Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Methods for improving limited field-of-view radiotherapy reconstructions using imperfect a priori images Kenneth J. Ruchalaa) TomoTherapy, Inc., Madison, Wisconsin 53717-1954 Gustavo H. Olivera TomoTherapy, Inc., Madison, Wisconsin 53717-1954 and Department of Medical Physics, University of Wisconsin–Madison, Madison, Wisconsin 53706-1532 Jeffrey M. Kapatoes and Paul J. Reckwerdt TomoTherapy, Inc., Madison, Wisconsin 53717-1954 Thomas R. Mackie TomoTherapy, Inc., Madison, Wisconsin 53717-1954 and Department of Medical Physics and Department of Human Oncology, University of Wisconsin–Madison, Madison, Wisconsin 53706-1532 !Received 10 January 2002; accepted for publication 19 August 2002; published 25 October 2002" There are many benefits to having an online CT imaging system for radiotherapy, as it helps identify changes in the patient’s position and anatomy between the time of planning and treatment. However, many current online CT systems suffer from a limited field-of-view !LFOV" in that collected data do not encompass the patient’s complete cross section. Reconstruction of these data sets can quantitatively distort the image values and introduce artifacts. This work explores the use of planning CT data as a priori information for improving these reconstructions. Methods are presented to incorporate this data by aligning the LFOV with the planning images and then merging the data sets in sinogram space. One alignment option is explicit fusion, producing fusion-aligned reprojection !FAR" images. For cases where explicit fusion is not viable, FAR can be implemented using the implicit fusion of normal setup error, referred to as normal-error-aligned reprojection !NEAR". These methods are evaluated for multiday patient images showing both internal and skin-surface anatomical variation. The iterative use of NEAR and FAR is also investigated, as are applications of NEAR and FAR to dose calculations and the compensation of LFOV online MVCT images with kVCT planning images. Results indicate that NEAR and FAR can utilize planning CT data as imperfect a priori information to reduce artifacts and quantitatively improve images. These benefits can also increase the accuracy of dose calculations and be used for augmenting CT images !e.g., MVCT" acquired at different energies than the planning CT. © 2002 American Association of Physicists in Medicine. #DOI: 10.1118/1.1513163$ Key words: CT, adaptive radiotherapy, radiotherapy verification, image reconstruction, therapy imaging I. INTRODUCTION 1,2 The process of adaptive radiotherapy !ART" is being investigated at many sites for the improvement of radiotherapy. A particularly useful tool for the development of ART is online CT imaging, and together these techniques are commonly referred to as image-guided adaptive radiotherapy. It is particularly useful to have CT capability integrated into the radiotherapy delivery system since this minimizes patient motion between imaging and treatment and potentially enables image acquisition during treatment. There are several approaches to online CT presently under development. One option is to use the megavoltage linear accelerator !linac" as the CT radiation source in conjunction with a detector system to measure the transmitted fluences.3–22 Online kVCT has been explored with an additional kV x-ray tube,23–25 using CT-on-rails,26 and with a dual-energy-capable linac head. 27 Their numerous differences aside, one characteristic com2590 Med. Phys. 29 „11…, November 2002 mon amongst many of these systems is a limited field-ofview !LFOV". This is a lateral truncation of collected data and reconstruction of these incomplete data sets will not typically show the patient’s full extent. The images produced from LFOV data sets can also contain artifacts, commonly known as ‘‘cupping’’ or ‘‘bowl’’ artifacts,28 –32 that may impair the utility of the images for both quantitative and qualitative purposes. The severity of the artifacts depends upon the hardware used and the patient specifics like patient size, anatomy inside and outside the FOV, and position within the scanner. Furthermore, the limitations imposed by these artifacts ultimately depend upon not only the artifact severity, but also on the location and characteristics of the structures of interest, the ability of the observer to ‘‘read-through’’ the distortion, and the desired imaging task. Some of the possible imaging tasks include subjective visual inspection, patient repositioning,26,33 delivery verification,34 dose reconstruction,35,36 treatment adjustment #Kapatoes, 2001 0094-2405Õ2002Õ29„11…Õ2590Õ16Õ$19.00 © 2002 Am. Assoc. Phys. Med. 2590 2591 Ruchala et al.: Methods for improving limited field-of-view #543,37 contouring,9,38 – 40 deformable fusion,40– 43 and deformable dose mapping.40 There are two prevalent reasons for FOV limits in the context of radiotherapy imaging. One limiting problem is the field size of any multileaf collimators !MLCs" attached to the linac. For example, with all leaves open, the NOMOS MIMiC has a field size of approximately 20 cm at isocenter. The MLC of the University of Wisconsin TomoTherapy Prototype has a FOV of approximately 40 cm at the axis, but this is still not sufficiently large to collect complete FOV data for all sites of all patients. The other common cause of LFOV images is the detector size. Some experimental online CT systems use detectors around 20!20 cm2 , providing around 15 cm of image FOV at isocenter. Systems utilizing EPIDs and FPIs around 40!40 cm2 provide FOVs of approximately 30 cm at isocenter. Use of a wider detector can improve the FOV, as can use of a laterally offset detector with overscanning over a larger angular range.6,44,45 It would be impractical to require that an online imaging system be able to completely encompass every possible slice of every patient. This is the motivation behind the development of LFOV data compensation techniques. One set of these techniques is geared toward collection of the missing data. The overscanning methods mentioned previously exemplify this. Analogously, on systems in which the MLC limits the FOV, overscanning could be performed with an offset MLC, or multiple scans could be performed for offset patient positions. Another set of techniques attempts to estimate the missing data based upon a priori knowledge or assumptions about the images or data sets, such as known physical characteristics or redundancies in the measured data. These techniques include: symmetric mirroring,46 use of redundant rays,47 polynomial extrapolation !potentially incorporating known boundaries",48,49 use of known internal geometry,50,51 and iterative techniques in image, sinogram, or Fourier space.52–58 Related techniques also exist for region-of-interest tomography !ROIT" in which the scan FOV is intentionally limited to reduce patient dose, even though complete FOV data sets are potentially available.59– 63 A variation of this theme has been applied to using selectively localized transmission data from radiotherapy treatments for image reconstruction. These data can be reconstructed in conjunction with other imaging data or a slightly modified treatment delivery,19,64,65 or by using other LFOV compensation techniques.9 The same principle can be applied to using treatment data to compensate for a reduced number of imaging angles as a means of reducing imaging dose.66 The goal of this work is to investigate new opportunities for the reconstruction of truncated data sets, whether systematically or intentionally limited. In previous studies, planning CT data have been evaluated for patient alignment67–71 and dose verification.72 Here it is evaluated as a potential source of a priori information for image reconstruction. Yet, if the planning CT images were always a perfect source of a priori information in that the patient’s anatomy never changed, then online CT would be extraneous. Therefore, this work focuses on methods for utilizing the planning CT data for LFOV Medical Physics, Vol. 29, No. 11, November 2002 2591 FIG. 1. Flowchart of the fusion-aligned reprojection !FAR" process. reconstruction with particular regard for a patient’s daily anatomical variations. Image improvement is evaluated for a range of indicative FOV sizes to demonstrate the opportunities for improvement when LFOV compensation is desired. II. MATERIALS AND METHODS A. Fusion-aligned reprojection and normal-error-aligned reprojection field-of-view compensation techniques In order to use the planning image data in conjunction with the online CT data it is helpful to know how the data sets relate. This can often be accomplished by automatic fusion, which will be discussed further in the following. Using this alignment, a method of image improvement called fusion-aligned reprojection !FAR" is investigated. The inputs to the FAR process !Fig. 1" are typically a planning CT image and a LFOV sinogram. This sinogram is initially reconstructed into a LFOV image set and then registered with the complete FOV planning CT. Based upon this result, the planning CT is aligned to the LFOV online image and then reprojected back into a sinogram. The sinogram data are used to estimate the data missing from the LFOV scan, yielding an augmented LFOV sinogram. This augmented sinogram is reconstructed into the FAR image. This completion process can be seen in sinogram space in Fig. 2. A truncated sinogram is shown in Fig. 2!a" for a FOV of 19.9 cm. The missing data are estimated from the aligned planning sinogram shown in Fig. 2!b". The resulting FAR sinogram is shown in Fig. 2!c". Figure 2!d" shows the difference between the complete FOV daily sinogram !not shown" and the aligned planning sinogram. The difference between the complete FOV daily sinogram and the FAR sinogram is shown in Fig. 2!e". There is little difference in the interior regions of this sinogram because the same data are being collected in both cases !aside from possible differences in scatter not included in this study". The peripheral regions of Fig. 2!e" necessarily resemble Fig. 2!d" because the FAR sinogram estimates those regions from the aligned planning sinogram. Figure 2!f" shows another type of FAR sinogram 2592 Ruchala et al.: Methods for improving limited field-of-view FIG. 2. The components of FAR in sinogram space, with lateral position representing detector position, vertical position representing gantry angle, and the sinogram values representing integral attenuation. !a" A 20 cm limited field-of-view !LFOV" online sinogram, !b" the planning CT sinogram, !c" the FAR sinogram obtained by merging !a" and !b". Differences are shown between the complete FOV online sinogram and !d" the planning sinogram, !e" the FAR sinogram, and !f" a FAR sinogram created with an inferior registration. for an inferior registration that will be discussed further in the following. The interior differences are again minimal because that data still come from the online sinogram, but the peripheral data have larger differences due to the inferior planning image alignment. Figure 2!f" underscores the principle that FAR ultimately depends upon an alignment between the planning and LFOV online images. Previous study of this topic has identified several automatic fusion algorithms that are generally successful under these circumstances, especially for FOVs that are approximately half the patient’s size or larger.73 The same study found that as the FOV decreased, the success rate of registrations with larger displacements was more impacted than for small displacements, so even using a very coarse manual fusion could improve automatic fusion results. Since FAR is not specific to the registration technique, other automatic, manual, or hybrid methods could also be used if desired, and FAR images could be created for multiple registrations if the fusion was inconclusive. Previous studies have demonstrated circumstances under which iterative reconstruction techniques could reduce image noise,3,18,74 and these might be helpful in cases where noise hindered the fusion. Any useful LFOV registration for FAR !aside from trialand-error", whether manual or automatic, implies that there is some information in those images in spite of any quantiMedical Physics, Vol. 29, No. 11, November 2002 2592 tative and qualitative degradation. In these cases, the goal of FAR is to quantitatively and qualitatively improve upon the information present by incorporating additional a priori information. Yet, as FOVs become more severely reduced, images may lose their utility for automatic fusion, manual fusion, and visual inspection. There are also a number of other reasons why automatic fusion may not provide the desired result, such as finding a local minimum. Another problem with fusion is that in the presence of anatomical changes there may not be an unambiguous correct alignment, as some structures may align well at the expense of others. It is important to investigate the application of FAR amidst these complications. One possibility is to substitute an implicit fusion for the explicit one. For example, the patient is often positioned within a few millimeters and a few degrees of the intended position using common radiotherapy patient setup protocols. This normal setup error is regarded as an implicit fusion and used for a type of FAR referred to as normal-error-aligned reprojection !NEAR". NEAR may be practical in cases where explicit fusion is unavailable or inaccurate. It is also useful, conceptually, for evaluating the robustness of FAR with regard to imperfect fusion results !e.g., cases that have misregistrations, or lack an unambiguous correct answer due to anatomical changes". A third benefit of NEAR is that it may enable a two #or more$ iteration version of FAR, referred to as NEAR2FAR. After creating a NEAR image, the quantitatively improved voxel values in the FOV might then enable an explicit fusion with the planning image, and a FAR image could be generated. NEAR and NEAR2FAR may be particularly beneficial when a LFOV causes severe quantitative and qualitative degradation of the images, whether because of a large patient, a small detector or MLC, or because a ROIT strategy is being pursued. B. Data sets and image generation As mentioned previously, a motivation of online imaging is that patient anatomy can change from the time of the planning CT to a given fraction. Thus, planning CT data are best evaluated as an imperfect source of a priori information due to anatomical variations and setup errors. It was also desired to have complete FOV images that could be truncated in sinogram space to create LFOV images, so that the results could be compared against a known standard. Appropriate multiday data were provided by William Beaumont Hospital.2 There were 17 daily prostate CT image sets, the first of which was designated the planning CT and the others were designated as daily online image volumes. Contours of the prostate, bladder, and rectum were also provided for each day. These image volumes were acquired at the William Beaumont Hospital on separate CT scanner and did not incorporate precise radiotherapy setup protocols. In order to realign these data sets to resemble normal setup errors for radiotherapy, 16 six-parameter offsets were randomly generated. Each of the three translational offsets were selected from a normal distribution with a zero-mean and standard deviation 2593 Ruchala et al.: Methods for improving limited field-of-view FIG. 3. The intentional misalignments of the planning image relative to the 16 online images, representing randomized normal setup error with standard deviations of 2 mm in each translational direction and 2° around each rotational axis. of 2 mm, and the three rotational errors were selected from a zero-mean normal distribution with a standard deviation of 2°. The net translational and rotational displacements are shown in Fig. 3. The planning CT image was then realigned 16 times, to create a planning CT image offset from each of the online images by one of the randomized six-parameter displacements. The 16 offset planning images were then forward projected, or reprojected, into sinograms with dimensions of 851 gantry angles by 512 detector positions. The 16 online images were also reprojected into sinograms, and then truncated in sinogram space to FOV sizes of 38.6, 29.3, and 19.9 cm. The missing sinogram data from the truncated LFOV images were completed using the corresponding offset planning sinograms. Phasing-in the planning CT sinogram data over the peripheral 15 pixels of each online sinogram’s FOV was heuristically found to reduce artifacts at the junction between the data sets by reducing high-frequency discontinuities at the boundary. This process produced NEAR sinograms for each of the 16 patient-days at the three different FOVs. All of the LFOV and NEAR sinograms were reconstructed using filtered back-projection into CT volumes. The entire process currently requires around 1 min/slice using nonoptimized PV-Wave !Visual Numerics, Houston, TX" codes on a 1.5 GHz Pentium III, primarily due to the reprojection and reconstruction times. Optimized CT reconstruction can currently be performed in subsecond times, and could reduce the time to approximately 1 s/slice. The FAR images were created by using automatic fusion to determine the alignment between each LFOV online image and the corresponding offset planning image. This fusion was done separately for each patient-day and each LFOV size. The particular automatic registration technique used was cropped mutual information !CMI",73 which is a variaMedical Physics, Vol. 29, No. 11, November 2002 2593 tion on the well-known mutual information algorithm.42,75– 80 The primary difference between CMI and mutual information is that peripheral points are discarded from the calculation, to avoid image points that generally have the most severe artifacts. In this case, the voxels within 1.9 cm of the FOV edge were disregarded. However, FAR is not specific to a single fusion algorithm !e.g., CMI" and others could be used as desired. Based upon these explicit fusion results, each offset planning image was realigned. For example, the planning image in randomized offset position 4 was automatically fused to the three sizes of LFOV online images from day 4 and then realigned to three different positions, based upon those results. Each of the 48 realigned planning images !16 patientdays times 3 FOVs" was then reprojected into a sinogram and used to augment the respective LFOV sinogram. The augmented LFOV sinograms were reconstructed into FAR images. The LFOV, NEAR, and FAR image volumes for each patient-day and FOV were then analyzed and compared to the complete FOV online images. Reconstructed images illustrate the qualitative improvements, while quantitative improvements are demonstrated with profile-plots, error histograms, and tabulated rms errors. C. Extensions and applications of NEAR and FAR Using the same basic methodology to create LFOV, NEAR, and FAR images, three additional experiments were performed: testing NEAR and FAR for dose calculations, testing iterative application of NEAR and FAR for severely limited FOVs, and testing FAR for a combination of kilovoltage and megavoltage CT images. 1. Dose calculations using FAR Dose calculations are typically based upon CT images and require reconstructed values that can be calibrated to electron densities. The artifacts and quantitative distortions introduced by FOV truncation may degrade this calibration, while the lack of peripheral information can impair the scatter and attenuation calculations often performed when computing dose. A treatment was optimized for the planning CT image, and the resulting dose distribution was calculated. The same treatment plan was then delivered to an aligned online image, as well as LFOV, NEAR, and FAR images. All dose calculations were performed with convolution/superposition using the CT images converted to electron density. The regions of interest contoured at the William Beaumont Hospital were used to generate DVH curves for the different image sets. 2. Iterative application of NEAR and FAR The use of NEAR and FAR was tested for a FOV of only 10.5 cm. This could reflect either a very small detector or MLC, or represent an intentionally limited field. NEAR images were created for a subset of four patient-days, and the 2594 Ruchala et al.: Methods for improving limited field-of-view FIG. 4. !a" Planning CT image slice with contours for the prostate !solid" and rectum !dotted"; !b" same image slice of an online CT, showing daily !white" and planning !black" contours; !c" another planning CT image slice showing the prostate and bladder !dashed", !d" corresponding slice of an online CT. results were used to test the creation of two-iteration NEAR2FAR images. Images, error histograms, rms values, and DVHs are presented. 3. Application of FAR to cross-energy CT images Since many online imaging systems being investigated use MVCT while planning CT images are almost universally acquired on scanners operating at kilovoltage energies, additional studies were performed to test the feasibility of FAR with cross-energy CT. In principle, FAR should enable the improvement of LFOV MVCT images with kVCT planning images to the extent that both represent the same physical characteristic, namely electron density. However, this assumption is not perfect, especially for kVCT in the presence of high-Z materials, so the goal was to investigate the opportunities for FAR under these circumstances. A CT performance phantom was first imaged to gauge the performance of the TomoTherapy Prototype MVCT. An anesthetized dog was then imaged first on a Siemens Hi-Q kilovoltage CT scanner and then using the TomoTherapy MVCT. The method was otherwise the same as described previously, with the one exception that since different CT systems were involved, the values in each image were converted from Hounsfield values to electron density values based on each scanner’s own calibration curve. The converted electron density images were reprojected and combined in sinogram space. The second cross-energy FAR test was for a human lung patient. Planning kVCT images were obtained on a General Electric PET-CT system, and online images were acquired with the TomoTherapy MVCT. III. RESULTS A. Prostate images using NEAR and FAR Figure 4 shows two representative images from the planning CT and the corresponding slices in two different online Medical Physics, Vol. 29, No. 11, November 2002 2594 FIG. 5. LFOV, normal-error-aligned reprojection !NEAR", and FAR images for field-of-view sizes of 38.6, 29.3, and 19.9 cm based upon the online image in Fig. 4!b". The contours from the complete FOV online image are overlaid for reference. CT image sets. The contours for the planning images are shown in black. The online images also show their own contours in white. The first online image #Fig. 5!b"$ is shown for three LFOVs in Fig. 5. The NEAR and FAR reconstructions from these same sinograms are also shown. The contours from the complete FOV image are overlaid for reference. As the FOV decreases, the artifacts become more severe in the LFOV images, while the NEAR and FAR images are less affected. The extent to which any of these images are sufficient for a given imaging task depends upon that task and possibly on a subjective judgment by a viewer. Central structures generally have less severe artifacts than peripheral structures, and patient size, anatomy, and position all affect artifacts, as does the imaging hardware. Thus, these images are not intended as a cataloging of the precise circumstances under which LFOV artifacts impair a given observer’s ability to perform any particular imaging task. Instead, these images are representative of how NEAR and FAR can utilize available information to qualitatively improve the reconstructions for a range of FOV sizes. In this particular case, there is little visual difference between the NEAR and FAR images. The similarity of NEAR and FAR images can occur for several reasons, such as because the normal setup error is small, the explicit fusion did not improve much upon the normal error, or because anatomical differences between the planning CT and the online CT are a more significant factor than the alignment between those images. Sometimes the differences between NEAR and FAR are more conspicuous, as can be seen in the similar layout of results !Fig. 6" for the online image from Fig. 4!d". In this case, the normal error included a larger rotation about the transverse axis, so the peripheral data in the NEAR image appear rotated relative to the complete FOV online image. 2595 Ruchala et al.: Methods for improving limited field-of-view 2595 FIG. 6. Limited field-of-view, normal-error-aligned reprojection, and fusionaligned reprojection images for field-of-view sizes of 38.6, 29.3, and 19.9 cm based upon the online image in Fig. 4!d". The contours from the complete FOV online image are overlaid for reference. This dissonance between the LFOV data and the offset planning image data results in artifacts, and these are most conspicuous in the 19.9 cm NEAR image. The FAR images improve upon the NEAR images in this case because the explicit fusion better aligned the planning data with the LFOV data. For both NEAR and FAR the junctioning method described earlier tempers these artifacts, but some remain due to anatomical differences between the online and planning images. This is the imperfection of the a priori knowledge and will be addressed more in the discussion. Profiles through the image slice shown in Fig. 5 are plotted in Fig. 7. Each plot contains profiles through the center of the LFOV, NEAR, FAR, and complete FOV images. For the 38.6 cm FOV, all four curves overlap almost indistinguishably aside from a narrow spike in apparent attenuation for the LFOV image at the outer extent of the imaged FOV. There are also differences between the online image and the NEAR and FAR images beyond this FOV extent, representing the imperfection of the planning CT data for estimating the missing peripheral data in the online image. For the FOV of 29.3 cm, the NEAR and FAR curves are again indistinguishable from the complete FOV online image profile inside the FOV. The LFOV image is seen to differ marginally at the center and more pronouncedly toward the limits of the FOV. Outside of the FOV, the NEAR and FAR images again differ from the complete online image, illustrating the differences between the planning and online images. This same pattern is amplified for the 19.9 cm FOV. The cupping artifacts in the LFOV image are increased, while the NEAR and FAR images remain indistinguishable from the complete online image within the scan FOV. However, the differences between the NEAR and FAR images and the complete online image increase outside of the FOV due to the increased reliance on planning CT data in lieu of actual scan data. Medical Physics, Vol. 29, No. 11, November 2002 FIG. 7. Profiles through a limited field-of-view image #black dotted$, complete FOV online image #dark gray solid$, fusion-aligned reprojection image #medium gray chain$, and normal-error-aligned reprojection image #light gray dashes$. Figures 5–7 present results for individual images. Additional quantitative analysis is shown in Fig. 8 in the form of histograms tallied over all 16 patient-days. Image value differences were taken between each LFOV, NEAR, and FAR image and the corresponding complete FOV online image. These difference values were accumulated for each type of reconstruction at each FOV for each ROI. As mentioned earlier, each ROI was defined by contours previously drawn at the William Beaumont Hospital, and the online image’s ROIs were applied to the corresponding LFOV, FAR, and NEAR images. Histograms generated with all ROIs shifted laterally by 2.8 mm were virtually identical to the nonshifted histograms, indicating a low sensitivity to the precise locations of the ROIs and interobserver variability. Figures 8!a" and 8!b" show that at 38.6 cm FOV there is only a marginal difference between the FAR histograms and LFOV histograms, approximately 5 Hounsfield units !HU" or 0.5%. The NEAR curves are indistinguishable from the FAR curves in both cases. There is a slightly greater separation between the curves at 29.3 cm FOV, seen in Figs. 8!c" and 8!d". The NEAR and FAR curves are still not differentiable, but the LFOV curves have shifted right, representing image values in the different ROIs increased by an average of 50– 100 HU !5%–10%" due to the cupping artifacts. For the 19.9 cm FOV #Figs. 8!e" and 8!f"$ this shift is approximately 250 HU, which is consistent with the profiles shown in Fig. 7. The dotted NEAR curve is vaguely differentiable from the FAR curve in Fig. 8!f" in that it is modestly shorter and wider. Additional quantitative analysis of these multiday images is presented in Table I, which contains the mean and standard deviation root-mean-squared !rms" values. The rms differ- 2596 Ruchala et al.: Methods for improving limited field-of-view 2596 FIG. 8. Histograms of the voxel differences between the complete field-of-view online CT !CFOV" and the LFOV and FAR images, tallied over 16 patient-days. The dotted curve for NEAR is only distinguishable from the FAR curves in !f". All horizontal axes have the same scale and units. ence was calculated between each complete FOV online image and the corresponding LFOV, NEAR, and FAR images at each FOV. The rms was scored in each contoured ROI along with the sets of all points in the FOV and all points. The means and standard deviations were taken over the set of 16 daily images. The mean and standard deviation rms values were also computed for the differences between the complete FOV online images and the sets of aligned and misaligned planning images. One important point about Table I is that it provides another evaluation of the quantitative limitations of LFOV reconstructions. For the largest FOV, all three reconstruction methods have very similar rms differences in the three contoured ROIs. The differences are larger for the sets of all points and all FOV points because these sets include the rind where the LFOV artifacts are most severe, whereas the contoured ROIs are further removed from the worst artifacts in this series of images. As the FOV is reduced, the rms is only increased by a few HU in the contoured ROIs for NEAR and FAR, yet the LFOVs rms increase by about 40 HU at 29.3 cm and over 250 HU at 19.9 cm. The rms values for all of the reconstruction methods are worse for the sets of all FOV points than for the contoured ROIs, yet even at the 19.9 cm FOV, the rms for FAR has only increased by about 10 HU, and for NEAR by about 15 HU. In comparison, the rms of the LFOV reconstructions increases from 109 to 383 to 949 as the FOV reduces from 38.6 to 29.3 to 19.9 cm. Table I also demonstrates the quantitative discrepancies between the online images and the planning images !both aligned and normal-error-misaligned". This underscores the benefits on online imaging over relying solely upon the planning CT, while also showing that even a LFOV online image augmented with NEAR or FAR can produce images that are quantitatively closer to the complete FOV online image than the planning image alone. The benefits of NEAR and FAR over the planning image were a reduction in rms error by TABLE I. The mean and standard deviation root-mean-square !rms" values average over 16 patient-days for three different reconstruction methods of three different image field-of-view !FOV" sizes. rms values were scored in rectal, bladder, and prostate regions, and in the sets of all FOV points and all points. The rms values between the aligned and misaligned planning images vs the complete FOV online images are shown for reference. Rectum points Prostate points FOV !cm" Reconstruction 38.6 LFOV NEAR FAR 23.3 22.6 22.5 3.2 3.4 3.4 16.3 16.1 16.1 1.9 2.0 2.0 21.8 21.8 21.8 29.3 LFOV NEAR FAR 76.1 23.2 23.0 4.3 3.4 3.6 54.3 16.8 16.8 2.8 1.9 1.9 19.9 LFOV NEAR FAR 425.7 26.3 24.7 32.6 4.8 4.2 297.5 22.2 21.6 123.3 124.8 116.4 114.2 45.1 48.0 Planning vs aligned daily Planning vs normal error daily a Bladder points Avg rms Std rms Avg rms Std rms Avg rms Std rms For 38.6, 29.3, 19.9 cm FOV, respectively. Medical Physics, Vol. 29, No. 11, November 2002 All points in FOV All points Avg rms Std rms 2.1 2.1 2.1 109.1 24.3 23.1 6.3 2.4 1.8 197.0 85.1 80.9 8.6 13.7 12.5 61.2 22.5 22.4 2.4 2.1 2.2 383.0 27.7 23.5 6.4 6.7 2.1 430.7 139.6 123.1 5.1 11.3 13.4 12.5 5.5 4.5 287.9 24.8 24.4 8.4 3.0 3.0 948.5 39.5 32.5 11.7 8.6 5.4 644.1 172.4 148.9 7.0 14.2 18.2 3.9 4.6 58.1 58.2 2.3 2.8 156.5 200.0 18.2 16.3 139.2, 108.2, 72.0a 23.4, 20.9, 13.6a 199.3, 179.1, 143.6a 20.8, 19.3, 14.1a Avg rms Std rms 2597 Ruchala et al.: Methods for improving limited field-of-view 2597 FIG. 9. Comparison of the DVHs for the dose optimized using the planning CT image !dashes" and the potential treatment dose, that would result from that plan being delivered to the regions of interest amidst anatomical changes !solid lines" based upon the complete FOV online CT. 23–32 HU in the bladder and 34 –36 HU in the prostate. The error was reduced by around 100 HU in the rectum because the NEAR and FAR images better reflected the anatomical changes between the planning and online images. The improvement was only marginal for the set of all points with 29.3 and 19.9 FOVs because the values beyond the scan FOV estimated by NEAR and FAR are essentially only as good as the planning values, so the net improvement is diluted. Finally, Table I quantifies the comparison between NEAR and FAR. For the three contoured regions, the differences are negligible. This indicates that even coarsely aligned peripheral data are generally sufficient to improve the reconstruction of internal regions. For the set of FOV points, the NEAR method results in slightly higher rms values and standard deviations, indicating that NEAR can be a little less reliable for peripheral points within the scan FOV. The biggest difference between NEAR and FAR is for the set of all points, indicating that for the outer points estimated from a version of the planning image, there is some benefit to having an aligned planning image over a normal-error misaligned version. B. Application of NEAR and FAR to dose calculations One important aspect of NEAR and FAR is the opportunity for quantitative improvements and artifact reduction. These benefits may also apply to applications that use these images, such as dose calculations. A dose distribution was optimized for the planning CT volume and the resulting DVHs are shown as the dashed ‘‘Opt’’ curves in Fig. 9. Also shown are the solid ‘‘Tx’’ curves indicating how this same plan would be delivered to one of the treatment CT volumes based upon rigid-body alignment of bony-anatomy. The dose calculations used convolution/superposition, the accuracy of which has been reviewed in many studies.81,82 Changes in the internal anatomy cause discrepancies between the sets of curves. This is both because the anatomical changes affect the dose distribution in physical space and because indiMedical Physics, Vol. 29, No. 11, November 2002 FIG. 10. Bivariate histograms tallying the percentage of error in dose vs percentage of error in image value for 19.9 cm limited field-of-view images and fusion-aligned reprojection images. vidual anatomical structures move relative to this dose distribution. However, DVHs !not shown" created using the two different dose calculations !for the planning and online images" but using the same ROIs !from the planning CT", indicate that the calculated dose distributions are very similar and that ROI changes are the predominant source of discrepancies seen in Fig. 9. The effects of organ motion on dose distribution are not unique to a particular delivery system, but instead underscore the importance of online CT for verification. Dose calculations were run for LFOV image volumes and FAR volumes at the three FOV sizes. By tallying the dosimetric error in each voxel versus the image value error in each voxel, bivariate histograms were created. These are shown for the LFOV and FAR dose distributions for the 19.9 cm images in Fig. 10. There is little correlation between dosimetric error and the image value error, indicating that image value errors analyzed earlier in this paper do not reliably predict the dosimetric error. This is attributable to the systematic changes in attenuation and scatter. Both the dosimetric and image value errors are generally smaller for FAR than for the LFOV image, but there is again no clear correlation between the two axes. The bivariate histogram for NEAR is not shown becomes of its closeness to the FAR histogram. For LFOV, FAR, and NEAR, errors were generally smaller for the larger FOVs, but this did not introduce obvious correlation between the axes. Ideally, the DVHs for the dose calculations to the LFOV and FAR images would exactly match the solid curves in Fig. 9, indicating that the dose delivered !or to-be-delivered" would be precisely known and not degraded by the LFOV. Even though this dose might not match the optimized dose !and in this case it does not due to internal organ motion" the 2598 Ruchala et al.: Methods for improving limited field-of-view 2598 FIG. 12. !a" LFOV reconstruction for a 10.5 cm FOV, !b" the NEAR reconstruction, and !c" the 2-iteration NEAR2FAR reconstruction. The contours from the complete FOV online image are overlaid for reference. strategy being pursued, and is a subject of future investigations in adaptive radiotherapy. FIG. 11. Comparison of the DVHs using the complete FOV online CT !CFOV" vs the three LFOV images at !a" 38.6 cm FOV, !b" 29.3 cm FOV, !c" 19.9 cm FOV, and !d" vs the FAR image based upon the 19.9 cm FOV. goal is to be correctly apprised of any problems so that appropriate modifications can be made based upon reliable information. The results of these dose calculations are shown in Fig. 11, comparing the dose calculated using the complete FOV online CT image with the dose calculation based upon LFOV and FAR images. The DVHs all used the same ROIs, from the complete FOV online image. The differences between the curves do not represent different doses being delivered, as the dose actually received by a patient is not affected by the CT FOV. Instead, the differences represent the error in the dose calculation from using LFOV or FAR images as opposed to the complete FOV online CT. It can be seen that there are discrepancies in the dose distributions calculated using the LFOV online images. These differences are small for the 38.6 cm LFOV image #Fig. 11!a"$, but become more substantial for the 29.3 cm #Fig. 11!b"$ and 19.9 cm #Fig. 11!c"$ LFOV images. In these three cases the dose is systematically overestimated for all structures with increased distortion as the FOV decreases. For this particular case, if only a LFOV online image was available, the resulting dose calculation might lead one to believe that the tumor was receiving an unnecessarily high dose, at the expense of overdosing the bladder and rectum. This could prompt modifications to the patient’s position or the treatment delivery in order to rectify the illusory overdosing. On the other hand, the DVHs based on the dose calculations to the FAR images are virtually indiscernible from the doses to the complete FOV planning image. This held true for all tested FOVs so only the DVHs for the smallest FOV are shown #Fig. 11!d"$. Dose distributions based upon NEAR images were also indistinguishable from the complete FOV calculations, an example of which is shown later. Thus, NEAR and FAR can improve upon the accuracy of dose calculations. The impact on clinical practice depends on the magnitude and nature of the inaccuracy and on the adaptive Medical Physics, Vol. 29, No. 11, November 2002 C. Iterative NEAR2FAR results An iterative application of NEAR and FAR was tested for a FOV of 10.5 cm that encompasses the entire prostate in each daily image, but not necessarily the rectum or bladder. The LFOV reconstruction of this data set is shown in Fig. 12!a", and the NEAR reconstruction is shown in Fig. 12!b". The contours from the complete FOV online image are again overlaid for reference. FAR was not immediately possible because the distortion of image values precluded a successful fusion. A NEAR image was created, and by fusing the interior scan region to the planning CT a NEAR2FAR image could be generated. The NEAR2FAR image is shown in Fig. 12!c". Streak and junctioning artifacts are reduced with the improved alignment. Histograms were tallied for the differences between the LFOV, NEAR, and NEAR2FAR images over the 4 patientdays for which NEAR2FAR images were generated. Figure 13 shows these for the prostate ROI and the set of all FOV points. Again the NEAR and NEAR2FAR histograms are virtually identical but the LFOV histograms are shifted by over 1000 HU !100%". The other ROIs are not included since they were not in the online scan FOV. FIG. 13. Histograms of the voxel differences between the complete field-ofview online CT and the LFOV, NEAR, and NEAR2FAR images tallied over four patient-days. 2599 Ruchala et al.: Methods for improving limited field-of-view 2599 TABLE II. The mean and standard deviation rms values average over four patient-days for three different reconstruction methods with 10.5 cm FOV. rms values were scored in the prostate region, and in the sets of all FOV points and all points. Other ROIs were not contained within this FOV. The rms values between the aligned and misaligned planning images vs the complete FOV online images are shown for reference. Prostate points All points in FOV All points FOV !cm" Reconstruction Avg rms Std rms Avg rms Std rms Avg rms Std rms 10.5 LFOV NEAR NEAR2FAR 1443.8 49.7 50.1 24.7 5.5 5.5 2037.5 96.0 111.0 15.4 11.9 38.8 758.7 182.9 150.5 4.0 16.0 37.6 58.1 58.2 2.3 2.8 76.7 138.4 19.1 16.5 156.5 200.0 18.2 16.3 Planning vs aligned daily Planning vs normal error daily Additional quantitative results can be seen in Table II. These values indicate the severity of the LFOV artifacts for the 10.5 cm FOV, and show the improvements when using NEAR or NEAR2FAR. The rms errors for these images are also lower than for the planning image given the normal setup error. The doses calculated using this LFOV image, NEAR, and NEAR2FAR are shown in Fig. 14. The DVHs use the known contours from the complete FOV online image. The LFOV dose calculation overestimates the prostate dose by approximately 15%, and the rectum and bladder doses have areas of both overestimation and underestimation. The dose distributions calculated using NEAR and NEAR2FAR produce DVHs indistinguishable from the full FOV dose calculation. D. FAR with cross-energy CT The CT performance phantom, imaged with the University of Wisconsin TomoTherapy Prototype is shown in Fig. 15. The dose was 2 cGy, and the complete scan FOV of 40 cm was reconstructed to a 512!512 matrix. This image shows a 20 cm solid water phantom !Gammex-RMI, Middleton, WI" that contains five different inserts and is surrounded by a bone-equivalent annulus. All air-holes sized 1.2 mm and larger are clearly differentiated. All three 29 mm contrast inserts are visible, with the lowest contrast being "3%. Slightly tighter plugs would remove any air that artificially enhances some of the contrast edges. The nominal slice thickness was 0.5 cm and the scan-pitch was 1.6. A solid FIG. 14. Comparisons of a dose calculation for complete FOV online images !CFOV" and !a" the 10.5 cm LFOV image, !b" the 10.5 cm-based NEAR image, and !c" the 10.5 cm-based NEAR2FAR image, for the rectal points, bladder points, and prostate points. Medical Physics, Vol. 29, No. 11, November 2002 2600 Ruchala et al.: Methods for improving limited field-of-view 2600 FIG. 15. A 2 cGy MVCT image on the University of Wisconsin Tomotherapy Prototype shows contrasts of 3% !top" and resolutions of 1.2 mm for air holes !bottom". These regions are expanded on the left with alternate window/level settings. The maximum energy was 3.5 MeV and the mean energy was 1.03 MeV. water plug accommodated an ion chamber !Exradin Model A12s, Standard Imaging, Middleton, WI" used to measure the dose to a slice accumulated during a 20-slice MVCT scan. Ion chamber measurements in the center of the phantom were within several percent of those taken in a peripheral portal of the phantom. Measurements with thermoluminescent dosimeters corroborated these dose values. The performance of this MVCT is further demonstrated by the lung images of a human patient shown in Fig. 16. One slice uses a window and level to accentuate soft-tissue con- FIG. 16. Two lung images from a human patient using 3 cGy TomoTherapy MVCT. The window and level are set to emphasize soft-tissue contrast in one image, and lung brachiation in the other. Medical Physics, Vol. 29, No. 11, November 2002 FIG. 17. !a" A slice from a canine CT on a commercial kilovoltage CT scanner, !b" the corresponding slice using 2 cGy MVCT, !c" a LFOV version of the same MVCT slice, and !d" the FAR reconstruction from the LFOV data, augmented with planning CT data. By reducing the LFOV artifacts, one of the eyes that was occluded in the LFOV reconstruction becomes visible in the FAR reconstruction. trasts, the other uses a wider window to include the brachiation in the lungs. In both images, the patient exceeded the 40 cm FOV. The pitch was 1.0 and the dose was 3 cGy, as confirmed by measurements in phantoms and TLDs on the patient. The canine images are shown in Fig. 17. Figure 17!a" shows a slice from a conventional kilovoltage CT of the dog. The same slice is shown in Fig. 17!b", acquired with a 2 cGy MVCT. A LFOV MVCT data set was created from the complete data set by discarding approximately half of the nonzero ray integrals !one-quarter from each side of the sinogram". The reconstructed LFOV MVCT slice is seen in Fig. 17!c" and the FAR image is shown in Fig. 17!d". Of particular interest is that these canine data sets were not only acquired on different CT systems but at different energies, requiring that FAR combine MV and kV data. The differences between these data sets are most pronounced for higher-Z materials !e.g., bone" because of the increased prevalence of photoelectric interactions at kV energies, which inflates the reconstructed values relative to electron density. This effect is also seen for high-Z endotrachial tube ventilating the dog during anesthesia: it is very bright in the kV image, but barely visible for the same window/level in the MV image. These results show that FAR can be used to augment MV data sets with kV planning CT data, even in the presence of high-Z structures inside, outside, and straddling the online FOV. 2601 Ruchala et al.: Methods for improving limited field-of-view 2601 FIG. 19. !a" A slice from the human lung CT on a commercial kilovoltage CT scanner, !b" the corresponding slice using 3 cGy MVCT, limited by the 40 cm FOV, !c" a 20.9 cm FOV version of the same MVCT slice, and !d" the FAR reconstruction from the 20.9 cm FOV data, augmented with planning CT data. FIG. 18. !a" Histograms of the voxel differences between the planning kVCT and the limited field-of-view #black dashed$, fusion-aligned reprojection #black solid$, and complete FOV MVCT #gray solid$ images. !b" Histograms of the voxel differences between the complete FOV online MVCT and the limited field-of-view #black dashed$ and fusion-aligned reprojection #black solid$ images. Another interesting structure is crystalline of eye that is clearly visible in the kVCT image and is also visible, albeit noisier, in the 2 cGy MVCT. The visibility of this structure is impaired by the limited FOV, but some of this loss is recovered by using FAR. In particular, one eye is again seen just inside the scan FOV through the use of FAR. Visibility of the other eye remains impaired because it lies on the junction of the two data sets. This illustrates the case that even if interior artifacts may not be considered problematic for visualization of central structures, the cupping artifacts may impair conspicuity of more peripheral structures of interest. Figure 18 presents difference histograms for these canine images. The differences in Fig. 18!a" are between the kVCT planning image and the complete FOV !solid gray curve", LFOV !dashed black curve", and FAR MVCT !solid black curve" images. The long negative tails are attributable to increased presence of photoelectric interactions in the kVCT images. However, the goal is not to match the planning CT image, but to better represent the complete FOV online CT in spite of differences relative to the planning CT. Figure 18!b" histograms the differences between the FAR and LFOV images versus the complete FOV MVCT image, showing the improvements with FAR. The rms difference between the complete FOV online MVCT and the planning kVCT is 237.3 HU, but use of this planning CT data in conjunction with the LFOV online MVCT data reduces the rms from 1152.8 HU for the LFOV image to 33.0 for the FAR image. The FAR results for the human lung images are shown in Fig. 19. A planning kVCT image containing the tumor is shown in Fig. 19!a" and the corresponding slice of the Medical Physics, Vol. 29, No. 11, November 2002 MVCT is seen in Fig. 19!b". There are some artifacts in this image as the patient exceeded the 40 cm FOV, and truncating the data to represent a 20.9 cm FOV exacerbates these artifacts #Fig. 19!c"$. FAR was applied to the 20.9 cm FOV image volume and the corresponding slice is shown in Fig. 19!d". The lack of a complete FOV online image for comparison precluded the computation of rms values and error histograms for this FAR image. IV. DISCUSSION As demonstrated previously, NEAR and FAR can be useful for qualitatively reducing cupping artifacts in LFOV images and quantitatively improving the reconstructed values. Like other LFOV compensation techniques, this method is useful in situations where the LFOV reconstruction artifacts qualitatively or quantitatively affect a desired imaging task or application. The quantitative benefits of NEAR and FAR have also been found to apply to the use of LFOV images for dose calculations. There are also several other benefits to NEAR and FAR. These methods can be applied automatically, if desired, or initiated at the request of a user. They do not require the collection of additional online data, which could increase patient dose, nor do they preclude use of the original LFOV images. There is no need to anticipate the extent to which a given LFOV reconstruction will be considered sufficient by a particular user, but instead NEAR and/or FAR can be applied after the scan in situations where improvements are sought. The majority of the computational time is for forward projection and reconstruction, and these steps can potentially be optimized to subsecond times per slice. Another benefit is that NEAR and FAR only require a priori information in the form of the planning CT images. The processing is performed in sinogram space but the current implementation uses reprojected data sets. Access to the original planning CT sinogram data is not necessary and may not be available for commercial CT scanners, and was not 2602 Ruchala et al.: Methods for improving limited field-of-view 2602 TABLE III. Applicability of FAR. Internal anatomy !from online CT, inside FOV" Peripheral anatomy !from planning CT, outside FOV" FAR applicability Same Same FAR unnecessary Patient unchanged so planning CT accurately represents anatomy and is applicable to quantitative applications !e.g., dose calculations" Same Different FAR marginal Online CT data provide little new information relative to planning CT, so FAR offers only limited improvement over planning CT Different Same Ideal FAR case—planning CT misrepresents internal anatomy, but FAR patient representation potentially perfect FAR-based quantitative information potentially perfect Different Different FAR beneficial !most real cases"—planning CT misrepresents internal and external anatomy, but FAR can reduce LFOV artifacts FAR can quantitatively improve image values Dose calculation with FAR can be more accurate than using LFOV CT or planning CT available for this study. Whether NEAR and FAR might benefit from access to this raw sinogram data is a topic for further investigation. Such a study would likely need to account for the differences in scan geometry, preprocessing steps, beam spectra, and calibration that are typically handled within a CT system. One limitation of these methods is that like with other LFOV compensation techniques using a priori information, the results depend upon the validity of those estimations. Since the LFOV online CT typically lacks peripheral data, the reconstruction of outer regions with FAR is estimated based upon the planning CT images and is only meant to provide a context of the patient’s shape. To the extent that it is imperfect a priori information, the outer contour and other soft-tissue structures outside the scan FOV in the FAR image may more closely resemble the planning CT than those structures at time of treatment. This problem is mitigated by positioning the structures of interest within the available FOV, as application of NEAR or FAR can reduce artifacts and quantitatively improve the reconstruction of this scan region, even given imperfect a priori information. It was also found that the peripheral regions, albeit imperfect, are sufficiently representative to enable improved dose calculations. Another limitation is that FAR requires the combination of different data sets, so artifacts can result. These are most commonly rings or arcs in the vicinity of where the online CT FOV ends !Fig. 6", and are commonly minor as compared to the LFOV artifacts. Different types of smoothing and junctioning were attempted to mitigate this effect, and the preferred method was a linear phase-in of the planning CT sinogram data over the peripheral 15 points of the online sinogram. Alternate filters or other improvements in FAR may further reduce artifacts, but their complete elimination for all cases is unlikely. As mentioned earlier, artifacts arise Medical Physics, Vol. 29, No. 11, November 2002 from inconsistencies in the data sinograms. To the extent that the patient’s outer anatomy has changed between the planning CT and the online CT or the data sets are misaligned there will likely be inconsistencies in the NEAR or FAR sinogram. This raises the issue of when FAR will be most and least useful. The applicability is summarized in Table III, in the context of whether inner and/or peripheral anatomy has changed from the time of the planning CT to that of the treatment-time CT. Obviously, anatomical changes are much more complex than the four categorizations used, but the aim is to illustrate the utility of FAR at the logical extrema. NEAR will be useful under similar conditions to FAR, but the normal error can impact results. The first row makes that point that if there are changes in neither the internal anatomy nor peripheral anatomy, then FAR is unnecessary, because the patient’s anatomy is identical to the planning CT. Similarly, if there are changes only to the peripheral anatomy beyond the FOV of the online CT system #Table III, row 2$, then FAR does not provide many benefits over using the planning CT since the collected data only contain marginal information about the peripheral regions. FAR would still be successful at improving the LFOV images but offers few benefits over using the planning CT itself, since the latter is sufficiently representative. In practice, a patient’s anatomy can change, which is precisely why online CT is beneficial, and these are the cases where FAR is most applicable. If only a patient’s internal anatomy changed, but the peripheral anatomy remained rigid #Table III, row 3$, then FAR could work perfectly. It could provide exact quantitative and qualitative information, benefiting both the patient representation and applications like dose calculations. This is because the data collected by the online CT could completely replace all of the changed por- 2603 Ruchala et al.: Methods for improving limited field-of-view tions of the planning CT sinogram. Yet, like the two cases mentioned previously, this is more of an illustrative extreme than a common clinical occurrence. Instead, there will commonly be some combination of internal and peripheral anatomy changes #Table III, row 4$. This is the situation for the cases evaluated here. The presence of internal anatomical changes motivates using the online CT data instead of the original planning CT. These anatomical changes also introduce inconsistencies between the online CT and planning CT sinograms, so the FAR image reconstruction will typically contain some artifacts, the magnitude of which can vary, as seen in Figs. 5, 6, 17, and 19. Yet, the images can still provide a substantial quantitative improvement and artifact reduction relative to the LFOV images. Likewise, since the images are not a perfect representation of the patient, the dose calculation will not be perfect; but in the cases studied !Figs. 11 and 14" this discrepancy was imperceptible. In terms of Table III, the biggest conceptual difference between NEAR and FAR is for row 3 !changed internal anatomy, identical peripheral anatomy", as the normal setup error would detract from this theoretically perfect case. For the multiday results presented, the differences between NEAR and FAR were less distinct, because anatomical changes were not localized to internal regions and the notion of a single ‘‘gold standard’’ image registration is compromised by these anatomical changes. There were some cases where FAR resulted in fewer artifacts than NEAR and FAR often had slightly better rms values. Nonetheless, NEAR was found to be a viable variation of FAR for situations where explicit fusion is not possible. These NEAR results likewise demonstrate the results that would be obtained if FAR was attempted with an inaccurate fusion result. NEAR could also be selectively followed by FAR iterations, if desired. Since neither NEAR nor FAR requires additional data collection, the choice between them does not need to be predetermined. Both can be tried, they can be combined with an iterative method, or FAR can even be tried multiple times with different registration results. A. Going FARther Finally, there are additional ways in which NEAR and FAR could be applied or extended. There are many other qualitative and quantitative methods for adaptive radiotherapy and these may benefit from LFOV compensation techniques such as the ones presented. NEAR and FAR could also be used for purposes beyond radiotherapy in cases where #potentially imperfect$ a priori information is available. In addition, while this study focused on a priori information in the form of a kilovoltage planning CT, it may be feasible to apply NEAR and FAR to multimodality images, such as creating a FAR image by combining an online CT !MV or kV" data set with a planning MRI image. In such cases, the MRI or other-modality image would need to be converted to values compatible with the LFOV data set. A complex mapping of values might provide the best results, but even using the alternate modality image to describe the Medical Physics, Vol. 29, No. 11, November 2002 2603 patient’s outer contour and using a water-equivalency assumption may provide benefits. This is particularly true considering the demonstrated robustness of FAR with regard to anatomical changes, imperfect alignments, and even the systematic differences in reconstructed values between MV- and kVCT images. This could present a practical method for developing upon long-standing LFOV completion principles.48,49,63 NEAR and FAR may also be extendable to other types of limited-data situations, such as limited-slice, limited-projection, or more complex patterns of limited data, such as reconstructing images collected during modulated treatment deliveries.19 V. CONCLUSION The use of planning CT data has been investigated as a means to improve limited field-of-view !LFOV" online CT images. The methods of fusion-aligned reprojection !FAR" and normal-error-aligned reprojection !NEAR" have been found to be successful at reducing artifacts and quantitatively improving image values with regards to the complete FOV images. These benefits can be helpful in cases where LFOV artifacts impair the desired imaging task. NEAR and FAR have also been found to improve dose calculations for LFOV images, as DVHs for the NEAR and FAR images are virtually identical to those based on complete FOV images. Tests of these methods with canine and and human data demonstrate their benefits to cross-energy cases, such as using a MVCT online imaging system with kVCT planning images. ACKNOWLEDGMENTS The authors are very thankful to Dr. Di Yan for providing the multiday patient CT images, structure sets, and accompanying manual fusion data collected by himself and others at the William Beaumont Hospital. We also wish to extend our appreciation to Dr. David Pearson, Eric Schloesser, Dr. Lisa Forrest, and Stewart Becker for their efforts in the collection of veterinary data at the UW hospital and to Dr. James Welsh, Dr. Minesh Mehta, Dr. John Balog, and Dr. Susanta Hui for their participation in the human patient clinical trials. a" Electronic mail: [email protected] D. Yan, F. Vicini, J. Wong, and A. Martinez, ‘‘Adaptive radiation therapy,’’ Phys. Med. Biol. 42, 123–132 !1997". 2 D. Yan, D. Jaffray, and J. Wong, ‘‘A model to accumulate fractionated dose in a deforming organ,’’ Int. J. Radiat. Oncol., Biol., Phys. 44, 665– 675 !1999". 3 H. Guan and Y. Zhu, ‘‘Feasibility of megavoltage portal CT using an electronic portal imaging device !EPID" and a multi-level scheme algebraic reconstruction technique !MLS-ART",’’ Phys. Med. Biol. 43, 2925– 2937 !1998". 4 D. A. Jaffray, D. G. Drake, M. Moreau, A. A. Martinez, and J. W. Wong, ‘‘A radiographic and tomographic imaging system integrated into a medical linear accelerator for localization of bone and soft-tissue targets,’’ Int. J. Radiat. Oncol., Biol., Phys. 43, 773–789 !1999". 5 B. M. Hesse, L. Spies, and B. A. Groh, ‘‘Tomotherapeutic portal imaging for radiation treatment verification,’’ Phys. Med. Biol. 43, 3607–3616 !1998". 6 S. Midgley, R. M. Millar, and J. Dudson, ‘‘A feasibility study for megavoltage cone beam CT using a commercial EPID,’’ Phys. Med. Biol. 43, 155–169 !1998". 1 2604 Ruchala et al.: Methods for improving limited field-of-view 7 M. A. Mosleh-Shirazi, P. M. Evans, W. Swindell, S. Webb, and M. Partridge, ‘‘A cone-beam megavoltage CT scanner for treatment verification in conformal radiotherapy,’’ Radiother. Oncol. 48, 319–328 !1998". 8 S. Loose and K. W. Leszczynski, ‘‘On few-view tomographic reconstruction with megavoltage photon beams,’’ Med. Phys. 28, 1679–1688 !2001". 9 E. Ford, J. Chang, H. Amols, D. Todor, G. Mageras, K. Sidhu, E. Yorke, and C. Ling, ‘‘Cone-beam CT for radiation treatment verification with megavoltage beams and an amorphous silicon imager,’’ Med. Phys. 28, 1232 !2001", abstract. 10 M. Partridge, P. M. Evans, and M. A. Mosleh-Shirazi, ‘‘Linear accelerator output variations and their consequences for megavoltage imaging,’’ Med. Phys. 25, 1443–1452 !1998". 11 B. Groh, ‘‘A study of the use of flat-panel imagers for radiotherapy verification,’’ Dipl.-Phys. Ph.D. thesis Medical Physics, Deutsches Krebsforschungszentrum !DKFZ"-Heidelberg, Germany, 2002. 12 D. G. Lewis, W. Swindell, E. J. Morton, P. M. Evans, and Z. R. Xiao, ‘‘A megavoltage CT scanner for radiotherapy verification,’’ Phys. Med. Biol. 37, 1985–1999 !1992". 13 L. Spies, ‘‘On scatter in megavoltage x-ray transmission imaging,’’ Dipl.Phys. thesis, Department of Medical Physics, Deutsches Krebsforschungszentrum-Heidelberg, 2000. 14 L. Spies, M. Ebert, B. Groh, B. Hesse, and T. Bortfeld, ‘‘Correction of scatter in megavoltage cone-beam CT,’’ Phys. Med. Biol. 46, 821– 833 !2001". 15 K. Nakagawa, Y. Aoki, A. Akanuma, Y. Onogi, A. Terahara, K. Sakata, N. Muta, Y. Sasaki, H. Kawakami, and K. Hanakawa, ‘‘Real-time beam monitoring in dynamic conformation therapy,’’ Int. J. Radiat. Oncol., Biol., Phys. 30, 1233–1238 !1994". 16 K. Nakagawa, Y. Aoki, M. Tago, A. Terahara, and K. Ohtomo, ‘‘Megavoltage CT-assisted sterotactic radiosurgery for thoracic tumors: Original research in the treatment of thoracic neoplasms,’’ Int. J. Radiat. Oncol., Biol., Phys. 48, 449– 457 !2000". 17 K. Mueller, J. Chang, H. Amols, and C. C. Ling, ‘‘Cone-beam computed tomography !CT" for a megavoltage linear accelerator !LINAC" using an electronic portal imaging device !EPID" and the algebraic reconstruction technique !ART",’’ Med. Phys. 27, 1439 !2000", abstract. 18 K. J. Ruchala, G. H. Olivera, E. A. Schloesser, and T. R. Mackie, ‘‘Megavoltage CT on a tomotherapy system,’’ Phys. Med. Biol. 44, 2597–2621 !1999". 19 K. J. Ruchala, G. H. Olivera, J. M. Kapatoes, E. A. Schloesser, P. J. Reckwerdt, and T. R. Mackie, ‘‘Megavoltage CT reconstruction during tomotherapy treatments,’’ Phys. Med. Biol. 45, 3545–3562 !2000". 20 K. J. Ruchala, G. H. Olivera, J. M. Kapatoes, E. A. Schloesser, P. J. Reckwerdt, and T. R. Mackie, ‘‘Megavoltage CT imaging as a by-product of multi-leaf collimator leakage,’’ Phys. Med. Biol. 45, N61–N70 !2000". 21 K. J. Ruchala, G. H. Olivera, E. A. Schloesser, R. Hinderer, and T. R. Mackie, ‘‘Calibration of a tomotherapeutic megavoltage CT system,’’ Phys. Med. Biol. 45, N27–N36 !2000". 22 R. Hinderer, J. M. Kapatoes, R. Jeraj, R. Schmidt, H. Keller, K. J. Ruchala, D. Pearson, G. Fang, G. H. Olivera, and T. R. Mackie, ‘‘A novel approach for a highly efficient detector for megavoltage photons,’’ Med. Phys. 28, 1232 !2001", abstract. 23 D. A. Jaffray and J. H. Siewerdsen, ‘‘Cone-beam computed tomography with a flat-panel imager: Initial performance characterization,’’ Med. Phys. 27, 1311–1323 !2000". 24 D. A. Jaffray, ‘‘Cone-beam computed tomography on a medical linear accelerator using a flat-panel imager,’’ Proceedings of the ICCR, The Use of Computers in Radiation Therapy, Heidelberg, Germany, edited by W. Schlegel and T. Bortfeld !Springer, Berlin, 2000", pp. 558 –560. 25 T. R. Mackie, T. Holmes, S. Swerdloff, P. Reckwerdt, J. O. Deasy, J. Yang, B. Paliwal, and T. Kinsella, ‘‘Tomotherapy: A new concept for the delivery of dynamic conformal radiotherapy,’’ Med. Phys. 20, 1709–1719 !1993". 26 C. Hua, M. Lovelock, G. Mageras, E. Lief, M. Katz, J. Mechalakos, T. Hollister, W. Lutz, M. Zelefsky, and C. Ling, ‘‘Online prostate position correction using 5 CT slices for prostate cancer patients undergoing image-guided external beam radiotherapy,’’ Med. Phys. 28, 1258 !2001", abstract. 27 P. Munro, Y. Cho, and M. Mulligan, ‘‘Kilovision: Kilovoltage imaging using a medical linear accelerator,’’ Med. Phys. 28, 1232 !2001", abstract. 28 R. N. Bracewell and S. J. Wernecke, ‘‘Image reconstruction over a finite field of view,’’ J. Opt. Soc. Am. 65, 1342–1346 !1975". Medical Physics, Vol. 29, No. 11, November 2002 2604 29 J. C. Gore and S. Leeman, ‘‘The reconstruction of objects from incomplete projections,’’ Phys. Med. Biol. 25, 129–136 !1980". 30 S. C. Huang, M. E. Phelps, and E. J. Hoffman, ‘‘Effect of out-of-field objects in transaxial reconstruction tomography,’’ in Reconstruction Tomography in Diagnostic Radiology and Nuclear Medicine, edited by M. M. Ter-Pogossian, M. E. Phelps, G. L. Brownell, J. R. Cox, D. O. Davis, and R. G. Evens !University Park Press, Baltimore, 1977". 31 B. E. Oppenheim, ‘‘Reconstruction tomography from incomplete projections,’’ in Ref. 30. 32 P. S. Tofts and J. C. Gore, ‘‘Some sources of artifact in computed tomography,’’ Phys. Med. Biol. 25, 117–127 !1980". 33 G. H. Olivera, K. J. Ruchala, J. M. Kapatoes, R. Jeraj, P. J. Reckwerdt, J. P. Balog, H. Keller, and T. R. Mackie, ‘‘Advantages and limitations in patient setup for prostate treatments using image based automatic fusion techniques,’’ Med. Phys. 28, 1260 !2001", abstract. 34 J. M. Kapatoes, G. H. Olivera, P. J. Reckwerdt, E. E. Fitchard, and T. R. Mackie, ‘‘Delivery verification in sequential and helical tomotherapy,’’ Phys. Med. Biol. 44, 1815–1841 !1999". 35 J. M. Kapatoes, G. H. Olivera, K. J. Ruchala, J. S. Smilowitz, P. J. Reckwerdt, and T. R. Mackie, ‘‘A feasible method for clinical delivery verification and dose reconstruction in tomotherapy,’’ Med. Phys. 28, 528 –542 !2001". 36 J. M. Kapatoes, G. H. Olivera, K. J. Ruchala, H. Keller, P. J. Reckwerdt, and T. R. Mackie, ‘‘Knowledge of the 3D treatment-position anatomy: A prerequisite for accurate dose reconstruction’’ Int. J. Radiat. Oncol., Biol., Phys. !submitted". 37 G. H. Olivera, E. E. Fitchard, P. J. Reckwerdt, K. J. Ruchala, and T. R. Mackie, ‘‘Delivery modification as an alternative to patient repositioning in tomotherapy,’’ Proceedings of the XIII International Conference on Computers in Radiotherapy, Heidelberg, Germany, edited by W. Schlegel and T. Bortfeld !Springer, Berlin, 2000", pp. 297–299. 38 M. Kass, A. Witkin, and D. Terzopoulous, ‘‘Snakes: Active contour models,’’ Int. J. Comput. Vis. 4, 321–331 !1988". 39 T. Coote, A. Hill, C. Taylor, and J. Haslam, ‘‘The use of active shape models for locating structures in medical images,’’ J. Image Vision Comput. 12, 355–366 !1994". 40 W. Lu, ‘‘Motion detection and correction for image guided radiation therapy,’’ Ph.D. thesis, Department of Medical Physics, University of Wisconsin–Madison, 2001. 41 H. Lester and S. R. Arridge, ‘‘A survey of hierarchical non-linear medical image registration,’’ Pattern Recogn. 32, 129–149 !1999". 42 D. Hill, P. Batchelor, M. Holden, and D. Hawkes, ‘‘Medical image registration,’’ Phys. Med. Biol. 46, R1–R45 !2001". 43 M. L. Kessler, K. Li, and C. Meyer, ‘‘Automated image registration using mutual information for both affine and thin-plate spline geometric transformations,’’ in Ref. 24, pp. 96 –98. 44 P. S. Cho, R. H. Johnson, and T. W. Griffin, ‘‘Cone-beam CT for radiotherapy applications,’’ Phys. Med. Biol. 40, 1863–1883 !1995". 45 D. Jaffray and J. Siewerdsen, ‘‘Cone-beam CT with a flat-panel imager: Dosimetric considerations,’’ Med. Phys. 28, 1259 !2001". 46 B. Ohnesorge, T. Flohr, K. Schwarz, J. P. Heiken, and K. T. Bae, ‘‘Efficient correction for CT image artifacts caused by objects extending outside the scan field of view,’’ Med. Phys. 27, 39– 46 !2000". 47 P. M. Margosian, ‘‘A redundant ray projection completion method for an inverse fan beam computed tomography system,’’ J. Comp. Assist. Tomogr 6, 608 – 613 !1982". 48 R. M. Lewitt, ‘‘Processing of incomplete measurement data in computed tomography,’’ Med. Phys. 6, 412– 417 !1979". 49 G. T. Herman and R. M. Lewitt, ‘‘Evaluation of a preprocessing algorithm for truncated CT projections,’’ J. Comput. Assist. Tomogr. 5, 127– 135 !1981". 50 L.-T. Chang, ‘‘Attenuation correction and incomplete projection single photon emission computed tomography,’’ IEEE Trans. Nucl. Sci. 26, 2780–2785 !1979". 51 T. E. Gureyev and R. Evans, ‘‘Tomography of objects with a priori known internal geometry,’’ Inverse Prob. 14, 1469–1480 !1998". 52 P. R. Edholm, R. M. Lewitt, and B. Lindholm, ‘‘Novel properties of the Fourier decomposition of the sinogram,’’ Proceedings of the International Workshop on Physics and Engineering of Computerized Multidimensional Imaging and Processing, Newport Beach, CA, edited by O. Nalcioglu !SPIE, Bellingham, 1986", pp. 8 –18. 53 J. S. Karp, G. Muehllehner, and R. M. Lewitt, ‘‘Constrained Fourier 2605 Ruchala et al.: Methods for improving limited field-of-view space method for compensation of missing data in emission computed tomography,’’ IEEE Trans. Med. Imaging 7, 21–25 !1988". 54 J. M. Ollinger and J. S. Karp, ‘‘An evaluation of three algorithms for reconstructing images from data with missing projections,’’ IEEE Trans. Nucl. Sci. 35, 629– 634 !1988". 55 K. C. Tam and V. Perez-Mendez, ‘‘Limited-angle three-dimensional reconstructions using Fourier transform iterations and radon transform iterations,’’ Opt. Eng. 20, 586 –589 !1981". 56 K. C. Tam, J. W. Eberhard, and K. W. Mitchell, ‘‘Incomplete-data CT image reconstructions in industrial applications,’’ IEEE Trans. Nucl. Sci. 37, 1490–1499 !1990". 57 K. C. Tam and V. Perez-Mendez, ‘‘Tomographical imaging with limitedangle input,’’ J. Opt. Soc. Am. 71, 582–592 !1981". 58 T. Sato, S. J. Norton, M. Linzer, O. Ikeda, and M. Hirama, ‘‘Tomographic image reconstruction from limited projections using iterative revisions in image and transform spaces,’’ Appl. Opt. 20, 395–399 !1981". 59 O. Nalcioglu, Z. H. Cho, and R. Y. Lou, ‘‘Limited field of view reconstruction in computerized tomography,’’ IEEE Trans. Nucl. Sci. NS-26, 546 –551 !1979". 60 O. Nalcioglu and R. Lou, ‘‘Post-reconstruction method for beam hardening in computerized tomography,’’ Phys. Med. Biol. 24, 330–340 !1979". 61 O. Nalcioglu, P. V. Sankar, and J. Sklansky, ‘‘Region-of-interest x-ray tomography !ROIT",’’ SPIE 206, 98 –102 !1979". 62 P. V. Sankar, O. Nalcioglu, and J. Sklansky, ‘‘Undersampling errors in region-of-interest tomography,’’ IEEE Trans. Med. Imaging 1, 168 –173 !1982". 63 W. Wagner, ‘‘Reconstructions from restricted region scan data—New means to reduce the patient dose,’’ IEEE Trans. Nucl. Sci. NS-26, 2866 – 2869 !1979". 64 K. J. Ruchala, ‘‘Megavoltage computed tomography for tomotherapeutic verification,’’ Ph.D. thesis, Department of Medical Physics, University of Wisconsin–Madison, 1999. 65 K. J. Ruchala, G. H. Olivera, J. M. Kapatoes, E. A. Schloesser, and T. R. Mackie, ‘‘Megavoltage CT imaging using the incomplete data collected during modulated treatment delivery,’’ Med. Phys. 26, 1064 !1999", abstract. 66 H. Guan, F. Yin, Y. Zhu, and J. Kim, ‘‘Adaptive portal CT reconstruction: A simulation study,’’ Med. Phys. 27, 2209–2214 !2000". 67 J. Kim, J. A. Fessler, K. L. Lam, J. M. Balter, and R. K. Ten Haken, ‘‘A feasibility study of mutual information based setup error estimation for radiotherapy,’’ Med. Phys. 28, 2507–2517 !2001". 68 L. J. Pisani, D. Yan, and J. W. Wong, ‘‘A CT-portal alignment method for 3-D patient localization in radiation treatment setup,’’ Med. Phys. 23, 1132 !1996", abstract. 69 M. L. Kessler and K. Li, ‘‘Image fusion for conformal radiation therapy,’’ in 3-D Conformal and Intensity Modulated Radiation Therapy, edited by J. A. Purdy, W. H. I. Grant, J. R. Palta, B. Butler, and C. Perez !Advanced Medical Publishing, Madison, WI, 2001". Medical Physics, Vol. 29, No. 11, November 2002 2605 70 M. van Herk, A. Bruce, and A. P. Guus Kroes, ‘‘Quantification of organ motion during conformal radiotherapy of the prostate by three dimensional image registration,’’ Int. J. Radiat. Oncol., Biol., Phys. 33, 1311– 1320 !1995". 71 K. G. Gilhuijs, K. Drukker, A. Touw, P. J. van de Ven, and M. van Herk, ‘‘Interactive three dimensional inspection of patient setup in radiation therapy using digital portal images and computed tomography data,’’ Int. J. Radiat. Oncol., Biol., Phys. 34, 873– 885 !1996". 72 T. R. McNutt, T. R. Mackie, P. J. Reckwerdt, and B. R. Paliwal, ‘‘Modeling dose distributions from portal dose images using the convolution/ superposition method,’’ Med. Phys. 23, 1381–1392 !1996". 73 K. J. Ruchala, G. H. Olivera, and J. M. Kapatoes, ‘‘Limited-data image registration for radiotherapy positioning and verification,’’ Int. J. Radiat. Oncol., Biol., Phys. 54, 592– 605 !2002". 74 H. Guan and R. Gordon, ‘‘Computed tomography using algebraic reconstruction techniques !ARTs" with different projection access schemes: A comparison study under practical situations,’’ Phys. Med. Biol. 41, 1727– 1743 !1996". 75 P. Viola and W. M. Wells III, ‘‘Alignment by maximization of mutual information,’’ Int. J. Comput. Vis. 24, 137–154 !1997". 76 W. M. Wells III, P. Viola, H. Atsumi, S. Nakajima, and R. Kikinis, ‘‘Multi-modal volume registration by maximization of mutual information,’’ Med. Image Anal 1, 35–51 !1996". 77 F. Maes, A. Collignon, D. Vandermeulen, G. Marchal, and P. Suetens, ‘‘Multimodality image registration by maximization of mutual information,’’ IEEE Trans. Med. Imaging 16, 187–198 !1997". 78 D. L. G. Hill and P. Batchelor, ‘‘Registration methodology: Concepts and algorithms,’’ in Medical Image Registration, edited by J. V. Hajnal, D. L. G. Hill, and D. J. Hawkes !CRC, Boca Raton, FL, 2001". 79 J. Pluim, A. Maintz, and M. Viergever, ‘‘Interpolation artifacts in mutual information-based image registration,’’ Comput. Vis. Image Underst. 77, 211–232 !2000". 80 C. Studholme, D. Hill, and D. Hawkes, ‘‘Automated three-dimensional registration of magnetic resonance and positron emission tomography brain images by multiresolution optimization of voxel similarity measures,’’ Med. Phys. 24, 25–35 !1997". 81 T. R. Mackie, P. Reckwerdt, T. McNutt, M. Gehring, and C. Sanders, 1996, ‘‘Photon beam dose computations,’’ Proceedings of the 1996 Summer School for the American Association of Physicists in Medicine, Vancouver, B.C., Canada, edited by T. R. Mackie and J. R. Palta !Advanced Medical Publishing, Madison, 1996", pp. 103–174. 82 M. R. Arnfield, C. H. Siantar, J. Siebers, P. Garmon, L. Cox, and R. Mohan, ‘‘The impact of electron transport on the accuracy of computed dose,’’ Med. Phys. 27, 1266 –1274 !2000".