Survey

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Survey

Document related concepts

Transcript

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 25, NO. 12, DECEMBER 2006 1573 Scatter Correction Method for X-Ray CT Using Primary Modulation: Theory and Preliminary Results Lei Zhu*, Member, IEEE, N. Robert Bennett, and Rebecca Fahrig Abstract—An X-ray system with a large area detector has high scatter-to-primary ratios (SPRs), which result in severe artifacts in reconstructed computed tomography (CT) images. A scatter correction algorithm is introduced that provides effective scatter correction but does not require additional patient exposure. The key hypothesis of the algorithm is that the high-frequency components of the X-ray spatial distribution do not result in strong high-frequency signals in the scatter. A calibration sheet with a checkerboard pattern of semitransparent blockers (a “primary modulator”) is inserted between the X-ray source and the object. The primary distribution is partially modulated by a high-frequency function, while the scatter distribution still has dominant low-frequency components, based on the hypothesis. Filtering and demodulation techniques suffice to extract the low-frequency components of the primary and hence obtain the scatter estimation. The hypothesis was validated using Monte Carlo (MC) simulation, and the algorithm was evaluated by both MC simulations and physical experiments. Reconstructions of a software humanoid phantom suggested system parameters in the physical implementation and showed that the proposed method reduced the relative mean square error of the reconstructed image in the central region of interest from 74.2% to below 1%. In preliminary physical experiments on the standard evaluation phantom, this error was reduced from 31.8% to 2.3%, and it was also demonstrated that the algorithm has no noticeable impact on the resolution of the reconstructed image in spite of the filter-based approach. Although the proposed scatter correction technique was implemented for X-ray CT, it can also be used in other X-ray imaging applications, as long as a primary modulator can be inserted between the X-ray source and the imaged object. Index Terms—Cone-beam CT, primary modulation, scatter correction, X-ray scatter. I. INTRODUCTION N X-ray system with a large area detector, such as is used for cone-beam computed tomography (CT) (CBCT), typically has high scatter. Scatter causes severe distortions and contrast loss in the reconstructed images [1]–[5], and developing an effective scatter correction method is still a major challenge. Many methods have been proposed in the literature and provide significant improvement in image quality, although drawbacks still exist. One category of the scatter control techniques includes hardware-based methods [6]–[11], such as anti-scatter grid, air gap, scatter measurement using a beam stop array etc. A Manuscript received April 27, 2006; revised August 29, 2006. This work was supported in part by the National Institutes of Health (NIH) under Grant NIH R01 EB003524 and in part by the Lucas Foundation. Asterisk indicates corresponding author. *L. Zhu is with the Department of Electrical Engineering, Stanford University, Stanford, CA 94305 USA (e-mail: [email protected]). N. R. Bennett and R. Fahrig are with the Department of Radiology, Stanford University, Stanford, CA 94305 USA. Digital Object Identifier 10.1109/TMI.2006.884636 These approaches suppress or correct for the presence of scatter by modifying the X-ray system. An alternative approach using software-based methods estimates and corrects for the scatter based on the system geometry and imaged object, and it has been shown for some applications that effective scatter control can be achieved [12]–[19]. To combine the strengths of different types of correction methods, hybrid approaches are also often used [9], [20], [21], which could provide well-balanced tradeoffs among correction effectiveness, calculation time, exposure loss, dose increase etc. To improve the performance of scatter correction, a new method is proposed here. We insert a calibration sheet with spatially variant attenuating materials between the X-ray source and the object to modulate the primary distribution and make scatter and primary distributions strongly separate in the Fourier domain. With subsequent image processing techniques, the scatter distribution of the modified X-ray system is estimated accurately and artifacts are substantially reduced in the reconstructed images. This method provides scatter correction using a single scan acquisition without the loss of real-time imaging capability. The concept of “primary modulation” makes this method distinct from the current scatter correction methods, and provides advantages in implementation as well as high scatter estimation accuracy. In this paper, we describe the system geometry and the scatter correction algorithm; the key hypothesis that the high-frequency components of the X-ray source distribution do not result in strong high-frequency signals in the scatter distribution is validated using Monte Carlo (MC) simulations; then we evaluate the performance of the correction method using MC simulations on a software humanoid phantom, and preliminary physical experiments are carried out on our tabletop CBCT system to provide an experimental verification of the algorithm. II. METHOD AND MATERIALS The methodology of scatter correction using primary modulation is first presented in this section in a heuristic manner. Assumptions are made in the development of the algorithm, and their accuracy will be validated by MC simulations. The evaluation of the algorithm will be presented in a later section of experiment results. A. Concept of Primary Modulation In the absence of scatter correction, Fig. 1 conceptually depicts the regions in the Fourier domain where the main energy of the scatter and primary distributions of one X-ray projection is located. The contribution of Rayleigh scattering is small in the scatter distribution [22]. If only Compton scatter is considered, low-frequency components dominate in the scatter distribution 0278-0062/$20.00 © 2006 IEEE 1574 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 25, NO. 12, DECEMBER 2006 Fig. 1. Conceptual illustration of typical primary and scatter distributions in the Fourier domain, before correction. The encompassed regions are where most (not all) of the primary or scatter energy is located. (more discussion later), while the primary distribution, closely related to the object, is more arbitrary but still has strong lowfrequency components. The supports of the scatter and primary distributions overlap completely, while a fraction of Fourier space is left almost blank. From an information channel point of view, a better use of this “channel” would be reallocating the primary and scatter distributions such that more primary resides in the region of Fourier space where there is no or less scatter (“water-filling” scheme as in information theory [23]). Manipulating the scatter distribution is difficult because of the complexity in the physics of the scattering process. In fact, we assume that adding high-frequency components in the X-ray source distribution does not result in strong high-frequency components in the scatter, even though it might change the low-frequency content. By contrast, if the beam-hardening effects are negligible, the detected primary distribution is linearly proportional to the incident X-ray distribution. Therefore, an X-ray source distribution with high-frequency components is able to “code the primary” before projection. Equivalently, the original primary distribution acquired using a uniform X-ray source distribution is partially modulated by a high-frequency function defined by the new X-ray source distribution, while the scatter distribution is still predominantly low-frequency due to the hypothesis. The information content of the modulated primary distribution reaches the detector without much contamination of scatter, and the scatter distribution can, therefore, be estimated and corrected. The hypothesis relating to the scatter is the key to the algorithm development, and further discussion and verification are presented in Sections II-F and III. One easy way to alter the incident X-ray spatial intensity distribution is to insert between the X-ray source and the object a calibration sheet with attenuating materials—a “primary modulator.” In order to introduce high-frequency components, the primary modulator consists of semi-transparent blockers that are arranged in a “checkerboard” pattern, i.e., semi-transparent and “transparent” blockers alternate. Fig. 2 shows the system geometry with the insertion of the primary modulator. Based on the main idea presented above, scatter correction of this modified imaging system can be implemented. Note that all of the data analysis and processing are done before the log operation on the projection data that calculates the images of the attenuation coefficient integrals along the projection lines (referred to Fig. 2. Geometric configuration of the X-ray system with the insertion of the primary modulator. This particular primary modulator has a “checkerboard” pattern as does its shadow on the detector. The projection data are downsampled at the centers of high and low intensity regions before the log operation that calculates the line integrals, and only these data are used in the scatter estimation and correction. as line integral images hereafter), and the images are processed independently for each projection angle. To avoid the edge effect of the blocker shadows, only the data at the centers of high and low intensity regions on the detector are used in the scatter estimation and correction, i.e., the projection image is downsampled with a sampling period equal to the diameter of the blocker shadow (see Fig. 2). Parameters and are the horizontal and vertical indices of the downsampled data, and and are the downsampled primary distributions with and without the primary modulator. Denoting the photon transmission of the semi-transparent blockers as , we have (1) as the discrete Fourier transform (DFT) of Denote and and , respectively, and as the coordinate system in the Fourier domain. Taking the DFT of both sides of (1), one obtains (2) Equation (2) shows that part of the primary is replicated (modulated) to the high-frequency region in Fourier space, . Notice, on the other hand, that with weighting if is not very small, i.e., the high-frequency components of the incident X-ray spatial distribution are not strong, low frequencies still dominate in the scatter distribution, due to the hypothesis discussed above. With the primary modulator in place, the Fourier domain distributions of primary and scatter are shown in Fig. 3. Now, the primary is more separately distributed, and it provides a possibility for accurate scatter estimation. Since the low-frequency ZHU et al.: SCATTER CORRECTION METHOD FOR X-RAY CT USING PRIMARY MODULATION Fig. 3. Conceptual illustration of the primary and scatter distributions in the Fourier domain, with the primary modulator in place. The solid line indicates the primary distribution before modulation; the dot-dashed line indicates the modulated primary; the dashed line around the origin indicates the scatter distribution, which is mainly concentrated in the low-frequency region before and after primary modulation; the center region encompassed by the dotted line indicates the support of the low-pass filter used in Step 3.3 of the scatter correction algorithm proposed in Section II-D; the shaded region indicates the support of the high-pass filter used in Step 3.4. region is the only region where primary and scatter distributions are overlapping, knowledge of the low-frequency compo(encompassed by the dotted line around nents of the primary the origin) suffices to extract the scatter distribution, while the low-frequency primary can be estimated from its replica in the high-frequency region (shaded area). The estimation accuracy of the low-frequency primary spatial spectrum from the high-frequency components is crucial in this approach. The Fourier domain distribution of the projection image before modulation (solid line in Fig. 3) should not heavily contaminate the region of high-pass filtering (shaded area in Fig. 3). Otherwise, the high-frequency spectrum of the initial projection is demodulated together with the modulated low-frequency spectrum, making the low-frequency primary estimation inaccurate. This requires that: 1) the initial primary before modulation does not have large high-frequency components; or 2) the modulation frequency is high relative to the major frequency content of the primary, and modulation weight is large. The latter requires careful choice of the system parameters, while the former can be improved by preprocessing techniques on the projection images. B. Auxiliary Boundary Detection As discussed above, the algorithm will generate relatively large scatter estimation error in those regions of image that have large high-frequency components. In the projection images before the log operation, most of the high-frequency components of the primary distribution result from the sharp transitions at the object boundaries. Since linear filtering techniques are used in the data processing, the resulting error in the scatter corrected reconstructed images not only resides at the object boundary but also spills into the inside of the object where the reconstruction is more sensitive to scatter estimation error due to the small primary signal. An auxiliary boundary detection algorithm is, therefore, introduced to improve the scatter correction. The main idea is to premultiply the initial scatter-plus-primary projections by a two-dimensional (2-D) boundary profile func- 1575 Fig. 4. The cutoff functions. The hard cutoff function hits a threshold when s has a value close to but less than q . The soft cutoff function uses an exponential function to smooth the transition. tion , to smooth out the sharp transitions at the edges of the object. The boundary profile function smoothly decreases from 1 to a small value around the boundaries going from the interior towards the exterior of the projection of the object, with the sharpest slope around the object boundary. It would seem reasonable to divide the calculated scatter estimate by to reverse the effect of premultiplication. However, based on experimental results, we found that the estimation error is boosted by the division by , where the value of is small; on the other hand, the large value of the primary outside the boundary makes the scatter estimation there less demanding. Therefore, the division step is omitted in the implementation. The boundary profile function depicts the shape of the object boundary on the projection image. The estimation of this function, in general, is complicated, especially without prior information of the object. Fortunately, the objects scanned in a clinical setting usually have simple quasi-cylindrical or quasiellipsoidal shapes. The boundary profile can, therefore, be estimated based on the first order derivative of the projection image. A 2-D Gaussian filter is then applied to smooth the transition. Note that the boundary detection is not crucial to the overall algorithm performance, especially when system parameters are chosen to accommodate more primary high-frequency components in the scatter estimation. From physical experiments, we also found that a very rough estimation of the boundary profile is sufficient to improve the image quality. Although the simple derivative-based estimation is error-prone in the presence of noise, the scatter correction algorithm is still stable. The details of the implementation of the auxiliary boundary detection and its effect on the scatter correction performance is presented in Section IV. C. The Soft Cutoff Function A second issue of importance to the scatter correction algorithm is the substraction of the estimated scatter from the measured projection data. The log operation that follows implicitly requires that the estimated primary be positive. A cutoff function, therefore, is needed to guarantee that the estimated scatter is always less than the measured projection data. Fig. 4 shows the cutoff functions. The abrupt transition of a hard cutoff function with simple thresholding results in streaking artifacts, especially when the scatter estimation is inaccurate. To make the 1576 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 25, NO. 12, DECEMBER 2006 scatter correction algorithm more tolerant to estimation error, a smooth soft cutoff function is used with an exponential function to smooth the transition. The exponent is chosen to make the first order derivative continuous. as the measured projection data, as the estiDenote as the value that will be subtracted from mated scatter, and . The relationship between and is defined by the soft cutoff function as: (3) where the parameter , and is chosen empirically. If the scatter estimation is accurate, a large can be used; otherwise, a small should be chosen to suppress the overshoot error when the scatter estimate is inaccurate and very large. D. The Scatter Correction Algorithm With the primary modulator in place, an efficient scatter correction can be obtained for this modified system. All of the above observations lead to the following steps for the new scatter correction algorithm, which can be embedded in the image processing stage of conventional X-ray CT. • Step 1: Do two prescans with and without the primary modulator, with no object, to obtain an estimate of the blocker shadow position on the detector and the blocker transmission . • Step 2: Acquire projection images of the object with the primary modulator in place. • Step 3: For each projection image (could be processed in parallel with step 2): — Step 3.1 estimate the boundary profile using the auxiliary boundary detection algorithm (in the estimation, use the prescan result to cancel the effect of the blockers on the projection image), multiply the projection image by ; — Step 3.2: downsample the image at the centers of the blocker shadows; — Step 3.3: low-pass filter the downsampled image to obtain an estimate of scatter plus low-frequency primary; — Step 3.4: estimate the high-frequency components of the modulated primary by high-pass filtering the initial downsampled projection (the high-pass filter has the same size of support as the low-pass filter in step 3.3), then demodulate to the low-frequency signal, and ; this provides an estimate of weight it by low-frequency primary; — Step 3.5: subtract the result obtained in step 3.4 from that in step 3.3; — Step 3.6: upsample to obtain a scatter estimate of the whole field, apply the soft cutoff function. — Step 3.7: subtract the scatter estimate from the initial projection image to obtain a primary estimate, and divide the result by the flat field distribution with primary modulator in place that is measured in the prescan. Take the negative log value. • Step 4: Reconstruct the object using the processed projection images. (The Fourier domain illustration of steps 3.3–3.5 is shown in Fig. 3) Considerations in the practical implementation of the algorithm and a step-by-step example are presented in Section IV. E. Choice of System Parameters While the boundary detection and the soft cutoff function improve the scatter correction, algorithm performance is mainly determined by the choice of system parameters. Recall that for improved scatter estimation, contamination of the high-pass filtering region in Step 3.4 by the unmodulated initial primary distribution should be minimized. Based on this, the choice of proper system parameters is discussed below, although the optimal values are still to be determined by simulations and/or physical experiments. 1) Blocker Size : The blocker size determines the pitch of the data samples that are used in the scatter correction, and hence the primary modulation frequency. To make the modulated and unmodulated primaries strongly separate, a small is preferred. On the other hand, in a practical CT system, cannot be too small because of the penumbra effects from the finite size of the X-ray focal spot. 2) Blocker Transmission : With the primary modulator in place, the ratio of modulated primary to unmodulated primary . A smaller is needed for a larger primary is modulation and, therefore, a better scatter correction. However, decreasing , or increasing the blocker attenuation also results in stronger high-frequency components in the primary, which would increase the high-frequency components in the scatter and, therefore, might undermine the critical assumption that the scatter profile is still dominated by low-frequency components even if a calibration sheet of high-frequency modulation is inserted. : (The 3) Cutoff Frequency of the Low-Pass Filter .) cut-off frequency of the high-pass filter is is mainly determined by the scatter distribution. It should be chosen large enough to cover most of the extent of the scatter also increases the in the Fourier domain, although a large risk that unmodulated high-frequency primary is demodulated back as an estimate of low-frequency primary. To make the for the two algorithm more general, we use the same orthogonal directions of the 2-D image. However, a combivalues for different directions could be used nation of in the algorithm, as long as the support of the low-frequency filter and its corresponding high-frequency filter does not have overlapping areas. The choice of the system parameters is a tradeoff among many factors. From a practical point of view, the blocker size and transmission deserve more consideration since they are hardware related and usually cannot be easily changed. We use simulations to find proper value ranges, with further refinement to be provided by physical experiments. F. Hypothesis of Scatter Distribution The key hypothesis in the primary modulation method is that the high-frequency components of the X-ray spatial distribution do not result in strong high-frequency signals in the scatter. In ZHU et al.: SCATTER CORRECTION METHOD FOR X-RAY CT USING PRIMARY MODULATION other words, for the algorithm to work, low-frequency changes in the scatter after the insertion of the modulator are acceptable, but the insertion should not introduce much high-frequency signal in the scatter. This assumed property enables the primary to be manipulated without significantly expanding the scatter spatial spectrum in the Fourier domain, and makes the separation of primary and scatter possible. X-ray scattering is a stochastic multi-interaction process that is dependent on the imaging geometry and the object, and an analytical validation of the main hypothesis of this algorithm is complicated. A MC simulation is, therefore, used to provide justification in Section III. However, two additional arguments support the hypothesis. Because scattering is a random process that removes the spatial information from the initial signal, the high-frequency content of the incident X-ray source distribution will be lost during the scattering process. The hypothesis could also be explained by the simplified analytical models of scattering as used in many convolution based scatter estimation algorithms [12], [13], [16] [17], [20], [21]. A typical example is in [12], where the point spread function (PSF) of the single scatter on a homogeneous object is derived analytically based on the Klein-Nishina formula, which characterizes the behavior of Compton scattering accurately. The resultant equation is a function of object thickness, object-to-detector distance (air gap), and the linear attenuation coefficients of primary and scattered radiation. The PSFs for different parameter values are typically low-frequency responses. As the object thickness and air gap increase, the PSF becomes smoother and wider, i.e., more constrained to low frequencies. Since multiple scatter can be considered as consecutive single scatters, it is expected to have more low frequency components in its PSF. Therefore, we can conclude that Compton scattering has a low-frequency response to the X-ray source distribution. Since Rayleigh scatter consists mostly of small angle scattering, its contribution has high-frequency components. However, as also argued in [12], the contribution of Rayleigh scatter is considerably smaller than that of Compton scatter, and it does not degrade the effectiveness of the proposed scatter correction algorithm, as will be shown with simulated and measured data below. III. MC VALIDATION OF THE KEY HYPOTHESIS The MC simulations were carried out using Geant4 MC software [24]. The implementation of the code was verified by comparison with the SIERRA MC code [25], [26] for several simple geometries. Fig. 5 shows the simulation setup. We chose parameters of system geometry that are typical for C-arm CT. The MC simulation is time-consuming and the simulation time is proportional to the total number of incident photons; on the other hand, generating a scatter distribution with good quantum noise statistics requires a large number of incident photons per ray. To save computation, we used a relatively large pixel diameter on the detector, 1.56 mm. Two simulations were done to investigate the sensitivity of scatter to the spatial variation of the incident photon intensity. The first simulation used a uniform intensity distribution, with 3 100 000 photons per ray. The second used a distribution as shown in Fig. 5. The detector was divided into 1577 Fig. 5. The simulation setup for the validation of the key hypothesis. The first simulation uses a uniform incident photon intensity distribution on the detector. In the second simulation, the incident distribution is divided into strips in the horizontal direction with width of 12.5 mm on the detector, and every other strip has no incident photons (blacker stands for the region with no incident photons). strips, and every other strip had no incident photons. This represents the maximum high-frequency content that could be added into the incident photon intensity distribution, given the spacing of the strips. To avoid losing high-frequency information due to insufficient sampling, the strip width was set to 12.5 mm, or 8 pixels on the detector. The number of photons per ray was doubled to make the total number of incident photons the same. For both simulations, the X-ray source was monochromatic operating at 50 keV. It illuminated only the region of detector, and a uniform water cylinder of diameter 200 mm was used as the object. Shown in Fig. 6 are the one-dimensional (1-D) profiles taken at the central horizontal lines of the simulated scatter distributions (solid lines). The difference of the two scatter distributions could be found mainly around the object boundary mm on the detector), because the object attenuation ( changes rapidly in this region. The comparison also illustrates, in the second simulation, that although the X-ray source distribution has strong spatial high-frequency components, the resultant scatter distribution still has large low-frequency signals. The disturbance in the scatter due to the nonuniform X-ray distribution contains high-frequency components, as well as low-frequency signals that are acceptable in the hypothesis. Although not shown here, a comparison of the scatter profiles from Rayleigh and Compton indicates that most of the high-frequency discrepancy between scatter profiles with and without the strip pattern stems from Rayleigh scattering, which is more closely correlated with incident primary photon distribution due to its narrow PSF. To illustrate the impact of these high-frequency variations on the scatter estimation, the estimated scatter profiles are also shown in Fig. 6 (dashed lines). This data was analyzed using the steps of the proposed algorithm. The scatter was first downsampled at the center of the low and high intensity regions on the detector. To suppress the noise, the downsampled data was taken as the average of the neighboring 4 pixel values (a step included in the practical implementation, see Section IV-A). Then a low-pass filtering 1578 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 25, NO. 12, DECEMBER 2006 Fig. 7. The RMS of the relative estimation error inside the object boundary . The solid (from 159:4 mm to 159.4 mm on the detector) for different ! line: using uniform X-ray source intensity distribution; The dashed line: using nonuniform distribution as shown in Fig. 5. 0 the RMS of the relative error less than 0.8%. These results indicate that although high-frequency variation exists in the scatter distribution when the high-frequency components are present in the X-ray source distribution, the resultant error in the scatter estimation using low-passing filtering is small. In other words, the low-frequency signals still dominate in the scatter distribution. IV. SCATTER CORRECTION EXPERIMENTS Fig. 6. The central horizontal profiles of the simulated scatter distributions and the estimated scatter distributions. The estimation process uses similar steps as in the scatter correction algorithm, with the cutoff frequency of the low-pass filter ! = (=2). Note that although a relatively large estimation error can be found around the object boundary ( 159:4 mm), the high primary value there reduces its impact on the reconstructed image. (a) using uniform X-ray source intensity distribution. The RMS of the relative estimation error inside the object boundary (from 159:4 mm to 159.4 mm on the detector) is 0.48%. (b) using nonuniform X-ray source intensity distribution as shown in Fig. 5. The RMS of the relative estimation error inside the object boundary is 0.54%. 6 0 Both simulation and physical experiments are carried out to fully evaluate the performance of the algorithm. Since the goal of this study is to develop an effective scatter correction technique for clinical use, especially when scatter-to-primary-ratio (SPR) and its variation are both very high, simulations of chest imaging with a humanoid software phantom are used to investigate the effect of the choice of the algorithm parameters on the reconstructed image quality, and to suggest proper parameter values. Physical experiments on a standard evaluation phantom also provides a solid validation of the algorithm in a practical environment. A. Algorithm Implementation Details with the cutoff frequency was performed, and finally, the result was upsampled back to the original length. We focus on the estimation error of the scatter due to the change in the scatter distributions only, and the primary data was not used in the estimation of the scatter. Fig. 6(b) shows that the scatter estimate matches the original scatter distribution, with a relatively larger error at the object boundaries. Note, however, that if a semitransparent strip pattern is used, i.e., high-frequency components in the X-ray source distribution are reduced, the error will be reduced as well and in addition the high primary value will alleviate the impact of the scatter estimation error outside the object boundary. Further verification is provided by examining the root-meansquares (RMSs) of the estimation error relative to the simulated scatter, calculated for the data inside the object boundary mm to 159.4 mm on the detector. The RMSs of from Figs. 6(a) and 6(b) are 0.48% and 0.54%, respectively. The reused in the low-pass filtering are shown sults for different in Fig. 7. Small differences are seen for the two X-ray source , both distributions result in distributions. When Several details must be considered for a practical implementation of the algorithm proposed in Section II-D. In the MC simulations, the knowledge of the blocker pattern was assumed. In the physical experiments, the blocker transmission and the sampling rate were estimated using the prescans (Step 1). We determined the centers of the blocker shadows by maximizing the correlation between the flat field image of the modulator and a grid function. In Step 3.1, the boundary was estimated line by line along the direction parallel to the plane of the source trajectory. We took the first derivative of the projection image along that line, and then found the positions where the derivatives were the most positive and the most negative. This gives a rough estimate of the object boundary. We assigned 1 to the pixels within the boundary and 0.01 outside. Finally, a 2-D Gaussian filter was applied to the 2-D profile. The standard deviation of the Gaussian kernel was chosen empirically. From experiments, we found that the value could range within 5%–20% of the coverage of the object on the detector in the horizontal direction, and the scatter estimation error was insensitive to the choice. A constant ZHU et al.: SCATTER CORRECTION METHOD FOR X-RAY CT USING PRIMARY MODULATION Fig. 8. The Zubal phantom. (a) Three-dimensional view; (b) axial view; (c) coronal vew; (d) sagittal view. The shadow shows the scanned area in the simulations. TABLE I PARAMETERS OF THE SIMULATION EXPERIMENTS 1579 images are free of beam hardening artifacts. The standard FDK algorithm [28] was used for the reconstruction. We chose to do a chest scan (see Fig. 8), since in the lung region, the high SPR and SPR variation make scatter correction more challenging. The plane of the source trajectory was on slice 108 of the phantom. The scatter distributions were generated using the Geant4 MC software. Since generating noiseless scatter distributions using MC simulation is very time-consuming, the Richardson-Lucy (RL) fitting algorithm was used such that accurate and noiseless scatter distributions could be obtained using a much smaller number of photons [29]–[31]. The acceleration of the MC simulation by this algorithm stems from the fact that scatter distributions are very smooth (low-frequency), so the high-frequency statistical noise in the simulation of relatively few photons can be removed after curve fitting. The primary projections were calculated separately using line integration [32] and weighted to match the SPR. Denoting and as the scatter and primary as the scatter distributions obtained from MC simulation, distribution after RL fitting, and as the primary distribution for each by line integral calculation, the weight factor K on projection is computed as follows: (4) standard deviation, 40 mm, was used for all the simulations and for the physical experiments. The above algorithm works for quasi-cylindrical objects as were used in this paper. For other types of objects, the algorithm could be developed similarly. To make the algorithm more robust to noise, in Step 3.2, the downsampled data was the average of a small neighborhood (less than the blocker shadow size) around the sample point. The simulation experiments were noise-free, and downsampling without averaging was used; in the physical experiments, we averaged an area of 10-by-10 pixels on the detector. Filtering techniques were used in several steps of the algorithm (Step 3.3 and Step 3.4), including the upsampling step (Step 3.6). To suppress the error from filtering a finite-length signal, we applied Hamming windows on the filters with the same width, and the images were zero-padded to twice signal length along both axes. To make the algorithm more general, in the two orthogonal we chose the same cutoff frequency directions of the image. in the soft cutoff function of Step 3.6. We used Even with the acceleration, MC simulations are still computationally intense. For the purposes of saving computation, we assumed that the error in the scatter estimation due to the insertion of the primary modulation was small (supported by the MC simulations in Section III), and the same scatter distributions obtained without the modulator were used for the simulations. In the implementation of the primary modulation method, a perfect knowledge of the blocker shadow position was also assumed. The simulations do not include quantum noise; the effect of noise is addressed in the physical experiments described in Section IV-C. 2) Determination of the Reconstruction Accuracy: Side-byside image comparisons are used to show the improvement of reconstruction accuracy using our scatter correction approach. In addition, the relative reconstruction error (RRE) is also used as a quantitative measure. Denoting the reconstructed image with scatter correction as , and the ideal reconstruction without scatter as , the RRE is defined as relative mean square error of the reconstructed images in the region of interest (ROI) B. Simulation Experiments 1) The Zubal Phantom and MC Simulation: The Zubal phantom [27] was used in the simulation experiments, shown in Fig. 8. It is a humanoid software phantom from head to hip, with total size 128-by-128-by-243 and 4-mm resolution. The simulation and reconstruction parameters are summarized in Table I. Note that a large detector was used to avoid truncation of the projection images and, therefore, the scatter was exaggerated as compared to a practical implementation of the CBCT system. The X-ray source was monochromatic, so that the reconstructed % (5) are the coordinates of the reconstructed image. where The ROI is chosen as the central reconstructed volume of size 128-by-128-by-64. Parameters and are in Hounsfield units (HU). To make the reconstruction linear rather than affine, the value is shifted by 1000 in the RRE calculation, such that air is 0 HU and water is 1000 HU. 1580 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 25, NO. 12, DECEMBER 2006 Fig. 10. The line integral images (after the log operation) with and without scatter correction using primary modulation, and difference images calculated relative to the scatter free images. The parameters used in the scatter correc= (=2), as used in tion are ds = 3:760 mm (5pixel), = :7; ! Fig. 13(b). Note a compressed display window is used for the difference image with scatter corrected using primary modulation. (a) Without scatter correction. Display window: [0 8]. (b) Difference image. Display window: [-4 0]. (c) With scatter correction using primary modulation. Display window: [0 8]. (d) Difference image. Display window: [-0.2 0.2]. Fig. 9. Axial, coronal and sagittal views of the reconstructed volumes, without scatter correction and with scatter corrected by different methods. The RRE values are shown with the images. Display window: [ 200 500]HU. (a) No scatter correction. (b) Scatter estimated by a prescan on an all-water object with the same boundary. (c) SPR suppressed by an anti-scatter grid, with primary transmission 64.2%, scatter transmission 6.3%, based on the technical data of the Siemens 12/40 type anti-scatter grid. (d) Ideal reconstruction without scatter present. 0 3) Reconstruction Without Scatter Correction and With Scatter Corrected by Other Methods: Fig. 9 shows the challenge of effective scatter correction. Fig. 9(a) is reconstructed without scatter correction. The cupping and shading artifacts caused by scatter are very severe in the image, and the RRE is 74.2%. We also estimate the scatter by a prescan on an all-water phantom with the same boundary, assuming the boundary of the object is known. The scatter artifacts are reduced slightly, but the overall image quality is still poor [Fig. 9(b)]. Image corruption is also found in Fig. 9(c), where an anti-scatter grid is used to suppress the SPR. The primary and scatter transmissions of the grid are 64.2% and 6.3%, respectively, based on the technical data of the Siemens 12/40 type anti-scatter grid for fluoroscopy imaging. The simulation uses a simple scaling on the scatter and primary. As a reference, Fig. 9(d) is the reconstruction with perfect scatter correction. 4) Reconstruction With Scatter Correction Using Primary Modulation: The scatter correction algorithm by primary modulation contains three parameters: the blocker size , the blocker transmission and the cutoff frequency of the low-pass filter . For simplicity, in the simulation results Figs. 10–15, we Fig. 11. The errors in the scatter corrected images of line integrals (after the log operation), using different schemes. The 1-D profiles are taken at the central horizontal lines of the projection images, and the view angle is the same as in Fig. 10. The object boundary is around 280 mm and 310 mm. The parameters = used in the scatter correction are ds = 3:760 mm (5pixel), = :7; ! (=2), as used in Fig. 13(b). 0 use the blocker shadow size on the detector in or detector pixels as a system parameter instead of the actual blocker size. The effect of the scatter correction in the projection space is demonstrated in Fig. 10, where the images after the log operation are shown. The difference images are calculated relative to the scatter-free images. Fig. 11 compares 1-D profiles taken at the central horizontal lines on the projection data with the same view angle. To illustrate the effect of the boundary detection algorithm, different schemes are used in the scatter correction: without boundary detection, with boundary detection and division by the boundary profile , and with boundary detection but the division step is omitted. Since linear filtering techniques are involved in the algorithm, if the boundary detection is not used, high-frequency errors that originate from the object boundary have a global effect on the estimated projection image. The line integral images are more sensitive to estimation errors where the primary is low, due to the log operation, and therefore, high-frequency errors appear in the middle of the profile, as shown in ZHU et al.: SCATTER CORRECTION METHOD FOR X-RAY CT USING PRIMARY MODULATION 1581 Fig. 12. Scatter corrected images using primary modulation, but without boundary detection. Display window: [ 200 500]HU. (a) ds = 3:760 mm = (=2), as used in Fig. 13(b). (b) ds = 5:264 mm (5pixel), = :7; ! (7pixel), = :5; ! = (=2), as used in Fig. 13(b). 0 Fig. 14. The 1-D profiles of different reconstructed images and the differences calculated relative to the scatter free reconstruction. The profiles are taken at the central vertical lines of the axial view (plane of source trajectory). The detailed description of the legend is as follows. no scatter correction: [Fig. 9(a)]; prescan correction: scatter estimated and corrected by the prescan [Fig. 9(b)]; anti-scatter grid: scatter suppressed by the anti-scatter grid [Fig. 9(c)]; primary modulation: scatter corrected using the primary modulation method, ds = = (=2) [Fig. 13(b)]; scatter free: ideal 3:760 mm (5 pixel), = :7; ! reconstruction without scatter present [Fig. 9(d)]. The profile obtained using primary modulation matches the scatter-free reconstruction in most of the area. (a) The reconstructed values. (b) Difference profiles calculated relative to the scatter free reconstruction. Fig. 13. Representative reconstructed images with the scatter correction using primary modulation. Display window: [ 200 500]HU. The blocker shadow size on the detector ds is used as the parameter, instead of the actual blocker size d. is the blocker transmission, and ! is the cutoff frequency of the low-pass filter used in the scatter correction algorithm. (a) ds = 2:256 mm (3pixel), = :5; ! = (=3). (b) ds = 3:760 mm (5 pixel), = :7; ! = (=2). = (=2). (d) ds = 5:264 mm (c) ds = 5:264 mm (7 pixel), = :5; ! (7 pixel), = :9; ! = (=3). 0 Fig. 11. As a result, the reconstructed image has high-frequency artifacts, typically ripples, around the object center. Fig. 11 indicates that the boundary detection algorithm is able to suppress these high-frequency errors effectively, at the price of estimation accuracy loss at the object boundary. Since the ROI is usually the interior of the object rather than the periphery, we use the boundary detection as a supplementary step in the algorithm. Our algorithm only includes the multiplication of the boundary profile which suppresses the sharp transitions of the boundary on the projection image. The overshoot error caused by the division by as the final step is also clearly shown in Fig. 11, therefore, this step is omitted in the algorithm. Fig. 12 shows the ripple effect that occurs in reconstructed images when no boundary detection is applied. The same sets of parameters were used as in Fig. 13(b) and (c), which are the scatter corrected images with the boundary detection. The comparison shows that more pronounced high-frequency artifacts are present in Fig. 12. Representative reconstructed images with scatter corrected using the proposed algorithm with different system parameters are shown in Fig. 13. With proper parameter values, the scatter artifacts are significantly suppressed, and the RRE value can be reduced to less than 1%. Fig. 14 compares the 1-D profiles of Fig. 9 and the scatter correction result of Fig. 13(b). The profiles are taken at the central vertical lines of the axial views. The comparison reveals that the scatter correction is accurate in most of the area, with a small error mainly located around the object boundary, due to the imperfection of the boundary detection. 1582 Fig. 15. The RRE values of the reconstructions using different system parameters. Parameter ds is the blocker shadow size on the detector, in detector pixels, is the blocker transmission, and ! is the cutoff frequency of the low-pass = (=2) (b) ! = (=3). filter. (a) ! To investigate the effects of the system parameters on the algorithm performance, simulations were also done with various sets of the parameter values. The RRE values of these simulations are summarized in Fig. 15, and the results illustrate the effects of the system parameters as we discussed in Section II-E. If the position of the primary modulator is fixed, the blocker size can be calculated simply by dividing the corresponding by the magnification factor from the primary modulator to the detector. Decreasing or increases the scatter correction accuracy, since a higher sampling rate and a stronger modulation make scatter and primary more separable in the Fourier doshould be case-dependent. When the main. The choice of small or is chosen to provide accurate scatter correction, a large should be used. If the ability of scatter correction is should be chosen more limited by the large or , then conservatively. Fig. 15 also indicates that for an RRE below 3%, the blocker mm. Considering the magnifishadow size can be as large as cation factor of the modulator on the detector, the actual blocker mm, if the system geometry of the size on the modulator is physical experiments in Section IV-C is used. On the other hand, if a smaller is achievable, can be as high as 90%, in which case the exposure loss is only 5% (an issue for systems with limited X-ray tube output). Note, however, that a simplified model is used in the simulations, and the performance of the algorithm will be limited in practice by the nonideal effects of the system, such as penumbra effect due to the finite size of the focal spot, beam hardening effect etc. Therefore, the determination of and is based on considerations of the system physical limitations as well; the choice does not change the physics of X-ray projection, and it of could be optimized as a final step. The impact of the nonideal effects on the algorithm performance with the suggested system parameters is currently under investigation. C. Physical Experiments The simulation experiments used a simplified model, mainly due to the computation complexity of MC simulation. Many issues that are important to practical implementations were not considered, such as the algorithm robustness to noise, effect of beam hardening on the algorithm etc. In particular, since a relatively low resolution (4 mm) phantom was used in the simulations, the resolution quality of the scatter corrected IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 25, NO. 12, DECEMBER 2006 Fig. 16. The primary modulator used in the physical experiments. Fig. 17. The blocker shadow pattern on the detector and the downsample grid. The center of the blocker shadow and its spacing were estimated from the prescan results. To suppress the effect of noise in the scatter estimation, the downsampled data were taken as the average of the neighboring 10-by-10 detector pixel values. reconstructed image merits a closer examination. To provide additional validation of the proposed scatter correction algorithm, preliminary physical experiments were carried out on the tabletop CBCT system in our lab. 1) The Primary Modulator: Our first primary modulator was machined from aluminum. The geometry of the blockers is shown in Fig. 16. The blocker pattern is different from the checkerboard pattern, due to manufacturing considerations. However, all the theories still apply, since we only process the downsampled data and the shape of blocker does not matter (see Fig. 17). The side length of the blocker square is 2 mm, its thickness is 1 mm, and the base of the modulator is 2 mm. Therefore, the downsampling period on the detector in Step mm times the magnification factor 3.2 of the algorithm is from the primary modulator to the detector. The transmission of the blocker is approximately 90% when the X-ray source operates at 120 kVp. Note that, a larger blocker size is used, as compared to those used in the previous section. This is because the ratio of the blocker thickness to diameter must be small to avoid edge effects, while the blocker thickness cannot be smaller in order to keep the blocker transmission less than 90%. The low sampling rate of the blocker pattern limits the scatter correction capability when SPR and SPR variation are both very high. A more attenuating material could be used to provide a primary modulator with a higher sampling rate. 2) The Tabletop System and Evaluation Phantom: The tabletop CBCT system consists of a CPI Indico 100 100 kW programmable X-ray generator (CPI Communication & Medical Products Division, Georgetown, ON, Canada), a Varian G-1590SP X-ray tube (Varian X-ray Products, Salt Lake City, UT), a rotation stage, a Varian PaxScan 4030CB flat panel a-Si ZHU et al.: SCATTER CORRECTION METHOD FOR X-RAY CT USING PRIMARY MODULATION 1583 Fig. 19. The Catphan600 phantom. Fig. 18. The tabletop cone beam CT system with the insertion of the primary modulator. TABLE II PARAMETERS OF THE PHYSICAL EXPERIMENTS large area X-ray detector, a workstation and shielding windows. The X-ray tube had an inherent filtration of 1.0 mm Al, and operated with a 0.6 mm nominal focal spot size. We mounted the primary modulator on the outside surface of the collimator (see Fig. 18), with a nominal distance to the X-ray focal spot of 231.0 mm. No anti-scatter grid was used in the experiments. The object was put on the stage, and it rotated during the scan to acquire data for different projection angles. The imaging and reconstruction parameters are summarized in Table II. Note was used in the experiments to that a relatively large compensate for the extra filtration due to the 2 mm aluminum base of the modulator. Again, the FDK algorithm was used in the reconstructions, with the same Hamming-windowed ramp filter. A standard evaluation phantom, Catphan600 (The Phantom Laboratory, Salem, NY), was used to investigate the performance of the scatter correction algorithm (see Fig. 19). A detailed description of the phantom can be found at: http://www. phantomlab.com/catphan.html. In order to show the possible impact on the image resolution of the algorithm, the plane of source trajectory was selected to coincide with the slice of the phantom that contains the resolution test gauge. The resolution test objects are located at a radius of 5 cm, and range in size from 1 to 21 line pairs per cm. 3) Preliminary Results: The prescans show that the blocker transmission has a mean value of 0.899 with a small variance. In the algorithm implementation, we assume that the transmission is the same for different blockers and the mean value is used as Fig. 20. Diagrammatic illustration of the scatter correction on a projection image of the Catphan600 phantom. The horizontal direction is parallel to the plane of source trajectory. Step 3.1: suppress the boundary; Step 3.2: downsample; Step 3.3: low-pass filter; Step 3.4: high-pass filter, demodulate and weight; Step 3.5: subtract; Step 3.6: upsample and soft cutoff; Step 3.7: calculate scatter corrected line integrals. (see Section II-D for detailed description of the algorithm). The downsampled data at the edges of the detector (partially outside the field of view) were not used in the processing chain. the parameter . The pattern of the blocker shadows obtained in the prescan also indicate a sampling period of 13.5 mm. For the scatter correction results shown below, we use as the cutoff frequency of the filter. A step-by-step illustration of the scatter correction algorithm on the projection images of the Catphan600 phantom is shown in Fig. 20, and the detailed description of each step is presented in Section II-D. To provide a fair comparison, another experiment is also carried out with a 2 mm aluminum filter on the collimator, the same thickness as the base of the primary modulator, but without the modulator and the scatter correction. Fig. 21(a) and (b) compare the two results, where the axial views are chosen to include the resolution test objects. As a reference, Fig. 21(c) is the same slice reconstructed from the projection data acquired with a narrowly opened X-ray source collimator with the aluminum filter as used in Fig. 21(a), and the beam width on the detector is about 10 mm. This imaging setup resembles a slot-scan geometry and the scatter is inherently suppressed [33], [34]. Note that a relative large display window is used to examine the visibility of the intense test objects. The comparison shows that the scatter is corrected effectively by the proposed algorithm using primary 1584 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 25, NO. 12, DECEMBER 2006 Fig. 22. One-dimensional profile comparison of Fig. 21, taken along the arc passing through the center of the small test objects. The angle of the arc starts from (=12) to (7=12), as shown in Fig. 21. Fig. 23. One-dimensional profile comparison of Fig. 21, taken at the central vertical lines. Fig. 21. Experimental results of the Catphan600 phantom. Display window: [0 0.07] mm . The white squares in the image indicate where the mean values and noise variances are measured. The arcs indicate the six sets of line pairs, of which 1-D profiles passing through the center are compared in Fig. 22. Note that the noise properties are similar in other scatter correction methods that do not suppress the statistical noise in the scatter distribution (see Section IV-C4). (a) No scatter correction, collimator widely opened; In the selected area (white square), mean value: 0.0148; noise variance: 2:180 10 . (b) Scatter corrected using the primary modulation method, with widely opened collimator; In the selected area, mean value: 0.0222; noise variance: 2:536 10 . (c) Scatter suppressed with a narrowly opened collimator ( 10 mm width on the detector); In the selected area, mean value: 0.0217; noise variance: 8:597 10 . 2 2 2 modulation. Referencing Fig. 21(c) as a “scatter-free” image, the primary modulation method reduces the reconstruction error in the central ROI (white squares) from 31.8% to 2.3%. Although the noise of the reconstructed image is changed (more discussion later), there is no noticeable impact on the resolution quality using the primary modulation. This conclusion is also supported in Fig. 22, which compares the 1-D profiles of the images in Fig. 21, taken along the arc passing through the center of the test objects with the focus at the center of the images. The angular sampling rate is chosen such that the arc length between the samples is 0.0977 mm, one fourth of the reconstructed voxel size. to (indicated by the white The angle starts from arcs in Fig. 21), one set of line pairs past the resolution supported mm at a radius of by the angular sampling of the system, 5 cm. The algorithm performance with respect to resolution is expected, since scatter is corrected by estimating low-frequency primary and, therefore, only low-frequency misestimation of primary is possible and the sharp transitions (high-frequency) between objects are preserved. A formal resolution study with a modulation transfer function (MTF) measurement is underway. To provide further verification of effective scatter correction using our proposed algorithm, Fig. 23 shows the 1-D profiles ZHU et al.: SCATTER CORRECTION METHOD FOR X-RAY CT USING PRIMARY MODULATION taken at the central vertical lines of the images in Fig. 21. A good match is found between the result with scatter corrected using primary modulation and that with scatter greatly suppressed by a narrowly opened collimator. Both Figs. 22 and 23 reveal that most of the difference is located in the dense objects and in their surrounding areas. This could be due to the residual scatter in the slot-scan geometry which results in more reconstruction error in the highly attenuating areas, the relatively large scatter estimation error around sharp transitions in the projection image, the beam-hardening effects that are not considered in the algorithm design etc. 4) Noise Analysis: Change of noise level is found in the selected ROI of reconstructed images in Fig. 21. This is because in the scatter estimation, low-frequency behavior of the scatter is assumed. Although the noise-free scatter distribution is usually smooth, the statistical noise of the scatter basically contains high-frequency signal and cannot be estimated using the proposed algorithm. The noise of scatter, therefore, is left in the primary estimate, increasing the noise in the reconstructed image. This problem is not specific to our proposed algorithm, and similar noise performance can be found in the reconstructed images of other scatter correction algorithms, as long as the same assumption of low-frequency scatter is made, and only a smooth (low-frequency) scatter estimate is removed from the recorded image. The mathematics of the noise analysis of the projection data and reconstructed images can be found in classic textbooks [35]. There are two types of noise to be considered in the projection image. One is the image independent noise due to the electrical and roundoff error; the other is the image dependent noise due to the statistical fluctuation of the X-ray photons that exit the object. We assume the noise of the first type is small, and only the second type is considered here. Denote and as the random variables of the scatter and primary on the detector, respectively; and, and are poisson distributed with mean values and , i.e., and . From the fact that the PSF of the scatter is broad and smooth, we assume is independent of , and therefore, . Denote as the constant incident photon intensity. If no scatter correction is applied, the line integral image calculated by taking the negative log of the measured projection image can be written as (6) Denote and as the statistical noise of the scatter and primary, respectively, and express and as (7) Plug into (6), and expand at (8) Assume then the noise of 1585 is small, and ignore the high order term, , and its variance are approximated as (9) (10) Similarly, if the proposed scatter correction algorithm is apis plied, the average profile is corrected effectively, while mostly intact. The noise in the line integral image, , and its variance can be obtained as (11) (12) Therefore, the ratio of the variance of and is (13) As a benchmark, an ideal scatter correction suppresses both and . Using the same method as above, we have similar expressions of the noise, (14) (15) and (16) Equations (13) and (16) describe the quantitative relationship of the noise in the projection images of line integrals. The noise in the reconstructed image is the summation of the errors from different views, and the analysis is more involved. As a special case, in the physical experiments, we used a cylindrical phantom that is approximately circularly symmetric; therefore, the SPR distributions have small differences from view to view. Assuming the noise of the scatter and the primary are independent for different views, similar expressions of the noise variance of the reconstructed images around the center can be obtained as those of the projection images of line integrals, with an additional scaling determined by the filter, total projection number etc. (see [35] for detailed derivation). Therefore, if the same filter kernel is used in the reconstruction, then (13) and (16) provide a rough approximation of the noise relationships in the central region of the reconstructed images. Since the slot-scan geometry inherently suppresses the scatter, it is close to the ideal scatter correction. Using (13) or (16), and the measured noise variances of the ROIs around the center of the reconstructed images, as shown in Fig. 21, we can calculate the corresponding SPRs as 2.41 or 1.95, respectively. On the other hand, based on the scatter estimation (step 3.6), the SPR around the , central region of the projection image is estimated to be in good agreement with MC simulations of a water cylinder with the same diameter. Equations (13) and (16) are approximately consistent with the experimental measurements. 1586 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 25, NO. 12, DECEMBER 2006 V. DISCUSSION AND CONCLUSION In this paper, we introduce a new practical and effective scatter correction method by primary modulation. A primary modulator with attenuating blockers arranged in a “checkerboard” pattern is designed and inserted between the X-ray source and the object. A scatter correction algorithm involving linear filtering techniques is developed for the modified X-ray system, with the hypothesis that the insertion of the primary modulator does not result in strong high-frequency components in the scatter distribution and only the primary distribution is modulated. Note that, although in this paper we implemented the proposed scatter correction technique for X-ray CT, it can also be used in other X-ray imaging applications, as long as a primary modulator can be inserted between the X-ray source and the imaged object. Recently, a similar frequency modulation based method was proposed by Bani-Hashemi et al. [36]. They inserted a modulation grid between the object and the detector and used nonlinear filtering techniques to correct the scatter, based on the fact that the scatter and primary have very different responses to the modulation. In our method, it might be counter-intuitive that the modulator is placed where no scattering of the imaged object has occurred; however, a different rationale is used in the algorithm development. Our method is expected to be more dose efficient since in Bani-Hashemi’s method, the modulation grid between the object and detector will absorb information-carrying photons. The key hypothesis of the algorithm was validated using a MC simulation, and the algorithm performance was evaluated by both MC simulations and physical experiments. The results show that the proposed method is able to suppress the scatter artifacts significantly. The MC simulations on the software humanoid phantom suggest system parameters for the physical implementation and reveal that the relative reconstruction error can be reduced from 74.2% to less than 1%. The scatter correction ability is also demonstrated by physical experiments, where a standard evaluation phantom was used. The proposed method reduces the reconstruction error from 31.8% to 2.3% in the central ROI. The image comparison also shows that the scatter correction method has no noticeable impact on the resolution of the reconstructed image. This is because in our method, only a low-frequency scatter estimate is subtracted from the original projection image, and the high-frequency content is kept intact. This observation also makes our approach different from Bani-Hashemi’s method. In their published studies, it has been reported that their technique will always produce a filtered image having at best 0.41 of the maximum detector resolution when maximum scatter rejection is desired. An analysis of the noise in the estimated projection images is also carried out. Since the high-frequency statistical noise in the scatter distribution can not be removed by our method, the noise of the line inte, gral image using our method increases by a factor of as compared to the noise of the image using perfect scatter correction. This noise increase is generally true for all of the scatter correction algorithms that assume smooth scatter distributions, such as most of the model-based correction methods, although it can be avoided in the scatter suppression methods that use the property of the scatter incident angle, such as anti-scatter grid, air-gap etc. Our implementation uses a calibration sheet with a “checkerboard” pattern, and only the data at the center of blocker shadows are used in the scatter correction to avoid edge effects. Note, however, that other patterns can also be designed based on different modulation functions. For example, a high-frequency sinusoidal function is free of edges, and it can be used to avoid the downsample/upsample steps in the algorithm. The proposed approach has several advantages. First, requiring only an insertion of a stationary calibration sheet, this approach is easy to implement. Second, the semi-transparency of the blockers reduces the loss of exposure, and it enables sampling of the field distribution with a higher sampling rate, as compared to the measurement based correction using a beam stop array. As a result, the scatter estimation is expected to be more accurate. Third, a single scan is necessary to acquire both the primary and the estimate of the scatter, and the primary modulator is placed between the X-ray source and the object, so the dose required for an acquisition is the same as for a conventional scan with no anti-scatter grid. Fourth, since the primary is modulated, not completely blocked, real-time imaging algorithms could be implemented so that the visual effects of the checkerboard pattern are not seen during image acquisition. Fifth, this approach uses linear filtering techniques, therefore, the computation can be very efficient. Future work will include a formal resolution study with MTF measurements, and the algorithm will be further evaluated on the tabletop CBCT system, using a more sophisticated phantom and a finer primary modulator. A hybrid method using primary modulation in combination with an anti-scatter grid will also be investigated. ACKNOWLEDGMENT The authors would like to thank S. Yoon for providing the forward projection code. They would also like to acknowledge Prof. J. Boone for insightful discussions and generous help with the MC simulation. REFERENCES [1] A. Malusek, M. Sandborg, and G. Carlsson, “Simulation of scatter in cone beam CT: Effects on projectin image quality,” Proc. SPIE, vol. 4320, pp. 851–860, 2001. [2] X. Tang, R. Ning, R. Yu, and D. Conover, “Investigation into the influence of X-ray scattering on the imaging performance of an X-ray flat panel imager (FPI) based cone beam volume CT (CBVCT),” in Proc. SPIE, 2001, vol. 4320, pp. 851–860. [3] J. Siewerdsen and D. Jeffray, “Cone-beam computed tomography with a flat-panel imager: Magnitude and effects of X-ray scatter,” Med. Phys., vol. 28, pp. 220–231, 2001. [4] P. Joseph, “The effects of scatter in X-ray computed tomography,” Med. Phys., vol. 9, pp. 464–472, 1982. [5] G. Glover, “Compton scatter effects in CT reconstructions,” Med. Phys., vol. 9, pp. 860–867, 1982. [6] J. Sorenson and J. Floch, “Scatter rejection by air gap: An empirical model,” Med. Phys., vol. 12, pp. 308–316, 1985. [7] U. Neitzel, “Grids or air gaps for scatter reduction in digital radiography: A model calculation,” Med. Phys., vol. 19, pp. 475–481, 1992. [8] J. Y. Lo, C. E. Floyd, J. A. Baker, and C. E. Ravin, “Scatter compensation in digital chest radiography using the posterior beam stop technique,” Med. Phys., vol. 21, pp. 435–443, 1994. [9] R. Ning, X. Tang, and D. Conover, “X-ray scatter correction algorithm for cone-beam CT imaging,” Med. Phys., vol. 31, no. 5, pp. 1195–1202, 2004. ZHU et al.: SCATTER CORRECTION METHOD FOR X-RAY CT USING PRIMARY MODULATION [10] R. Ning, X. Tang, R. Yu, and D. Conover, “X-ray scatter suppression algorithm for cone beam volume CT,” in Proc. SPIE, 2002, vol. 4682, pp. 774–781. [11] L. Zhu, N. Strobel, and R. Fahrig, “X-ray scatter correction for cone-beam CT using moving blocker array,” Proc. SPIE, vol. 5745, pp. 251–258, 2005. [12] J. Boone and J. Seibert, “An analytical model of the scattered radiation distribution in diagnostic radiology,” Med. Phys., vol. 15, no. 5, pp. 721–725, 1998. [13] J. Seibert and J. Boone, “X-ray scatter removal by deconvolution,” Med. Phys., vol. 15, pp. 567–575, 1988. [14] C. E. Floyd, A. H. Baydush, J. Y. Lo, J. E. Bowsher, and C. E. Ravin, “Scatter compensation for digital chest radiography using maximum likelihood expectation maximization,” Investigat. Radiol., vol. 28, pp. 427–433, 1993. [15] M. Honda, K. Kikuchi, and K. Komatsu, “Method for estimationg the intensity of scattered radiation using a scatter generation model,” Med. Phys., vol. 18, no. 2, pp. 219–226, 1991. [16] J. Wiegert, M. Bertram, G. Rose, and T. Aach, “Model-based scatter correction for cone-beam coputed tomography,” in Proc. SPIE, 2005, vol. 5745, pp. 271–282. [17] D. G. Kruger, F. Z. W. W. Peppler, D. L. Ergun, and C. A. Mistretta, “A regional convolution kernel algorithm for scatter correction in dualenergy images: Comparison to single-kernel algorithms,” Med. Phys., vol. 21, pp. 175–184, 1994. [18] M. Bertram, J. Wiegert, and G. Rose, “Potential of software-based scatter corrections in cone-beam volume CT,” Proc. SPIE, vol. 5745, pp. 259–270, 2005. [19] S. Molloi, Y. Zhou, and G. Wamsely, “ Scatter-glare estimation for digital radiographic systems: comparison of digital filtration and sampling techniques,” IEEE Trans. Med. Imag., vol. 17, no. 6, pp. 881–888, Dec. 1998. [20] L. A. Love and R. A. Kruger, “Scatter estimation for a digital radiographic system using convolution filtering,” Med. Phys., vol. 14, no. 2, pp. 178–185, 1987. [21] Y. Kriakou, P. Deak, T. Riedel, L. V. Smekal, and W. A. Kalender, “Combining deterministic and monte-carlo methods in the simulation fo X-ray attenuation and scatter,” Eur. Radiol., vol. 15, no. 1, p. 306, 2005. 1587 [22] J. M. Boone and J. A. Seibert, “Monte carlo simulation of the scattered radiation distribution in diagnostic radiology,” Med. Phys., vol. 15, no. 5, pp. 713–720, 1988. [23] T. Cover and J. A. Thomas, Elements of Information Theory. New York: Wiley, 1991. [24] S. A. , “Geant4: A simulation toolkit,” Nucl. Instrum. Meth. A, vol. 506, pp. 250–303, 2003. [25] J. Boone, M. Buonocore, and V. Cooper, “Monte carlo validation in diagnostic radiological imaging,” Med. Phys., vol. 27, pp. 1294–1304, 2000. [26] J. Boone and V. Cooper, “Scatter/primary in mammography: Monte carlo validation,” Med. Phys., vol. 27, pp. 1818–1831, 2000. [27] G. Zubal, G. Gindi, M. Lee, C. Harrell, and E. Smith, “High resolution anthropomorphic phantom for Monte Carlo analysis of internal radiation sources,” in Proc. 3rd Annu. IEEE Symp. Computer-based Medical Systems, Jun. 1990, pp. 540–546. [28] L. A. Feldkamp, L. C. Davis, and J. W. Kress, “Practical cone-beam algorithm,” J. Opt. Soc. Am., vol. A 1, pp. 612–619, 1984. [29] A. P. Colijn and F. J. Beekman, “Accelerated simulation of cone-beam X-ray scatter projections,” IEEE Tran. Med. Imag., vol. 23, no. 5, pp. 584–590, May 2004. [30] W. Richardson, “Bayesian-based iterative method of image reconstruction,” J. Opt. Soc. Amer., vol. 62, pp. 55–59, 1972. [31] L. Lucy, “An iterative technique for the rectification of observed distributions,” Astronom. J., vol. 74, pp. 745–754, 1974. [32] R. L. Siddon, “Fast calculation of the exact radilological path for a three-dimensional CT array,” Med. Phys., vol. 12, no. 2, pp. 252–255, 1985. [33] J. Boone, J. Seibert, C. Tang, and S. Lane, “Grid and slot scan scatter reduction in mammography: Comparison by using monte carlo techniques,” Radiology, vol. 222, no. 2, pp. 519–527, Feb. 2002. [34] E. Samei, “Comparative scatter and dose performance of slot-scan and full-field digital chest radiography systems,” Radiology, vol. 235, no. 3, pp. 940–949, Jun. 2005. [35] A. C. Kak and M. Slaney, Principles of Computerized Tomography. Piscataway, NJ: IEEE Press, 1987. [36] A. Bani-Hashemi, E. Blanz, J. Maltz, D. Hristov, and M. Svatos, “Cone beam X-ray scatter removal via image freqency modulation and filtering,” Med. Phys., vol. 32, no. 6, p. 2093, 2005.