwestern Medical Center at Dallas, TX) for their advice and collaboration. Dr. Robert E. Lenkinski acknowledges the support of the Missing Link Award from the Cancer Prevention Research Institute of Texas (CPRIT), 2011e2015. Dr. Nikolai Slavine would like to thank Dr. W. P. Segars (Duke University Medical Center, NC) for C-programming computer installation details of CT simulation tool 4D XCAT and Drs. Ananth Madhuranthakam and Zhongwei Zhang for valuable help with multiple DICOM CT files operation in MATLAB®. Special thanks to Jocelyn Chafouleas, Glenn Katz and Saleh Ramezani for technical assistance with the data presentation. References [1] Mehta D, Thompson R, Morton T, Dhanantwari A, Shefer ED. Iterative model reconstruction: simultaneously lowered computed tomography radiation dose and improved image quality. Med Phys Int J 2013;1(2):147e55. [2] Kordolaimi SD, Argentos S, Pantos I, Kelekis NL, Efstathopoulos EP. A new era in computed tomographic dose optimization: the impact of iterative reconstruction on image quality and radiation dose. J Comput Assist Tomogr 2013;37(6):924e31. €ve A, Olsson ML, Siemund R, Stålhammar F, Bjo €rkman-Burtscher IM, [3] Lo € derberg M. Six iterative reconstruction algorithms in brain CT: a phantom So study on image quality at different radiation dose levels. Br J Radiol 2013;86(1031):20130388. [4] Pan X, Sidky EY, Vannier M. Why do commercial CT scanners still employ traditional, filtered back projection for image reconstruction? Inverse Probl 2009;25:1e36.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009
N.V. Slavine et al. / Physica Medica xxx (2015) 1e9 €l BP, Fingerle AA, Renger B, Münzel D, Rummeny EJ, Dobritz M. Initial [5] Noe performance characterization of a clinical noise e suppressing reconstruction algorithm for MDCT. Am J Roentgenol 2011;197:1404e9. [6] Desai GS, Uppot RN, Yu EW, Kambadakone AR, Sahani DV. Impact of iterative reconstruction on image quality and radiation dose in multidetector CT of large body size adults. Eur Radiol 2012;22(8):1631e40. [7] Kligerman S, Mehta D, Farnadesh M, Jeudy J, Olsen K, White C. Use of a hybrid iterative reconstruction technique to reduce image noise and improve image quality in obese patients undergoing computed tomographic pulmonary angiography. J Thorac Imaging 2013;28(1):49e59. [8] Noel P, Fingerle A, Renger B, Rummeny EJ, Dobritz M. A clinical comparison study of a novel statistical iterative and filtered back-projection reconstruction. In: Pelc J, Ehsan S, Nishikawa R, editors. Medical imaging: physics of medical imaging. Bellingham: SPIE; 2011. [9] Ghetti C, Ortenzia O, Serreli G. CT iterative reconstruction in image space: a phantom study. Phys Med 2012;28:161e5. [10] Scibelli A. iDose4 iterative reconstruction technique (Philips Healthcare Web site), www.clinical.netforum.healthcare.philips.com/global/Explore/WhitePapers/CT/iDose4-iterative-reconstruction-technique [accessed 11.03.11]. [11] Pickhardt PJ, Lubner MG, Kim DH, Tang J, Ruma JA, del Rio AM, et al. Abdominal CT with Model-Based Iterative Reconstruction (MBIR): initial results of a prospective trial comparing ultra-low dose with standard-dose imaging. Am J Roentgenol 2012;299(6):1266e74. [12] Ge Wang, Vannier MW, Skinner MW, Cavalcanti M, Harding G. Spiral CT image deblurring for cochlear implantation. IEEE Trans Med Imag 1998;17:251e62. [13] Tipnis S, Ramachandra A, Huda W, Hardie A, Schoepf J, Costello P, et al. Iterative reconstruction in image space (IRIS) and lesion detection in abdominal CT. 2010. SPIE Press, Med, Imaging; Bellingham, WA, Proc SPIE 7622. [14] Catphan®500 and 600, manual. Cambridge, NY: The Phantom Laboratory; 2012. p. 1e40. [15] Segars WP, Mahesh M, Beck TJ, Frey EC, Tsui BM. Realistic CT simulation using the 4D XCAT phantom. Med Phys 2008;35(8):3800e8. http://dmip1.rad.jhmi. edu/xcat/.
9
[16] Antich P, Parkey R, Seliounine S, Slavine NV, Tsyganov E, Zinchenko A. Application of expectation maximization algorithms for image resolution improvement in a small animal PET system. IEEE Trans Nucl Sci 2005;52(3): 684e90. [17] E. Tsyganov, A. Zinchenko, N.V. Slavine, P. Antich, S. Seliounine, O. Oz, et al. Reconstruction algorithm with resolution deconvolution for 3D image in a small animal PET imager. Springer Science Media Inc, 163e176, [Chapter 8]. [18] IDose4 premium package, © Koninklijke Philips N.V. 2014. http://www. healthcare.philips.com/main/products/ct/products/idose4_premium_ package/. [19] AIDR 3D, Toshiba Medical System Corporation. 2015. http://www. toshibamedicalsystems.com/tmd/english/products/dose/lowdose/ reconstruction.html. [20] Silva AC, Lawder HJ, Hara A, Kujak J, Pavlicek W. Innovations in CT dose reduction strategy: application of the adaptive statistical iterative reconstruction algorithm. Am J Roentgenol 2010;194:191e9. [21] Matsuki M, Murakami T, Juri H, Yoshikawa S, Narumi Y. Impact of adaptive iterative dose reduction (AIDR) 3D on low-dose abdominal CT: comparison with routine-dose CT using filtered back projection. Acta Radiol 2013;54: 869e75. [22] Price RG, Vance S, Cattaneo R, Schulz L, Elsheikh MA, Chetty IJ, et al. Characterization of a commercial hybrid iterative and model-based reconstruction algorithm in radiation oncology. Med Phys 2014;41:081907. [23] Snyder DL, Miller M, Thomas L, Politte DG. Noise and edge artifacts in maximum-likelihood reconstructions for emission tomography. IEEE Trans Med Imaging 1987;6:228e38. [24] Kudo H, Courdurier M, Noo F, Defrise M. Tiny a priori knowledge solves the interior problem in computed tomography. Phys Med Biol 2008;53: 2207e31. [25] Pontana F, Duhamel A, Pagniez J, Flohr T, Faivre JB, Hachulla AL, et al. Chest computed tomography using iterative reconstruction vs FBP (part 2) image quality of low dose CT examinations in 80 patients. Eur Radiol 2010;21: 636e43.
Please cite this article in press as: Slavine NV, et al., An iterative deconvolution algorithm for image recovery in clinical CT: A phantom study, Physica Medica (2015), http://dx.doi.org/10.1016/j.ejmp.2015.06.009