Effect of scattered megavoltage x‐rays on markerless tumor tracking using dual energy kilovoltage imaging

Abstract Purpose To determine the effect of megavoltage (MV) scatter on the accuracy of markerless tumor tracking (MTT) for lung tumors using dual energy (DE) imaging and to consider a post‐processing technique to mitigate the effects of MV scatter on DE‐MTT. Methods A Varian TrueBeam linac was used to acquire a series of interleaved 60/120 kVp images of a motion phantom with simulated tumors (10 and 15 mm diameter). Two sets of consecutive high/low energy projections were acquired, with and without MV beam delivery. The MV field sizes (FS) ranged from 2 × 2 cm2–6 × 6 cm2 in steps of 1 × 1 cm2. Weighted logarithmic subtraction was performed on sequential images to produce soft‐tissue images for kV only (DEkV) and kV with MV beam on (DEkV+MV). Wavelet and fast Fourier transformation filtering (wavelet‐FFT) was used to remove stripe noise introduced by MV scatter in the DE images (DEkV+MVCorr). A template‐based matching algorithm was then used to track the target on DEkV, DEkV+MV, and DEkV+MVCorr images. Tracking accuracy was evaluated using the tracking success rate (TSR) and mean absolute error (MAE). Results For the 10 and 15 mm targets, the TSR for DEkV images was 98.7% and 100%, and MAE was 0.53 and 0.42 mm, respectively. For the 10 mm target, the TSR, including the effects of MV scatter, ranged from 86.5% (2 × 2 cm2) to 69.4% (6 × 6 cm2), while the MAE ranged from 2.05 mm to 4.04 mm. The application of wavelet‐FFT algorithm to remove stripe noise (DEkV+MVCorr) resulted in TSR values of 96.9% (2 × 2 cm2) to 93.4% (6 × 6 cm2) and subsequent MAE values were 0.89 mm to 1.37 mm. Similar trends were observed for the 15 mm target. Conclusion MV scatter significantly impacts the tracking accuracy of lung tumors using DE images. Wavelet‐FFT filtering can improve the accuracy of DE‐MTT during treatment.


INTRODUCTION
In radiation therapy of lung cancer, precise localization of the tumor in the presence of breathing motion is vital for treatment accuracy. 1 Markerless lung tumor tracking using kilovoltage (kV) images has been previously reported, negating the potential complications and cost associated with implanting markers. [2][3][4][5][6] One of the barriers to markerless tumor tracking (MTT) is the inferior tumor visibility on single energy (SE) kV images in cases where the tumor is obstructed by bony structures. 6 Our group and others have explored dual energy (DE) kV imaging to address the limitations of MTT using SE images. [7][8][9][10][11][12] DE imaging involves obtaining kV images at high (i.e., 120 kVp) and low (i.e., 60 kVp) energies followed by weighted logarithmic subtraction to suppress bone. These prior studies indicate the potential of DE imaging to enhance the visualization of lung tumors and provide accurate intra-fractional imaging for MTT. [8][9][10] In a clinical scenario, the MV (therapeutic) beam will be delivered while kV images are acquired for MTT. The conventional linear accelerator design has an onboard imager (OBI) that is located at right angles with respect to the MV beam. During the treatment, scattered x-rays from the MV beam are incident on the kV detector. This process introduces "horizontal stripe noise" on kV images caused by the accumulation of scatter signal of the MV treatment beam pulses which are delivered asynchronously to the read out sequence of the kV imager rows. This additional scatter signal not only adds beam noise but also introduces horizontal structures in the kV images when the MV beam is on. 13 Luo et al. demonstrated that MV scatter reduces the contrast-to-noise ratio (CNR) of kV images up to 30%. 13 This contrast degradation, as well as the horizontal line structures introduced by the MV pulses, can cause detection errors or lack of detection of the tumor, making MTT challenging. Hence, there is a clinical need to mitigate the effect of MV stripe noise on kV images to allow for real-time tumor tracking.
Several solutions have been proposed to improve kV image quality during the MV beam treatment delivery. Van Herk et al. proposed an alternating sequence between the kV pulse and MV pulse such that the kV source was enabled during every other frame. This approach was demonstrated for a volumetric modulated radiotherapy (VMAT) delivery and showed that cone beam computed tomography (CBCT) reconstruction was nearly identical to that obtained without the MV beam, with the exception of some unavoidable noise from MV scatter. 14 Ling et al. also examined the effect of VMAT delivery on CBCT image quality. They divided each treatment arc control point into two, the first for the MV radiation delivery and the second for acquiring kV projections. This technique repeatedly pauses the MV beam delivery to acquire scatter-free kV projections. Using this approach, they were able to produce nearly scatter-free CBCT images, however, clinical implementation would require synchronization of kV and MV beams. 15 Another approach places emphasis on calculating the MV scatter map. 16,17 For these studies, three images were typically obtained: (1) kV only images (without MV beam irradiation), (2) concurrent kV images during MV beam irradiation (MV+kV images), and (3) images containing MV-scatter only (MV-scatter map). Each MV-scatter map was separately subtracted from the MV+kV images to obtain MV scatter corrected images. 17 The image contrast was improved by this scatter measurement-based correction method. However, this method was not able to eliminate stripe noise from kV images acquired during MV beam irradiation. 17 An alternative approach using post-processing techniques may be useful in removing stripes. Several techniques have been proposed in other disciplines including moving average filtering, 18 interpolation approaches, 19 frequency filtering with fast Fourier transform (FFT), 20 as well as approaches using wavelets. 21,22 The purpose of this study is twofold. First, we experimentally evaluate the impact of MV scatter and subsequent stripe noise on the accuracy of MTT with DE imaging. Second, we investigate a post-processing technique to remove stripe noise on DE images and determine if this technique is sufficient to reduce the image degradation due to MV scatter. To our knowledge, this is the first study to investigate both concepts related to MTT using DE imaging.

Phantom
The CIRS dynamic thorax motion phantom (CIRS Inc. Norfolk,VA) was used in this study.The phantom approximates an average human thorax in size and with the embedded structures (lung cavities,ribs,and spine).Two different spherical target inserts with diameters of 10 and 15 mm were used in this study. A lung equivalent cylinder containing the spherical target was inserted into the lung equivalent lobe of the phantom. Subsequently, the cos 4 waveform 23 with amplitude = 30 mm and period = 5 s was applied to produce target motion in the superior-inferior direction.

Data acquisition
The on-board imager (OBI) of a TrueBeam linear accelerator (Varian Medical Systems, Palo Alto, CA) was used in Developer mode to acquire fast-kV switching images for DE imaging. The OBI system consists of a For these studies, the phantom was placed on the treatment table with the tumor in its static position located at the isocenter. Two types of images were obtained: (1) kV images without MV beam irradiation for reference (kV only images) and (2) kV images when the MV beam is on (kV+MV images). Both sets of images were acquired over a 180 • arc (gantry angles: 360−180 degrees; imaging angles: 90−270 degrees) using fast kV-switching as discussed above. Note that the difference in the gantry/imaging angles is due to the 90-degree offset of the OBI with the gantry. The mA setting for each pulse was adjusted to minimize the difference in air exposure between the 60 kVp (60 mA, 20 ms) and 120 kVp (15 mA, 20 ms) acquisitions. [8][9][10] For the MV irradiations, a 6X-FFF beam was delivered over the same gantry angles used previously with a dose rate of 1400 MU/min, which is the highest possible dose rate for this energy on TrueBeam, representing the worst case scenario. At this particular dose rate, the MV pulses are delivered with 360 pulses per second, meaning that during the 16 ms required to read out the two halves of the kV imager area, 5−6 MV pulses are delivered. The field size was varied from 2 × 2 cm 2 to 6 × 6 cm 2 in steps of 1 × 1 cm 2 to cover typical field sizes used for lung stereotactic body radiotherapy (SBRT). A summary of the experimental parameters is shown in Table 1.

Dual energy weighted logarithmic subtraction
To create the DE images, weighted logarithmic subtraction (WLS) was performed offline on paired 60/120 kVp images to reduce bone in the resultant soft-tissue images. [8][9][10] WLS was performed to generate DE kV and DE kV+MV as follows 24 : (2) where I H kV , I L kV are the intensities of individual pixels on the high and low energy projections, respectively without MV beam irradiation, while I H kV+MV , I L kV+MV are the high and low energy projections, respectively, when the MV beam is on. The weighting factors, w 1 and w 2 , were used to produce DE kV and DE kV+MV images. Weighting factor optimization was performed using a custom graphic user interface (GUI) in MATLAB R2022a (MathWorks, Natick, MA, USA) to show the resulting DE images interactively as the weighting factor is manually adjusted in steps of 0.01 until the bone was visually removed and the soft tissue image remained. Once the best weighting factor was set for one DE image, it was applied to the next 20 projections, which corresponded to approximately 13 degrees of imaging. A previous study demonstrated that the optimal weighting factor does not change significantly over this range. 9 This process was applied to all images acquired over the 180 • arc to separately create DE kV and DE kV+MV images.

MV stripe noise removal
The core algorithm of combined wavelet and FFT filtering used in the present study was proposed by Münch et al., 22 and termed wavelet-FFT filtering. The algorithm for stripe removal is shown in Figure 1. First, a twodimensional discrete wavelet decomposition (2D-DWT) breaks down the image into three components: diagonal, vertical, and horizontal direction data. The FFT is then used to separate out the horizontal stripes, which are identified by their linear structure in the wavelet. Subsequently, the horizontal coefficients are multiplied by a Gaussian function to eliminate the stripe noise in the frequency domain. The restored image is then generated by applying the 2D-DWT reconstruction. This multi-step process eliminates the unwanted stripes from the image while preserving the other image details. This algorithm was implemented for stripe noise removal using in-house software in MATLAB R2022a (MathWorks, Natick, MA, USA). The wavelet-FFT algorithm has several parameters that require optimization, the first of which is the decomposition level (ℓ) to select the frequency band where the filter is applied. We investigated various parameters to find the best combination that produces a visually satisfying result with a minimal amount of stripe artifacts while still preserving the overall image quality. After assessing various F I G U R E 1 Flowchart of the wavelet-FFT algorithm for stripe removal.
parameters, it was determined that a decomposition value of ℓ = 7 and a db43 wavelet offered the most satisfactory outcome for all DE images. Sigma (σ) defines the width of FFT coefficients which are damped within each horizontal detail band. A larger value of σ will filter more stripes from the image but may also remove more image information. To maintain a balance between stripe removal and the desired image structure preservation, σ = 7 (for 2 × 2 cm 2 and 3 × 3 cm 2 ) and σ = 11 (4 × 4 cm 2 -6 × 6 cm 2 ) pixels was used for this study. Qualitatively, the selected parameters successfully removed horizontal stripes and improved the overall image quality, making it more suitable for tracking across all MV field sizes. It is important to note that this MV stripe removal method was applied to the entire image.
A serendipitous finding of this study demonstrated that vertical stripe removal also improved the quality of DE images (both with and without stripes). The reason is that the individual high/low energy images constituting the DE image are acquired approximately 67 ms apart during which time the gantry moves ∼0.4 • . 9 A previous study has shown that rigid registration between these images can be used, however there are still some subtle misregistration artifacts that can negatively affect tracking. 9 We observed that the vertical stripe removal method by Münch et al. 22 is helpful in removing these artifacts and improves image quality.Therefore,we applied the vertical stripe removal method to both DE kV and DE kV+MV images. The optimized parameters to correct the misregistration artifacts (ℓ = 10, σ = 10, and a db43 wavelet) were subsequently used.  25 In the first step, the RapidTrack-Planning (RTP) software was used to generate templates from a CT scan. Our departmental CT simulator (Somatom Open AS, Siemens Healthineers, Germany) was used to obtain a CT scan of the phantom reconstructed with a 0.6 mm slice thickness. Next, the Eclipse software (version 15.5, Varian) was used to contour the targets and place the treatment isocenter at the center of the static target. CT images and contours were then imported into RapidTrack Planning v1.12 (Varian) to create the individual templates. Twodimensional (2D) reference templates were created for every 1 • of gantry rotation. The parameters used to generate these templates were taken from a recent publication. 26 After creating these templates, for each DE image, the template associated with the imaging angle nearest to the projection was selected. To find the best match between the template and the image, the normalized cross correlation (NCC) of 2D template locations within a specified search window on the image was calculated as a degree of similarity. The NCC was calculated as follows:

Markerless tumor tracking
where n is number of pixels in an image and template, and f (x, y) is the intensity within the search region f of the DE image at location x and y. Similarly, t(x, y) is the intensity value for template t. The parameters f andt denotes the mean value of the DE image and the template,respectively.Last,σ f and σ t are standard deviations over the corresponding regions. The tracked location was determined from the peak value of the NCC match surface.

Tracking metrics
To perform quantitative evaluation of template-based tumor tracking, two metrics were considered: tracking success rate (TSR) and mean absolute error (MAE) with standard deviations (SDs). The TSR is based on the difference between the actual and tracked locations. A successful tracking event for a particular frame is the difference between the tracked location and ground truth (GT) positions of <2 mm. Absolute error refers to the magnitude of difference between the tracked location and the GT. MAE is calculated as the sum of all these absolute errors, divided by the number of frames as given by: where x i and x t are actual tracked location and GT, respectively, and N is the total number of frames for the particular data set. GT positions were obtained based on the cos 4 programmed motion function with fixed amplitude (30 mm) and period (5 s).

RESULTS
The individual images of the high energy (HE) and low energy (LE) kV beams with the MV beam off and on, when a 10 mm target is used, are displayed in Figure 2a-d. By visual inspection, the relative magnitude of the MV stripes is higher in the LE image versus the HE images. Figure 2e,f shows representative intensity profiles through the HE image and the corresponding LE image with (kV+MV) versus without (kV) stripe noise. In both cases, increasing the field size increases the number of counts in the images due to MV scatter accumulation in the row readout of the kV signal. Of note, the effect of MV scatter on the HE images results in a simple vertical shift of the signal. However, on LE images, the MV scatter degrades the image such that much of the signal structure is lost even at field sizes as small as 3 × 3 cm 2 . Although the MV signal is approximately the same in both images, since the kV only signal is significantly lower in the LE image (vs. HE images), the addition of MV scatter overwhelms the kV signal. Therefore, the initial approach to remove stripes on the individual images was not successful for the LE image, since much of the image structure is lost due to MV scatter. As such, we focused on removing stripes on the resultant DE image which provided more consistent results. Figure 3 shows a region of interest (ROI-8 × 8 cm 2 around isocenter) of DE kV and DE kV+MV images with a FS of 2 × 2 cm 2 -6 × 6 cm 2 for a simulated tumor size of 10 mm at a gantry angle of 315.4 • . Qualitatively, the stripe noise is intensified with increasing FS, which significantly degrades the quality of DE kV+MV images. The simulated tumor is clearly visible in Figure 3a. However, with increasing MV field size (Figure 3b-f ),the target visibility is reduced due to the addition of horizontal stripes, particularly in Figure 3e,f . Figure 4 graphically shows the effect of stripe noise on tracking accuracy. The tracked location (green circle) of the target was correctly overlaid on the DE kV image as shown in Figure 4a . The introduction of stripe noise leads the tracking algorithm to misidentify or not be able to locate the target as shown in Figure 4b. Figure 4c shows the restoration of image quality following stripe removal and is similar to that of the reference DE kV . Moreover, the simulated tumor was accurately tracked on the DE corr kV+MV image. The additional benefit of this stripe removal algorithm is that it may also further suppress some bone edges (Figure 4b) that were not completely removed by DE processing indicated by the yellow arrow. Figure 5 shows the TSR for each of the two targets (10 and 15 mm) for DE kV, DE kV+MV , and DE Corr kV+MV images. DE kV images have the highest TSR values of 98.7% for the 10 mm target and 100% for the 15 mm target shown by solid black bars. The diagonal pattern bars show the TSR values for the uncorrected images (DE kV+MV ) as a function of FS. For the 10 mm target, the TSR is reduced to 86% for the 2 × 2 cm 2 FS. As the FS is increased (i.e., 4 × 4 to 6 × 6 cm 2 ), the amount of stripe noise increases (Figure 3) resulting in a more significant reduction in TSR to 69.4%. Similarly for the 15 mm target, DE kV+MV for FS (2 × 2 to 4 × 4 cm 2 ) have TSR values of 92.1% to 89.5%, but for the largest FS it decreases to 77.7% as shown in Figure 5b.
The improved tracking accuracy for the corrected images as compared to MV-scatter images is denoted by the black bars with white dots ( Figure 5). Use of the wavelet-FFT filter on DE kV+MV images was able to increase the TSR to 96.9% for 2 × 2 cm 2 FS and 94.3% for 3 × 3 cm 2 FS for the 10 mm simulated tumor. For the larger FS, it was able to accurately track almost 24% more frames as compared to DE kV+MV images (Figure 5a). A similar observation was made for the 15 mm simulated tumor in Figure 5b. For larger FSs, the stripe removal algorithm requires a higher level of wavelet decomposition that not only removes stripes but may also remove some other features in the image. However, greater improvements were observed for larger FSs, that is, the TSR is >93% for 6 × 6 cm 2 FS for both 10 and 15 mm targets. Of note, in a clinical scenario, 10 mm tumors will typically be treated with FSs between 3 × 3 cm 2 and 4 × 4 cm 2 , while  15 mm tumors will likely be treated with FSs between 4 × 4 cm 2 and 5 × 5 cm 2 . For these field sizes, the wavelet-FFT algorithm worked well to remove the stripes, without significantly impacting tracking accuracy. Subsequently, the TSR for both cases were very close to that of DE kV alone and provided a significant improvement in tracking accuracy as compared to DE kV+MV images.
The MAE of the tracked frames with SDs for DE kV , DE kV+MV, and DE corr kV+MV images are presented in Figure 6. The MAEs of the DE kV+MV images were higher than those of the reference image (DE kV  Other approaches to address the MV scatter issue have been presented in the literature, such as the work of Ling et al., which proposed dividing gantry control points into two parts: one for MV-radiation delivery and the second for acquiring kV images. 14 This method, however, has the limitation of prolonging the treatment delivery time. Van Herk et al. proposed an alternative technique, which used alternating kV imaging and MV treatment beam delivery to avoid the MV scatter; this resulted in a reduction in the overall number of imaging frames. 14, 15 Results similar to ours were obtained, in that both correction methods resulted in the improved contrast-to-noise ratio (CNR) relative to kV-only CBCT images, but the signal was not completely recovered. Additionally, both studies 14,15 utilized flattened-filter (FF) MV beams. However, due to the higher dose rate, FFF beams are more suitable for SBRT. Unfortunately, the above-mentioned techniques are not highly practical for FFF beams due to the high pulse rate. Iramina et al. showed that the MV scatter value factor for FFF beams was larger than that of FF beams. 17 Their method used scatter maps to decrease the effect of MV scatter, but the stripe noise introduced was not completely eliminated when utilizing FFF beams. 17 In comparison to these prior studies, the postprocessing algorithm presented in this study was able to remove most stripe noise and resulted in improved tracking accuracy as compared to uncorrected images. The present method improved the tumor tracking accuracy with TSR > 90% for both the targets across all FSs as shown in Figure 5. For smaller field sizes, the TSR was > 95% for both targets, which is comparable to tracking results obtained in a recent study. 27 An advantage of this technique is that it does not require additional hardware or modifications to the treatment unit. Moreover, this approach was able to remove the stripe noise from the DE images while using FFF beams.
Our eventual goal is to clinically implement MTT for real-time tracking during MV beam irradiation with minimal image degradation. Further studies and improvements are needed for this technique before it can be implemented clinically. Based on our current and prior results, the tracking accuracy is generally higher for larger versus smaller targets. 9,10 Therefore, the tracking accuracy of different types of tumors (size, shape, and density) needs to be explored relative to the interplay with the MV beam. As discussed previously, the implemented stripe removal algorithm was able to restore the image quality close to that of the DE kV images for the smallest FS. However, this method may require some refinement to recover the image quality for larger FS to improve tracking accuracy during treatment delivery. Future studies will also consider hardware techniques combined with the present algorithm to mitigate MV scatter on kV images to improve MTT. Lastly, all the images were processed offline. Clinical implementation will require synchronous processing with limited latency. The computational time to implement the stripe removal algorithm in clinical practice plays an important role. A benefit of the wavelet-FFT algorithm is that it is computationally efficient. 22 Moreover, this algorithm can be parallelized on GPUs and hence can be performed in near-real time during treatment delivery. 28

CONCLUSION
The present study considered the issue of MV scatter and its effects on the tracking accuracy of simulated lung tumors when using DE imaging.Our results demonstrated that stripe noise, introduced by MV scatter, degrades DE image quality, and reduces target tracking accuracy, relative to images obtained without MV scatter. The addition of a wavelet-FFT filter,as a post-processing technique, removes this stripe noise and restores the tracking accuracy to values that are comparable to those obtained without MV scatter.

AU T H O R C O N T R I B U T I O N S
All listed authors contributed to the work and to writing the article.

AC K N OW L E D G M E N T S
Research reported in this publication was supported by the National Cancer Institute of the National Institutes of Health under Award Number R01-CA207483. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

C O N F L I C T O F I N T E R E S T S TAT E M E N T
Mathias Lehmann, Daniel Morf, Liangjia Zhu, and Michal Walczak are employed by Varian Medical Systems.