Technical Note: Synthetic treatment beam imaging for motion monitoring during spine SBRT treatments – a phantom study

Abstract

Purpose:

One of the biggest challenges in applying megavoltage (MV) treatment beam imaging for monitoring spine motion in stereotactic body radiotherapy (SBRT) is the small beam apertures in the images due to strong beam modulations in IMRT planning. The purpose of this study is to investigate the feasibility of a markerless motion tracking method in spine SBRT delivery using a novel enhanced synthetic treatment beam (ESTB) imaging technique.

Methods:

Three clinical spine SBRT plans using 6XFFF beams and sliding window IMRT technique were transferred to a thorax phantom and delivered by a TrueBeam machine. Before delivery, the phantom was aligned to the plan isocenter using CBCT setup and verified with a second CBCT, and then 2 mm shifts were introduced in both the cranio-caudal (CC) and the left–right (LR) directions with the couch. During beam delivery, MV images were continuously taken with an electronic portal imaging device (EPID) and automatically grabbed by Varian iTools Capture software with a frame rate of 11.6 Hz. After preprocessing for scatter correction and beam intensity compensation, every 50 frames of MV images were combined to generate a series of ESTB images for each beam. The ESTB images were then registered to the projections of the verification CBCT at the matched beam angles to detect the 2 mm shifts.

Results:

Compared to snapshot MV images, the ESTB images had significantly enlarged fields of view (FOVs) and improved image quality. Based on 2D rigid registration, the ESTB image to CBCT projection matching showed submillimeter accuracy in detecting motion. Specifically, the root mean square errors in detecting the LR/CC shifts were 0.35/0.28 mm, 0.32/0.35 mm, 0.63/0.44 mm, 0.55/0.51 mm, and 0.69/0.42 mm at gantry angles 180, 160, 140, 120, 100, respectively.

Conclusion:

Our results in the phantom study suggest that ESTB images from a sliding window IMRT plan can be used to detect spine motion, with submillimeter precision in the 2D plane perpendicular to the beam.

1. INTRODUCTION

Stereotactic body radiotherapy (SBRT) has been proven highly efficacious in treating spinal metastases [1-3], however, due to the high radiation dose per fraction (8~24Gy per fraction) and sharp dose gradient (5-10% per mm), spinal cord motion during spine SBRT treatment is a big concern, even with the implementation of an immobilization device [4]. Constant monitoring of the spine position during beam delivery is critical to avoiding radiation induced neurologic toxicity to the spinal cord or cauda. There are several systems clinically available for quantitative monitoring of spine motion during radiation treatment, including: 1) Xsight spine tracking on the CyberKnife platform (Accuray Inc, Sunnyvale, CA) [5-6], and 2) ExacTrac X-ray 6D system (BrainLAB AG, Feldkirchen, Germany), both of which employ an orthogonal pair of kilovoltage (kV) x-ray imaging systems mounted in fixed positions [7-9]. However, for conventional linacs equipped with an on-board kV imaging system and megavoltage (MV) electronic portal imaging device (EPID), it is more convenient and cost-effective to utilize the linac imaging systems to perform intra-fractional motion tracking [10].

Gurney-Champion OJ et al. proposed to use on-treatment kV projection images from different gantry angles to generate digital tomosynthesis (DTS) images, and then to use multiple DTS images to triangulate spine motion [11]. Similarly, Hazelaar C et al. developed a markerless spine tracking algorithm, also based on kV projection images. They matched the on-treatment kV images with corresponding templates generated from the planning CT, and with multiple registrations at different gantry angles they were able to determine the spine motion [12]. Both methods had inherent latencies of a few seconds due to the need for kV projection images from different gantry angles. Furthermore, the authors also pointed out some other drawbacks of using a single imaging system for motion tracking: 1) the accuracy of detecting movements occurring in the direction parallel to the kV radiation is low [11, 13]; 2) poor image quality due to increased attenuation when the x-ray beam rotates to the patient’s lateral direction can lead to poor detection accuracy [12]. Since the MV image axis is orthogonal to the kV counterpart, if acquired simultaneously and combined with the kV images, on-treatment MV images can be a natural solution to these problems. There have been several publications related to using MV images for monitoring lung and prostate tumor motion in real-time during treatment [14-20]. However, these techniques are mostly applied to 3D static or 3D conformal/arc plans, which have relatively large field apertures in MV images to facilitate image matching between the MV image and digitally reconstructed radiography (DRR) from planning CT. On the contrary, for spine SBRT treatment, since highly intensity modulated beams are used to create a sharp dose falloff to spare the radiosensitive spinal cord or cauda while maintaining adequate target coverage, the beam apertures are often small, which renders a very challenging task of tracking spine motion with the treatment beam images. In this paper we present a method to enlarge the field of view (FOV) of MV treatment beam images, called the enhanced synthetic treatment beam (ESTB) imaging, and illustrate its feasibility in detecting motion in spine SBRT treatment delivered with sliding window IMRT.

2. METHODS AND MATERIALS

2.A. Treatment plans

At our institution, an automated approach to intensity modulated treatment planning, the Expedited Constrained Hierarchical Optimization (ECHO) engine, has been incorporated into the Eclipse treatment planning system and is used to generate clinical plans for spine SBRT treatments [21]. A typical ECHO plan adopts a sliding window IMRT technique and is highly modulated, resulting in fast dose falloff. In this study, three clinical paraspinal treatment plans were randomly selected and transferred to a thorax anthropomorphic phantom "LUNGMAN" (Kyoto Kagaku Co., Japan). The phantom contained synthetic bones made of epoxy resin and soft tissue made of polyurethane. The phantom size was 43 x 40 x 48 cm³ with chest girth of 94 cm, and the weight was approximately 18 kg. The treatment plans used 6XFFF beams with a dose rate of 1400 MU/min, arranged at 9 posterior angles: 180⁰, 160⁰, 140⁰, 120⁰, 100⁰, 260⁰, 240⁰, 220⁰, 200⁰, respectively. In Fig. 1 we showed the phantom and one of the treatment plans in the axial view. The water equivalent depth in the anterior-posterior (AP) direction through the plan isocenter was about 20 cm. The total MU of ECHO plans are usually kept around 5-7 times of the prescription dose per fraction (in the unit of cGy). For example, for the plan shown in Fig. 1, the prescribed dose was 27 Gy in 3 fractions, and the average MU per beam for the plan was 593.3. Each field had 166 control points, and the delivery time was 25.4 seconds per field on average.

Fig. 1.

LUNGMAN phantom (left) and a typical ECHO paraspinal SBRT plan transferred to the phantom (right). The plan contained 9 posterior beams. The water equivalent depth in anterior-posterior direction at plan isocenter was about 20 cm.

2.B. Image acquisition

A Varian TrueBeam linear accelerator (version 2.7) with a PerfectPitch 6-DoF couch (Varian Medical Systems, Palo Alto, CA) was used to deliver the treatment plans. Before each plan delivery, a CBCT image of the phantom was taken and matched to the planning CT image to determine the corrected phantom position. The best alignment was achieved by applying translational corrections in AP, cranio-caudal (CC), and left–right (LR) directions, and rotational corrections in yaw, roll, and pitch directions using the 6-DoF couch. A second CBCT scan (the verification CBCT) was followed to ensure the image matching errors were within 0.2 mm and 0.2 degrees in all six directions. The kV projection images of the verification CBCT at every gantry angle were saved for later use as motion detection references. Full trajectory scan protocol was used for the CBCT scans and each scan resulted in about 896 projection images, with projection angle intervals less than 0.5 degrees.

After the initial alignment of the phantom, 2 mm shifts in LR and CC directions were introduced by moving the couch and the treatment plans were then delivered. During beam delivery, a Varian aS1200 EPID was used to acquire MV images continuously (Varian Medical Systems, Palo Alto, CA) at 150 cm source to imager distance (SID). The active area of the EPID was 43 × 43 cm² with 1280 × 1280 pixel arrays and pixel pitch of 0.336 mm. The images were automatically grabbed with VARIAN iTools Capture software at a rate of 11.6 Hz. Conventional corrections were applied by the software, including dark field and flood field calibrations, profile correction and bad pixel corrections.

2.C. Enhanced synthetic treatment beam imaging

With sliding window IMRT delivery, each MV frame usually contains many small MLC segments. In order to have a meaningful FOV size, it is often necessary to combine multiple frames into one large image, which is often referred to as a synthesizing process. Our ESTB process in this work had three major considerations: 1) beam intensity spatial variation correction; 2) scatter reduction; and 3) beam intensity temporal variation correction.

2.C.1. Beam intensity spatial variation

Fig. 2(a) is an example of a single MV frame acquired during irradiation of the thorax phantom. As we can see, some pixels appeared brighter than other pixels in this image. The intensity variation of the pixels was not only due to different attenuation at different locations, but also largely due to the variation of the local MLC leaf pair opening: where the MLC leaf pair had relatively larger opening, the pixels showed stronger image intensity. To further see this, we delivered the same beam to a uniform phantom (15 cm slabs of solid water), and the resulting single-frame MV image for the same MLC control point is shown in Fig. 2(b). Although the attenuation was uniform across the FOV for the solid water phantom, its image intensity showed similar variations to Fig.2(a) for the pixels at the corresponding locations. To eliminate the dependency of image intensity on beam aperture size, we used the solid water phantom images to normalize the on-treatment MV images by taking the pixel-wise ratio of the corresponding frames of the two. The gantry angle for the solid water irradiation was always kept at zero degrees (beam perpendicular to the phantom surface), and the source-to-surface distance (SSD) at 100 cm. Other plan parameters such as the collimator angle, MU, and dose rate were identical for the two deliveries. To synchronize the acquired MV frames between the two deliveries, the MLC control points of the frames were matched, which can be easily located from the frame header grabbed by the Varian iTools Capture software. The pixel intensity of the resulting image after normalization represented the relative attenuations of the thorax phantom (i.e. relative to the attenuation of 15 cm solid water for 6XFFF beam). The impact of this normalization will be discussed further in the following sections.

Fig. 2.

Impact of MLC opening on image intensity. Single-frame MV images were acquired with the same beam for the thorax phantom (a) and the solid water phantom (b). The pixel intensity variation in (b), although similar to (a), was mainly due to the different widths of the MLC leaf pair openings at different pixels, since the attenuation was uniform across the FOV for the solid water phantom.

2.C.2. Scatter

Since scatter increases with the irradiated volume, for the regions imaged with small MLC apertures, the scatter contribution can be considered very low. For examples, the scatter-to-primary ratios (SPRs) are about 0.01 for single-ray CT,0.05 to 0.15 for fan-beam and spiral CT and may be as large as 0.4 to 2.0 in CBCT [22]. SPRs for an MV beam are typically smaller than a kV beam for the same geometry [23]. The scatter contribution outside the MLC aperture can be easily removed by a simple thresholding method. Specifically, a binary mask was first created for each frame based on the pre-treatment solid water phantom images in the following way: for any frame, a threshold was set to a certain percentage of the frame’s maximum intensity αI_w,MAX. The value for any pixel with an intensity higher than the threshold was set to one, and the rest of the pixels’ values were set to zero, which effectively resulted in a binary mask for each frame. The masks were then applied to both the solid water phantom images and the thorax phantom images to remove the scatter contributions. In this study, we found empirically that the threshold should be set to at least 75% of the maximum frame intensity in order to avoid scatter-induced artifacts. A higher threshold will remove more scatter, leading to better image quality but smaller FOV. Fig. 3 shows an example of single-frame MV images before and after the thresholding. We can see that the pixels with relatively weak intensities were removed from the image and hence are not used in later image synthesizing process.

Fig. 3.

Single-frame MV image before (a) and after (b) 75% maximum intensity thresholding. Only those pixels with relatively strong signals were retained and later used to formulate the synthetic image.

To deal with the remaining scatter inside the MLC apertures after thresholding, we made the following assumption: within each small aperture, the scatter contribution to any pixel S_i is approximately proportional to the primary signal: S_i = r_iP_i, where r_i is dependent on the shape of field apertures, therefore the total signal is I_i = (1 + r_i)P_i. By taking the ratio of the pre-treatment solid-water phantom image to the on-treatment MV image, the scatter contribution can be effectively cancelled out. In our experiments, we first delivered the IMRT beams to 15 cm slabs of solid water, and the resulting MV images were stored and used to normalize the thorax phantom images. Specifically, the process can be described with Eq. (1), where $I_w(\overrightarrow{x},t)$ and $I_p(\overrightarrow{x},t)$ are MV images after thresholding for solid water phantom and thorax phantom at frame t, respectively, and $a(\overrightarrow{x},t)$ is the normalized image for frame t. Note $a(\overrightarrow{x},t)$ represented the relative attenuation between the solid-water phantom and the thorax phantom.

$$a(\overrightarrow{x},t)=\frac{I_w(\overrightarrow{x},t)}{I_p(\overrightarrow{x},t)}$$ (1)

2.C.3. Beam intensity frame variation

For any pixel located at $\overrightarrow{x}$, since there are multiple values $a(\overrightarrow{x},t)$ measured at different frames t, we can use the mean value over the frames to get the best estimation of the pixel value $\tilde{a}(\overrightarrow{x})$. However, due to the potential variation of the radiation beam output from frame to frame, it is necessary to apply a frame dependent correction before finding the mean values. This step can be summarized with Eqs. (2) and (3), where C(t) is the correction factor for frame t, N(t) is total number of non-zero pixles in each frame t, $T(\overrightarrow{x})$ is the total number of frames that have non-zero values at pixel $\overrightarrow{x}$, and $A(\overrightarrow{x})$ is the final synthetic image. To understand the meaning of Eq. (2), we first calculated the non-zero average value over frames for all pixels, which represented the expected value of each pixel from multiple measurements, $\tilde{a}(\overrightarrow{x})=\frac{1}{T(\overrightarrow{x})}{\sum }_{t}a(\overrightarrow{x},t)$. For individual frame t, the fluctuation relative to the average value at each pixel is $\sigma (\overrightarrow{x},t)=\frac{a(\overrightarrow{x},t)}{\tilde{a}(\overrightarrow{x})}$, and by averaging the fluctuation across all pixels in frame t, we can get the overall correction factor due to the beam output fluctuation for frame t: $C(t)=\frac{1}{N(t)}{\sum }_x\sigma (t,x)$. As an example, Fig. 4 shows the plot of frame-to-frame correction factors for one of the beams delivered. The factors typically ranged from 0.98 to 1.03 in our experiments. Although the frame-to-frame correction factors were small, synthesizing without the correction could lead to noticeable artifacts in the final images, see the example in Fig. 5(a).

Fig. 4.

An example of frame-to-frame correction factors.

Fig. 5.

Artifacts in the synthetic image due to frame-to-frame beam output variation. Compared with the ground-truth kV x-ray image (c), the synthetic image without the frame-to-frame correction (a) showed artificial structures as indicated by the arrow; (b) is the image generated with the proposed ESTB imaging method, which is free of such artifacts.

$$C(t)=\frac{1}{N(t)}{\sum }_x\frac{a(\overrightarrow{x},t)}{\frac{1}{T(\overrightarrow{x})}{\sum }_{t}a(\overrightarrow{x},t)}$$ (2)

$$A(\overrightarrow{x})=\frac{1}{T(\overrightarrow{x})}{\sum }_{t}a(\overrightarrow{x},t)∕C(t)$$ (3)

2.D. Analysis of ESTB image-based motion detection

For any static-gantry IMRT beam at a particular gantry angle, a series of on-treatment ESTB images are generated at different time points, and successively compared with the reference CBCT projection to see if any motion has occurred since the initial setup. Note the reference projection is from the verification CBCT scan and at the matched gantry angle. For example, for a treatment beam at a 100-degree gantry angle, the reference 2D CBCT projection should be acquired at a 190-degree gantry angle, so that the kV radiation central axis is aligned to the MV radiation central axis. Since the CBCT scans are acquired with full trajectory protocol (~896 projections for 360-degree gantry rotation), there always exists a reference projection for any treatment beam angle with an accuracy <0.3 degrees. To quantify any motion during treatment, a rigid 2D/2D image registration based on maximum mutual information (MI) was performed between each ESTB image and the corresponding reference CBCT projection. In this study, three clinical SBRT plans were delivered, resulting in a total of 760 ESTB images (about 85 ESTB images per beam angle). The ESTB images obtained from the same beam angle were used to determine the average shifts for this particular angle, and the root mean square errors (RMSEs) of the calculated shifts were also assessed.

3. RESULTS

In Fig. 6 we show several examples of the ESTB images acquired with one of the delivered beams, which had total of 350 frames of MV images. From left to right on the top panel, are the ESTB images generated with frames 1-100, 101-200, 201-300, and 1-350, respectively. On the bottom are the reference images, which came from the same CBCT projection at the matched beam angle. Each generated ESTB image was registered to the reference image, and the red contours indicated the field apertures projected onto the kV projection image based on the registration. Although the contrast resolution is not as good as the CBCT projection, many anatomical features such as the vertebral body edges and the intervertebral discs are clearly visible in the ESTB MV images.

Fig. 6.

Comparisons between enhanced synthetic treatment beam (ESTB) images (top) and corresponding kV projection images (bottom): the vertebral body edges and intervertebral disks are clearly visible in the ESTB images. The red contours are the ESTB image apertures projected to the kV images based on image registration to show the corresponding anatomical structures.

To further test the feasibility of using ESTB images for spine motion tracking, we analyzed the images acquired for several gantry angles after the 2 mm LR/CC couch shifts were made. For each beam, ESTB images were generated using 50 frames at 10 frame intervals, i.e., the first ESTB image frame was generated from original frames 1-50, the second ESTB image frame was generated from frames 11-60, and so on. On average, about 85 ESTB images per beam angle were obtained with the three IMRT plans, and the images were registered to the corresponding CBCT projections to calculate 2D translations. The results are summarized in Table I. We found the errors between the detected shifts and the designed shifts are all within a millimeter for all gantry angles. No clear trend of angle dependency for the tracking accuracy was found in this phantom study, partially due to 1) uncertainties such as gantry sag, EPID sag, and central axis (CAX) motion; 2) the limited phantom size hence similar image quality for all gantry angles.

Table I.

Motion tracking accuracy vs. gantry angle. The designed shifts were 2 mm in both LR and CC directions, and the detected shifts are the averaged results for the beams with the same gantry angles from three IMRT plans, expressed in the form of (LR shift, CC shift). The RMSEs are expressed similarly.

Gantry angle	180	160	140	120	100
Detected shifts in LR/CC directions (mm/ mm)	2.31/2.25	1.94/ 2.32	2.47/2.29	2.13/2.39	2.28/2.33
RMSEs in LR/CC directions (mm/mm)	0.35/0.28	0.32/0.35	0.63/0.44	0.55/0.51	0.69/0.42

4. DISCUSSION:

Compared with the motion monitoring techniques that use kV imaging only [11, 12], the proposed ESTB MV imaging has the same level of submillimeter precision in detecting spine motion. The technique also has a significant advantage in the sense that it can catch potential movements that have larger dosimetric impact to the patient, which kV imaging may fail to capture. For example, if a patient moves about 2 mm laterally during a posterior-anterior (PA) beam delivery, this 2 mm motion could hardly be detected by 2D kV imaging since the kV radiation is also from the lateral direction. If the spinal cord moves into the treatment beam due to the motion, a severe myelopathy could happen to the patient. On the contrary, the 2 mm lateral motion can be easily detected by the MV imaging since the motion is perpendicular to the MV beam. Although MV imaging is insensitive to the motion along the radiation beam direction, the dosimetric impact of such motion is minor, because it is mainly governed by the inverse square law and the motion magnitude is significantly small compared to the source to target distance (STD).

The problem of single-frame MV imaging for highly modulated IMRT treatment is twofold: (1) the aperture may be too small to reveal any meaningful anatomical features for reliable matching, which is the main reason that MV imaging was rarely used to monitor IMRT treatments in the past; (2) noise and scatter are imminent in single-frame MV images. Although simple integrals of MV frames can increase FOV, this often results in a low-quality image. The proposed ESTB imaging method significantly reduces scatter, and at the same time, also looks for the best statistic estimation of the pixel value from multiple frames, leading to improved image quality necessary for intensity-based image matching.

The FOV of ESTB images combined from 50 frames (about 4.3 s latency) is usually large enough to provide adequate anatomical features for image matching based on our phantom study. The more frames of MV images used for generating the ESTB image, the more anatomical structures the ESTB image can reveal. If all frames acquired from a single beam are combined to formulate a single ESTB image, the image would become similar to a beam’s-eye-view (BEV) MV image acquired for a 3D conformal plan. However, an inherent limitation of the ESTB imaging is its temporal resolution. Continuous motion will blur the ESTB images, making them unsuitable for motion detection. For this reason, the proposed ESTB imaging is only aimed to monitor patient random bulk shifts. Whenever a motion is detected by an ESTB frame, the treatment can be stopped, and the patient can be re-aligned. Furthermore, the selection of the number of frames in this study was to make sure the FOVs were large enough for all ESTB images, therefore not optimized for the best temporal resolution for individual beams. Clearly, the fewer frames of MV images used in ESTB imaging, the better temporal resolution that may be achieved for intrafraction motion monitoring, but it is always limited by the FOV for robust image matching. To combat this challenge, we are also investigating a machine learning-based image matching method, which could potentially enhance the sensitivity and robustness of image matching, making it possible to use images with smaller FOVs for motion detection, leading to improved temporal resolution.

The focus of this technique is to demonstrate that ESTB imaging could generate an image feasible for 2D image matching, however, there are several factors that can potentially affect the matching results, such as gantry and EPID sag, coincidence problems between MV and kV isocenters, and CAX wobbling [24-25], which are not studied because they do not have a direct impact on the ESTB image quality. Part of these factors are systematic and therefore can be corrected. A more rigorous performance evaluation should include systematic calibrations/corrections for these factors for both kV and MV imaging systems, which is out of the scope of this technical note. Our current work only showed the feasibility of using ESTB imaging to provide motion information in a 2D plane, but it can also be combined with on-treatment kV imaging to provide more accurate 6D motion information.

5. CONCLUSION:

ESTB imaging can significantly increase the FOV of MV images acquired during sliding window IMRT delivery. A phantom study based on three clinical plans suggested that ESTB images, when used to detect spine motion, can achieve submillimeter precision in the direction perpendicular to the beams.