Automated Lung Tumor Detection and Quantification for Respiratory Gated PET/CT Images Jiali Wang∗a, Misael del Vallea, Juan Franquizb, Anthony McGorona a Department of Biomedical Engineering, Florida International University, Miami, Fl; b Radiological Physics of South Florida, Inc. Miami, FL ABSTRACT Purpose: To develop and validate an automatic algorithm for the detection and functional assessment of lung tumors on three-dimensional respiratory gated PET/CT images. Method and Materials: First the algorithm will automatically segment lung regions in CT images, then identify and localize focal increases of activity in lung regions of PET images at each gated bin. Once the tumor voxels have been determined, an integration algorithm will include all the tumor counts collected at different bins within the respiratory cycle into one reference bin. Then the total activity (Bq), concentration (Bq/ml), functional volume (ml) and standard uptake values (SUV) are calculated for each tumor on PET images. Validation of the automatic algorithm was demonstrated by conducting experiments with the computerized 4D NCAT phantom and with a dynamic lung-chest phantom imaged using a GE PET/CT System at Baptist Hospital of Miami. Tumor variables to be controlled were: volume, total number of counts (activity), maximum and average number of counts. These values were the gold standard to which the results of the algorithm were compared. The tumor’s motion was also controlled with different respiratory periods and amplitudes. Results: Validation, feasibility and robustness of the algorithm were demonstrated. With the algorithm, the best compromise between short PET scan time and reduced image noise can be achieved, while quantification and clinical analysis become faster and more precise. Keywords: Lung tumor, PET, CT, Standard Uptake Value (SUV)
1.
INTRODUCTION
Lung cancer is one of the most common types of cancers in the United States, with more than 175,000 deaths per year [1]. Since the early treatment of small lung tumors has a high probability of curability, the accurate definition of tumor volume and activity is especially important [2]. Molecular imaging with 18FDG-PET can provide significantly higher sensitivity (87%) and specificity (91%) than CT in detection and characterization of malignant lung nodules [1 - 2], it has become a popular imaging modality for lung cancer diagnosis, staging, and for differentiating tumor recurrence. But one major disadvantage of 18FDG-PET imaging is the respiratory motion artifacts because of the relatively long acquisition time. Many clinical and research studies have shown that respiratory motion degrades the PET images by reducing SUV (standardized uptake value) and tumor-to-background ratio, distorting the real size, shape and location of the tumor and other structures. These artifacts negatively impact the application of 18FDG-PET for the detection of small tumors, monitoring response to treatment and radiation therapy planning. Gating in PET was proposed to solve the problem of respiratory motion artifacts. Varian Medical Systems (Palo Alto, CA) has developed the Real-Time Position Management (RPM) Respiratory Gating System [3]. They demonstrated more accurate quantification and definition of PET tumors by gating. Anzai Medical (Japan) developed the AZ-773V system employing a strain gauge sensor to detect the mechanical expansion of the thoracic cavity resulting from respiratory motion. Wang et al reported utilizing a solid-state thermometer to detect the temperature difference of the nostril air flow due to inhalation and expiration [4]. All these methods have demonstrated the quantitative and qualitative benefits of reducing the blurring of tumors by taking images at discrete bins within the respiratory cycle. However, PET images collected at discrete bins can be significantly affected by noise, unless the total PET acquisition time is prolonged. ∗
Jiali Wang is with Department of Biomedical Engineering, Florida International University, 10555 West Flagler St. EC 2610, Miami, FL 33174, USA (telephone: 305-348-6950, email:
[email protected]) Medical Imaging 2008: Image Processing, edited by Joseph M. Reinhardt, Josien P. W. Pluim, Proc. of SPIE Vol. 6914, 69144G, (2008) 1605-7422/08/$18 · doi: 10.1117/12.772862 Proc. of SPIE Vol. 6914 69144G-1
2008 SPIE Digital Library -- Subscriber Archive Copy
Several image-based and projection-based algorithms have also been developed to correct for motion artifacts in PET. Affine transformation has been applied to PET data in list mode [8], deconvolution method has been used to correct for respiratory motion without the need for gating [9], and also global optical flow has been implemented to reduce lung motion artifacts with positive results [10]. To date, such approaches are still being developed and have not been implemented clinically. These methods are promising because they need not interfere with the current operation of the imaging session, but methods based on external optical tracking are further in clinical application since they are less computationally intensive. Since gating methods inherently decrease the image signal-to-noise ratio, they should be combined with some imaging-based methods to transform the set of gated images into a single image for analysis. From such a view point, we developed an automatic computer algorithm to detect and quantitatively assess lung tumors by conducting experiments with the computerized 4D NCAT phantom and with a dynamic lung-chest phantom imaged using a GE Discovery LS PET/CT System at Baptist Hospital of Miami. In addition to the automatic detection, quantification and definition of malignancy of lung tumors, this method will also provide a quantitative, fast technique for assessing the response of lung cancer to therapy and improving target definition in radiation treatment planning.
2.
METHODS
2.1 Physical Phantom and Computer Phantom Experiments were conducted using a dynamic lung-chest phantom (Model ECT/LUNG/P) which is a fully tissue equivalent anthropomorphic phantom, including a large, body-shaped cylinder with lung, spine and liver inserts (Figure 1). The lungs can be filled with Styrofoam beads or air to simulate lung tissue density. Hollow spheres filled with 18FDG were used to simulate lung tumors. The movement of the sphere is driven by a stepper motor controlled by a PIC microcontroller that allows the user to select different tumor motion parameters, i.e., different frequency and different amplitude to simulate different respiratory periods and amplitudes. Detail parameters of the experiments can be found in Table 1. Another stepper motor was used to simulate the movement of the chest.
+
III iii
Tumor FDG dose diluted into 1 liter water as tumor FDG concentration added into hollow spheres
GE Discovery LS PET/CT
+ Background FDG dose diluted into the phantom as background concentration
Spheres controlled by a step motor
Figure 1. Diagram for physical phantom experiments. Tumor FDG dose was first diluted into 1 liter water and then added into hollow spheres to simulate tumor FDG concentration and background FDG dose were diluted into the phantom to simulate background FDG concentration.
Proc. of SPIE Vol. 6914 69144G-2
Table 1. Parameters for physical phantom experiments
Sphere diameter (mm) Sphere volume (ml) Sphere position Respiratory cycle (second) Respiratory amplitude (mm) Tumor/background ratio Image Acquisition
17.69, 14.43, 11.89 2.0, 1.0, 0.5 Left lung 4 20 (5.98, 6.35, 6.01 ) ≈ 6 Ungated & Static PET: 5 min Gated PET: 10 min for 10 gates
A computer four-dimensional (4D) phantom named NURBS-based cardiac-torso (NCAT) phantom developed by Segars and Tsui at the University of North Carolina was also used to simulate different types of respiratory cycles, and different tumor positions [12 - 13]. This 4D NCAT phantom is a well established simulation program and is widely used as a gold standard in nuclear medicine imaging research. It can be used to simulate 18FDG distributions of activity and lung tumors. Tumor variables to be controlled were: volume, total number of counts (activity), maximum and average number of counts. These values were the gold standard to which the results of the algorithm were compared. The respiratory periods and amplitudes of the tumor movements were also controlled. Original NCAT phantom data is noise free. By selecting regions of interest at different slices, the mean and standard deviation (SD) of counts could be determined for each structure of interest (tumor, lung, soft tissue). To simulate real PET data with noise each voxels v(i) of each structure was assigned to the value: mean RND (*) × SD , ± v (i ) = n n where mean is the average value of all the counts in each structure, SD = mean , n is the number of time bins used to simulate respiration and RND(*) is a random number with normal distribution (mean equal to zero and standard deviation equal to one). Finally, the blurring effect due to the finite resolution of PET images was induced by the convolution of each transaxial slice with a Gaussian filter in which the FWHM of 5mm in the x and y directions are applied, corresponding to the approximate resolution of the PET camera used. The axial blurring was performed by the convolution of the images in the axial direction with a one-dimensional Gaussian filter. 2.2 PET/CT Scan All experiments were conducted using a Discovery LS PET/CT Scan (GE Medical Systems) and Varian Respiratory Gating System. This Hybrid system includes in the same instrument, the GE Light Speed multi-slice CT and the Advance Nxi PET scanner. Emission and transmission images are automatically registered and the CT map is used for attenuation correction of PET data. 2.3 Tumor Segmentation The initial stage of the motion tracking and integration algorithm for processing and analysis of gated PET images is the computer-assisted lung tumor segmentation. Since the respiratory cycle will be divided into at least six or eight time bins, the manual identification of the tumor in each time bin would be a time consuming and tedious task. The computer-assisted automatic identification of tumors would make the analysis and processing of gated PET images easier and faster. To accomplish this objective we have developed a technique for automatic identification of transaxial CT slices corresponding to lung regions, segmentation of lung areas, and detection and localization of tumor volumes in un-gated PET images. Figure 2 shows a flow diagram of the general strategy for automatic identification of lung tumors
Proc. of SPIE Vol. 6914 69144G-3
on gated PET Images. The final output of the automatic localization algorithm is a text file with the voxel coordinates and number of counts of each voxel for each respiratory time bin. This is the basic information that will be used in the motion track and integration algorithm. Preliminary studies with this method were presented at the Medical Imaging 2004 Conference in the International society for Optical Engineering (SPIE) [11]. All of the software code was developed using the Interactive Data Language (IDL), Research Systems, Inc (Boulder, CO). The computer-assisted algorithm was preliminarily assessed using physical phantom studies performed at Baptist Hospital of Miami.
Segmented lungs
Binary images of CT lungs are dilated to account for diffusion in PET. The binary templates are used to segment lungs in PET
Find seed voxel
Original CT Images
Region Growing Segmented lung regions excluding tumors Binary images of CT lungs are eroded to remove the entire tumor. This binary template is used to segment lung background in PET
Seed Background
Find average background Figure 2. Flow diagram for automatic tumor segmentation in PET images
2.4 Tumor Registration/Integration Once the tumor voxels have been determined at each gated time bin, a registration/integration algorithm will include all the tumor counts collected at different time bins within the respiratory cycle into solely one reference time bin. The basic idea is that, given a moved volume Vi at the bin i and a corresponding reference volume V0 at the reference time bin, find a transformation T such that the transformed volume T(Vi) matches as close as possible the reference volume V0. Since the number of voxels involved in each tumor is relatively small and all the voxels are compacted around a maximum value, only simple matching methods will be considered. There are two integration strategies used in this work, one is to register directly the target bin to the reference bin as shown on the left of Figure 3, and the other is to register each bin one by one as shown on the right of Figure 3. The second one could be more accurate as the discrepancy between each bin is less compared to the first one, but the registration process might take more computation time.
Proc. of SPIE Vol. 6914 69144G-4
Bin 1
T Transform
Bin 1
......
Bin 2
......
Bin 2
T Transform
......
Bin i-2 ......
T Transform
Bin ref*
Bin i-1
Bin i T Transform
Bin ref*
Bin i
Bin ref*
Bin ref*
Bin ref*
......
Bin ref*
Sum Bins *
Sum Bins * (a)
(b)
Figure 3. The process flow of two registration/integration strategies. (a) Register directly the target bin to the reference bin and (b) Register each bin one by one.
2.4.1
Intensity Based Registration
This method takes into account voxel values. The voxel with the maximum number of counts at Vi will be registered with the voxel with the maximum number of counts at V0. The same will be performed for the second maximum and subsequent voxels. The main inconvenience of this procedure is that Vi and V0 would include different numbers of voxels. Since respiratory motion is a continuous process, PET lung tumors within a discrete time bin can also present blurring. In this case, some deformation in the extended volume needs to be included. One deformation that can be assessed is the compression of the peripheral voxels into the volume. That is, to shrink the volume of the tumor by adding the peripheral pixels to their neighbors with lower counts. The shrinking process will start with those voxels of lowest activity. This process can be applied previous to any other registration approach. 2.4.2 Centroid Based Registration The centroid of each tumor volume will be calculated from:
c ( x, y , z ) =
∑ w a ( x, y , z ) ∑w i.
i
where a is the x, y, or z coordinate of the i-th voxel and wi is the voxel value of the i-th voxel. The motion vector (dx, dy, dz) calculated from the 3D distance between the two centroids, will move all the voxels in Vi into Vo: Vi ( x + dx, y + dy , z + dz ) = V0 ( x, y, z ) Optimal matching will be determined from the minimal value of the sum of the squared count differences (S) between integrated voxels into the reference time bin and those obtained with static acquisitions when the simulated tumor (either in the physical phantom or in the computerized phantom) has the fixed position corresponding to the reference bin:
1 ∑ (Vi 0 ( x, y, z) − V0 ( x, y, z)) 2 n where n is the total number of voxels included in the comparison, Vi0(x, y, z) is the number of counts of voxels integrated into the reference bin, and V0(x, y, z) is the number of counts of the voxel corresponding the static position of the tumor in the reference bin. S=
Proc. of SPIE Vol. 6914 69144G-5
The registered bins were also evaluated with reference to the non-registered bins using the correlation coefficient, which is defined as: ∑i [(Vi 0 ( x, y, z ) − m) × (V0 ( x, y, z ) − n)] correlatio n = ∑ (Vi 0 ( x, y, z ) − m) 2 × ∑ (V0 ( x, y, z ) − n) 2 i
i
where Vi0(x, y, z) and Vi(x, y, z) are the values from two datasets in the comparison and m, n are the mean values from the same datasets.
3.
RESULT
3.1 Tumor Segmentation The result of segmenting the lung region in CT images and tumors in PET images is shown in the Figure 4.
Figure 4. Original CT image of the lung-chest phantom (left), segmented lung region (center) and position and extension of simulated PET tumors (right) determined automatically by the algorithm
3.2 Tumor Registration of NCAT Phantom The registration/integration algorithm has been validated using computer simulated NCAT phantom data for different size tumors. Parameters for the NCAT phantom is shown in Table 2. The results are evaluated with reference to the non-registered bins as shown in Figure 5, we can see with two registration methods cross correlation values improve. Table 2. Parameters for NCAT phantom experiments
Sphere diameter (mm)
6, 8.5, 10, 20, 25
Sphere position
Left lung
Respiratory cycle (second) Respiratory amplitude (mm) Tumor/background ratio
5 2 2.5
Proc. of SPIE Vol. 6914 69144G-6
Centroid Based Registration 1
0. 8
0. 8
Correlation Result
Correlation Result
Intensity Based Registration 1
0. 6
0. 4
Before Registration After Registration
0. 2
0
6
8. 5
10
Tumor Size (mm)
20
Before Registration After Registration
0. 6
0. 4
0. 2
0
25
6
8. 5
10
Tumor Size (mm)
(a)
20
25
(b)
Figure 5. NCAT phantom results of two registration methods: before registration vs. after registration, (a) intensity based registration and (b) centroid based registration. After registration, correlation values increase for difference size tumors with both methods.
3.3 Tumor Registration of Physical Phantom The registration/integration algorithm has also been validated with the physical phantom experiments for different sizes of tumors. We can see from Figure 6 below, the activity concentration and tumor volume after registration are closer to the real values. Also, the correlation and noise level, after registration, are higher and lower respectively. Here the noise level is estimated by the standard deviation of the tumor region over the average value of the tumor region.
Tumor Volume
Activity Concentration (Normalized) 4
1
Tumor Volume (ml)
Activity Concentration (Normalized to Static PET)
1. 2
0. 8 0. 6 0. 4 0. 2
Static PET Intensity Registration Centroid Registration Gated PET Ungated PET
3 2. 5
Real Volume Intensity Registration Centroid Registration Gated PET Ungated PET
2 1. 5 1 0. 5 0
0 11. 89
`
3. 5
14. 43
Tumor Size (mm)
17. 69
11. 89
(a)
14. 43
Tumor Size (mm)
17. 69
(b)
Figure 6. Physical phantom results: (a) Comparing activity concentration of two registration methods, static PET, gated PET and ungated PET images, all the values normalized to the static PET (gold standard). (b) Comparing tumor functional volume of two registration methods with real volume, gated PET and ungated PET images.
Proc. of SPIE Vol. 6914 69144G-7
Correlation
Correlation
0. 8
23%
21.1%
19.5%
17.1%
18.5%
15.9%
0. 6 0. 4
Before Registration Intensity Registration Centroid Registration
0. 2 0 11. 89
14. 43
Tumor Size (mm)
Standard Deviation / Mean
Noise (Std Dev / Mean)
1
Before Registration Intensity Registration Centroid Registration
0. 8 0. 6
23.8% 33.3%
42.7% 48.9% 0. 4 39.5% 39.9% 0. 2 0 11. 89
17. 69
(a)
14. 43
Tumor Size (mm)
17. 69
(b)
Figure 7. Physical phantom results: (a) Correlation results of two registration methods comparing before registration. (b) Noise level after two registration methods comparing with before registration
4.
DISCUSSION AND CONCLUSION
Improving the detectability of a malignant lung tumor in its initial stage using 18FDG-PET will impact positively lung cancer patient care. However, because of the long duration of whole body PET scans, tumor and organ motion due to respiration can be a major challenge for accurate localization and quantification of PET images as the image will be blurred and the tumor smeared. The overall goal of this study is to develop and validate a fast and practical solution to the problem of respiratory motion for the accurate interpretation and quantitation of 18FDG uptake of lung PET images. Several papers appeared in recent years describing different image processing algorithms compensating for motion artifacts. There is a comprehensive review paper of motion correction methods in PET being published by Rahmim [7], but most are for brain and heart studies. Qiao et al have achieved motion correction by successfully applying non-rigid motion compensation to computer simulated list-mode PET data [14]; similar to it, Lamare et al also apply affine transformation on list-mode PET data [8]. While list mode collection is not generally implemented on clinical cameras, it is probably not a limiting obstacle. Deconvolution has been reported to correct lung motion artifact with positive results [9]. The method depends on an estimate of patient motion measured from 4D CT images. Generally, deconvolution methods are accurate for noise-less data, but it tends to amplify the noise in real PET data. Another method proposed by Dawood et al [10] utilizes a global optical flow algorithm for motion correcting images in individual gates. The method uses four assumptions to perform the deformable registration: intensity similarity, incremental transformation, smoothness, and error minimization, which cause to suffer from inaccuracy in presence of high noise. All these methods perform the registration on the entire PET image data which require more computation time and complicate the algorithm than the method proposed here. The innovative aspect of this project is to develop an automatic motion track and integration algorithm that includes all the counts collected in the respiratory cycle into solely one reference bin. This method has the advantages: (1) the computer-assisted automatic algorithm would simplify the following integration/registration algorithm which is only performed on the segmented tumor region, and would facilitate the 3D quantitation of activity and the introduction of the procedure to the clinical practice; (2) PET scan time doesn’t need to be increased to reducie statistical noise and increase signal to noise ratio, the integration of the information of different time bins into one set of tomographic slices, would make easier, faster and more reproducible the clinical interpretation of 18FDG-PET scans. Our experiments with both the NCAT software phantom as well as with physical phantom showed significant improvement in motion corrected data. For tumor activity concentration and tumor volume, we can get the value closer to the static tumor (real value) with reference to the ungated PET after the two registration methods (Figure 6). The improvements before and after correction are also quantified using the correlation coefficient. A 26.6% average improvement with the centroid-based registration and 21.0% average improvement with intensity-based registration were
Proc. of SPIE Vol. 6914 69144G-8
achieved for NCAT phantom data, and 21.2% average improvement with centroid-based registration and 17.7% average improvement with intensity-based registration were achieved for physical phantom data as shown in Figure 5 and Figure 7(a) respectively. The noise level is greatly reduced by an average of 35.4% and 40.7% after the two registration methods respectively as can be readily seed in Figure 7(b). In this study all the image processing algorithms were performed on CT attenuation corrected PET data. Error could occur here because only one CT image is taken during respiration while PET acquisition is a continuous process. For better results, attenuation correction should be applied after the integrated PET image is created for the reference bin, and the reference bin should be selected to match the CT transmission image, or by using respiratory gated CT images. Another parameter that could be optimized is the gating scheme. In all these experiments we use the first time bin as the reference bin and perform registration methods directly to the reference bin (Figure 3(a)). If we select the middle bin (e.g. 5th bin out of 10 bins, which is the average amplitude of respiratory motion), or use the successive registration method (Figure 3(b)), we can reduce the discrepancy between the target and reference gates and can probably get better results after integration. In this research project, we developed and validated a computer-assisted method that can automatically localize tumors in lung PET images of discrete bins within the breathing cycle, followed by an algorithm that integrated all the information of a complete respiratory cycle into a single reference bin. In this way, the detection and quantification of metabolic active volume of lung tumor can be done automatically. Comparing the results of the tumor before registration with after registration we can see higher correlation values and lower noise. This method allowed more accurate estimation of tumor functional volume, more accurate three-dimensional quantitative analysis of activity concentration, and also the best compromise between short PET scan time and reduced image noise can also be achieved. It will provide a more reliable method for assessing the response of lung cancer to therapy with 18FDG-PET with minimal interaction of an operator.
ACKNOWLEDGEMENT This work was supported by a grant from the NIH (R15CA118284-01).
Proc. of SPIE Vol. 6914 69144G-9
REFERENCES [1] H. Steinert, M. Hauser, F. Allemann, H. Engel, T. Berthold, G. K. Schulthess, and W. Weder, “Non-small cell lung cancer: Nodal staging with FDG PET versus CT with correlative lymph node mapping and sampling”, Radiology 202, 441–446, 1997. [2] Shim SS, Lee KS, Kim BT, Chung MJ, Lee EJ, Han J, Choi JY, Kwon OJ, Shim YM, Kim S. “Non–Small Cell Lung Cancer: Prospective Comparison of Integrated FDG PET/CT and CT Alone for Preoperative Staging”, Radiology 236(3): 1011-9, 2005. [3] S.A. Nehmeh, Y.E. Erdi, K.E. Rosezweig, O.D. Squire, L.E. Braban, E. Ford, K. Sidhu, G. S. Mageras. “Reduction of respiratory motion artifacts in PET imaging of lung cancer by respiratory correlated dynamic PET: Methodology and comparison with respiratory gated PET”, J. Nucl. Med. 44: 1644 – 1648, 2003. [4] Y. Wang, H. Baghaei, H. Li, Y. Liu, T. Xing, J. Uribe, R. Ramirez, S. Xie, S. Kim, and W.-H. Wong, "A Simple Respiration Gating Technique and Its Application in High-Resolution PET Camera," IEEE Trans. Nucl. Sci., vol. 52, 2005. [5] S.A. Nehmeh, Y.E. Erdi, C.C. Ling, K. E. Rosenzweig, O.D. Squire. “Effect of respiratory gating on reducing lung motion artifacts in PET imaging of lung cancer”, Med. Phys. 29: 366 – 371, 2002. [6] Chu JZ, Tsui BMW, and Segars WP. “A simulation Study of the Effect of Gating Scheme on Respiratory Motion Blurring in FDG Lung PET”, in a Supplement to JNM 43(5), p. 208, 2002. [7] A. Rahmim, "Advanced Motion Correction Methods in PET," Iran J Nucl Med, vol. 13, 2005. [8] F. Lamare, T. Cresson, J. Savean, C. Cheze Le Rest, A. J. Reader, and D. Visvikis, "Respiratory motion correction for PET oncology applications using affine transformation of list mode data," Phys Med Biol, vol. 52, pp. 121-40, 2007. [9] M. Dawood, N. Lang, X. Jiang, and K. P. Schafers, "Lung motion correction on respiratory gated 3-D PET/CT images," IEEE Trans Med Imaging, vol. 25, pp. 476-85, 2006. [10] I. El Naqa, D. A. Low, J. D. Bradley, M. Vicic, and J. O. Deasy, "Deblurring of breathing motion artifacts in thoracic PET images by deconvolution methods," Med Phys, vol. 33, pp. 3587-600, 2006. [11] J.M. Franquiz, S. Vaddahi, G. Soler. "Computer-aided lung nodule detection and assessment by using a hybrid PET/CT scanner," Proc. SPIE, 5369-49. San Diego, CA, February 14 to 19, 2004. [12] W. P. Segars, B. M. Tsui, and A. J. D. Silva, "CT-PET image fusion using the 4D NCAT phantom with the purpose of attenuation correction," EEE Trans. Nuclear Science, 2002. [13] W. P. Segars and B. M. Tsui, "Study of the efficacy of respiratory gating in myocardial SPECT with the new 4D NCAT phantom," IEEE Trans. Nucl. Sci., vol. 49, 2002. [14] F. Qiao, J. Clark, T. Pan, and O. Mawlawi, "Expectation Maximization Reconstruction of PET Image with Non-rigid Motion Compensation," Conf Proc IEEE Eng Med Biol Soc, vol. 4, pp. 4453-6, 2005. [15] J. Wang, J. Byrne, J. Franquiz, and A. McGoron, "Evaluation of amplitude-based sorting algorithm to reduce lung tumor blurring in PET images using 4D NCAT phantom," Comput Methods Programs Biomed, vol. 87, pp. 112-22, 2007.
Proc. of SPIE Vol. 6914 69144G-10
