Markerless tumor tracking using fast‐kV switching dual‐energy fluoroscopy on a benchtop system

Abstract

Purpose

To evaluate markerless tumor tracking (MTT) using fast‐kV switching dual‐energy (DE) fluoroscopy on a bench top system.

Methods

Fast‐kV switching DE fluoroscopy was implemented on a bench top which includes a turntable stand, flat panel detector, and x‐ray tube. The customized generator firmware enables consecutive x‐ray pulses that alternate between programmed high and low energies (e.g., 60 and 120 kVp) with a maximum frame rate of 15 Hz. In‐house software was implemented to perform weighted DE subtraction of consecutive images to create an image sequence that removes bone and enhances soft tissues. The weighting factor was optimized based on gantry angle. To characterize this system, a phantom was used that simulates the chest anatomy and tumor motion in the lung. Five clinically relevant tumor sizes (5–25 mm diameter) were considered. The targets were programmed to move in the inferior‐superior direction of the phantom, perpendicular to the x‐ray beam, using a cos⁴ waveform to mimic respiratory motion. Target inserts were then tracked with MTT software using a template matching method. The optimal computed tomography (CT) slice thickness for template generation was also evaluated. Tracking success rate and accuracy were calculated in regions of the phantom where the target overlapped ribs vs spine, to compare the performance of single energy (SE) and DE imaging methods.

Results

For the 5 mm target, a CT slice thickness of 0.75 mm resulted in the lowest tracking error. For the larger targets (≥10 mm) a CT slice thickness ≤2 mm resulted in comparable tracking errors for SE and DE images. Overall DE imaging improved MTT accuracy, relative to SE imaging, for all tumor targets in a rotational acquisition. Compared to SE, DE imaging increased tracking success rate of small target inserts (5 and 10 mm). For fast motion tracking, success rates improved from 23% to 64% and 74% to 90% for 5 and 10 mm targets inserts overlapping ribs, respectively. For slow moving targets success rates improved from 19% to 59% and 59% to 91% in 5 and 10 mm targets overlapping the ribs, respectively. Similar results were observed when the targets overlapped the spine. For larger targets (≥15 mm) tracking success rates were comparable using SE and DE imaging.

Conclusion

This work presents the first results of MTT using fast‐kV switching DE fluoroscopy. Using DE imaging has improved the tracking accuracy of MTT, especially for small targets. The results of this study will guide the future implementation of fast‐kV switching DE imaging using the on‐board imager of a linear accelerator.

Short abstract

1. Introduction

Lung tumor motion during radiation therapy (RT) may be significant and unpredictable. Examples of methods to reduce tumor motion include abdominal compression^{1, 2} and breath‐hold.^{3, 4} Additionally, methods have been proposed to track lung tumors during treatment for online adaptation. These methods include the use of external surrogates,^{5, 6} radiopaque fiducial markers,^{7, 8, 9} and electromagnetic transponder beacons.^{10, 11, 12} In the latter cases, the implantation of these markers is invasive and has been associated with an increased risk of pneumothorax.⁷ Furthermore, marker migration may result in surrogacy errors.^{13, 14, 15}

Markerless tumor tracking (MTT) using x‐ray imaging may overcome the risks associated with invasive methods. Several researchers have been considering MTT in the lung using MV,^{16, 17, 18} and kV imaging.^{19, 20, 21, 22, 23} However, a major difficulty with MTT is that tumor overlapping bone may not be detectable on x‐ray projections.¹⁷ This effect makes MTT especially challenging for rotational acquisitions as many projections may have tumor/bone overlap. In these cases, improved localization methods are required. For example, Yang et al.²¹ obtained cone beam computed tomography (CBCT) projections, and subtracted digitally reconstructed radiographs (DRRs) of bone to highlight the tumor for tracking. In a separate study, van Sornsen De Koste et al.²² applied a band‐pass spatial filter to enhance tumor visibility of CBCT projection images.

An imaging technique that may mitigate the effect of tumor/bone overlap is dual‐energy (DE) fluoroscopy. Briefly, this method consists of acquiring two sets of x‐ray images at different energy settings, for example, 120 and 60 kVp. These images are subsequently combined to produce bone suppressed images using weighted logarithmic subtraction (WLS).^{24, 25} The resulting images improve the visibility of tumors that are hidden due to bony obstructions.^{24, 26, 27} Previously, researchers have evaluated the potential of sequential DE imaging for MTT using the on‐board imager (OBI) of a linear accelerator.^{28, 29} In these studies, fluoroscopic images were obtained sequentially at two different x‐ray energies and retrospectively processed to create a DE subtraction sequence. While this approach demonstrated the efficacy of DE fluoroscopy, the clinical use of sequential fluoroscopy images is limited due to respiratory and organ motion artifacts. An alternative method involves DE fluoroscopy using fast‐kV switching in which the x‐ray tube potential is alternated at a high frame rate. Such an approach would limit motion artifacts and provide real‐time DE imaging for MTT. In this study, MTT results of DE fluoroscopy using fast‐kV switching on a bench top system are presented.

2. Materials and Methods

2.A. Phantom

To simulate lung tumor motion, the CIRS dynamic thorax motion phantom (CIRS Inc. Norfolk, VA) was utilized. The phantom approximates the human thorax both in size and structure. The body of the phantom is composed of tissue and lung‐equivalent epoxy materials. Additionally, the phantom contains a three‐dimensional (3D) anthropomorphic spine and ribs constructed with cortical and trabecular bone. A lung‐equivalent epoxy rod contains a soft tissue target insert to simulate the tumor (Fig. 1). Five different spherical target inserts with diameters of 5, 10, 15, 20, and 25 mm were used in this study. A motion controller is connected to the lung insert, and is capable of producing complex 3D motions by utilizing simultaneous translational and rotational movements with preprogrammed periodic waveforms. In this study, the cos⁴ waveform was used to simulate respiratory motion in the inferior‐superior direction.

Figure 1.

CIRS motion phantom assembly; torso phantom with lung insert, motor with independent surrogate control and actuator box. [Color figure can be viewed at wileyonlinelibrary.com]

2.B. Bench top system

Images were acquired using a fast‐kV switching real‐time fluoroscopy prototype developed at Varian Medical Systems (Palo Alto, CA, USA). This system was implemented on a bench top that includes a turntable stand, amorphous silicon (a‐Si) flat panel detector (PaxScan 4030CB, Varex Imaging, Salt Lake City, UT), and x‐ray tube (GS 1542, Varex Imaging, Salt Lake City, UT) with custom firmware and software (EPS 45‐80, EMD Technologies, Saint‐Eustache, Quebec, Canada). The system emulates the hardware on a clinical OBI. The fast‐kV switching firmware enables consecutive x‐ray pulses that alternate between programmed energies. In this study, alternating 60 and 120 kVp x‐rays³⁰ were used. These energies were chosen based on a previous study by Haytmyradov et al.³⁰ which demonstrated that energy combinations of 140–60, 130–60, and 120–60 kVp produced the highest (and statistically equivalent) signal‐difference‐to‐noise ratios (SNDR) per unit dose.

The maximum frame rate of 15 frames per second (fps) was used in all studies. This results in DE images with an effective frame rate of 7.5 Hz. The mA setting for each pulse was adjusted to balance the air dose between the 60 kVp (60 mA) and 120 kVp (15 mA) settings. The detector and x‐ray source were fixed on the bench top, whereas the motion phantom was mounted vertically on the turntable stand as shown in Fig. 2. The turntable rotates at 6 degrees per second, allowing a full 360ᵒ scan in approximately one minute. The source‐to‐axis distance (SAD) and source‐to‐image distance (SID) were 100.8 and 148.3 cm, respectively, and were fixed for all experiments. Acquired x‐ray images were encoded in a 16 bit unsigned‐integer having dimensions of 1024 × 768 and a pixel size of 0.388 mm (2 × 2 binned). These images were subsequently saved in the Varian proprietary XIM format.

Figure 2.

Experimental setup of fast‐kV switching imager prototype. Left, flat panel detector. Middle, CIRS motion phantom positioned vertically and mounted on turntable to simulate cone‐beam acquisition. Right, fast‐kV switching x‐ray tube. [Color figure can be viewed at wileyonlinelibrary.com]

Following image acquisition, DE images were produced offline using in‐house software developed in MATLAB R2018a (MathWorks, Natrick, MA, USA). Consecutive high (120 kVp) and low (60 kVp) energy images were combined to yield a bone suppressed image utilizing the WLS method defined as follows:

$$\text{ln}{I}_{\mathit{DE}}=\text{ln}{I}_{\mathit{High}}-w_s·\text{ln}{I}_{\mathit{Low}}$$ (1)

where ln I _DE, ln I _High, and ln I _Low represent the natural logarithm of reconstructed DE high and low kVp x‐ray projections, respectively. For comparison purposes, single‐energy (SE) imaging sequences were obtained by selecting only the 120 kVp images. Thus, the corresponding SE frame rate is 7.5 Hz, which is the same as DE imaging.

2.C. Markerless tumor tracking

To assess the performance of MTT on a fast‐kV switching DE images, template matching software (RapidTrack v3.0.3, Varian, Palo Alto, CA — noncommercial research software) was utilized. The MTT software is based on the work by Mostafavi et al.³¹ which localizes the position of the target in x‐ray projections³¹ using a template generated for each imaging angle. The location of the tumor is identified by scanning the template throughout the search window defined at the isocenter of the x‐ray projection within a 23 × 23 mm bounding box. For each iteration, the normalized cross‐correlation (NCC) between the template and image is computed, which results in a match score surface (Fig. 3). The offset at which the NCC score has the maximum value represents potential match position. The strength of this peak relative to NCC values away from the peak, called side lobe values, is quantified by the peak‐to‐side lobe ratio (PSR). Block et al.³² showed that PSR values can be used as a predictor of successful template matches, and PSR > 3 can reduce the false detection rate.³²

Figure 3.

Demonstration of markerless tumor tracking software. Left, tumor template generated offline is scanned across the search window. Right, the normalized cross‐correlation (NCC) surface after the search. The match location is identified at the peak location of the NCC surface. [Color figure can be viewed at wileyonlinelibrary.com]

2.D. Experimental methods

Several experiments were performed on the bench top system to optimize imaging parameters as well as compare tracking accuracy of DE vs. SE imaging. All experiments were performed using target motion that was programmed to follow cos⁴ waveform trajectory in the inferior‐superior direction of the thorax that oscillates 15 mm peak‐to‐peak amplitude in 5 s, unless otherwise stated. The direction of the x‐ray beam was perpendicular to the direction of insert motion. The fast‐kV switching x‐ray generator settings were fixed to 120 kVp, 15 mA, 20 ms and 60 kVp, 60 mA, 20 ms for all experiments.

2.D.1. Image registration and weighting factor optimization

Although fast‐kV switching considerably reduces motion artifacts, there is non‐negligible time delay (~67 ms) between high and low x‐ray pulses. For a cone beam type acquisition, the gantry will rotate ~0.4° during this time period. As such, subtraction of the high/low energy pair can result in residual bone images due to the offset between these images. To remove these artifacts, a rigid image registration algorithm was used to align the high/low kVp images before DE subtraction using mutual information (MI) metric.³³ MI has been successfully used in mono‐ and multimodality image registration tasks^{34, 35} and does not require the definition of landmarks or assumes a linear relationship among gray values in the images. The MI metric was found fit well to our problem due to gray value differences between high and low kVp images. The registration procedure was fully automated by maximizing MI between high and low kVp images.

Once the images were aligned, the optimal weighting factor ($w_S$) defined in Eq. (1), was determined using an iterative method by minimizing contrast between bone and background regions of interest (ROI).^{26, 29, 30} In‐house software was developed to interactively adjust the weighting factor to produce the highest quality bone suppressed image for each pair of high/low images. Due to the rotational geometry, the weighting factors can change as the x‐rays penetrate larger thicknesses of the phantom having different density materials (such as vertebrae). To accommodate these variations, the weighting factor was optimized after every 20 frames (~8°).

2.D.2. Impact of slice thickness on tracking

The effect of CT slice thickness on template quality and therefore MTT performance was evaluated using a fixed angle fast‐kV switching projections obtained for all five simulated tumors at an oblique angle of 45ᵒ. To generate templates, the CIRS phantom was scanned on the Siemens SOMATOM Open AS (Siemens Healthineers, Forchheim, Germany). These images were subsequently contoured using the Eclipse software (Varian, Palo Alto, CA) by a trained physicist and the contours then were used to generate templates for tracking software using noncommercial RapidTrack‐Planning software (RTP version 1.12.2, Varian Medical Systems). To evaluate the impact of the CT slice thickness on template generation, these images were reconstructed using 0.75, 1.0, 1.5, 2.0, and 3.0 mm slice thicknesses. For both DE and SE image sequences, MTT was evaluated for each tumor size/CT slice thickness combination.

2.E. Tracking on rotational acquisitions

Motion and location of tumor (e.g., overlapping high‐density objects such as vertebrae or ribs) can affect tracking accuracy of MTT.²² To study these effects we considered three different tumor motion scenarios using rotational acquisitions as follows: static target (amplitude 0), slow‐motion target (peak‐to‐peak amplitude 15 mm, period 5 s), and fast motion target (peak‐to‐peak amplitude 15 mm, period 2.5 s). For each tumor insert, the three described motions were evaluated separately on a cone‐beam acquisition by rotating the turntable 360°. Similar to above, MTT software was used to track each of the tumors on SE and DE images.

2.E.1. Metrics

To perform quantitative assessment of template‐based tumor tracking, three metrics were considered: tracking accuracy, PSR score, and MTT success rate. Each of these is described below.

The tracking accuracy for SE and DE imaging was calculated using the mean absolute error (MAE) between the predicted and expected location of the tracked tumor demonstrated in Fig. 4. To calculate MAE, two‐dimensional (2D) tracking points were extracted from the tracking software report and fit to the cos⁴ function with the amplitude fixed based on the programmed motion. The period was fixed to the insert's motion in the corresponding experiment. The phase and offset of the fit line were left floating and extracted from the fit automatically. Subpixel tracking accuracy was then calculated using MAE of the tracked and expected points.

Figure 4.

Demonstration of mean absolute error calculation. Measured target location (green‐triangle) is produced from markerless tumor tracking software report. Expected target location (red‐dotted line) is extracted by fitting observation to cos⁴ waveform with fixed amplitude and period. [Color figure can be viewed at wileyonlinelibrary.com]

The PSR score is related to the strength of template matching algorithm.³² For each iteration, the normalized cross‐correlation (NCC) between template and image was computed to produce a 2D match score surface. The PSR values for SE and DE images were compared among the different experiments. A successful match was defined when PSR > 3 and match score threshold > 0.3. MTT success rate was then calculated as the ratio of successful matches to the total number of frames in the experiment.

3. Results

3.A. Image registration and weighting factor optimization

Prior to DE subtraction, low and high kV images were aligned using a rigid image registration technique. Briefly, the low‐energy image was iteratively shifted, in the units of pixels, in the horizontal direction to maximize mutual information between the shifted frame and the high‐energy image. For each frame, the search for the required shifts used the results from the previous frames as a starting point. In general, the range of shifts necessary to register the two images was ±2 pixels, depending on the direction of rotation. This method was sufficient to register rotational acquisitions where motion artifacts are dominated by the vertical edges of the bony structures. Figure 5 depicts an example of this approach.

Figure 5.

Demonstration of single energy (120 kVp), dual energy (DE) subtraction, and DE subtraction after registration for 20 mm target. Left, SE image tumor is barely visible behind vertebrae. Middle, DE subtraction enhances visibility of tumor, but motion artifacts are seen due to rotation. Right, rigid registration eliminates motion artifacts.

The optimal weighting factor, w_s used to subtract the low‐energy images from the high‐energy images ranged from 0.62 to 0.79 across the cone‐beam projections (Fig. 6). On average, the weighting factor was 0.69 ± 0.04 and was periodically changing based on the position of target inserts relative to bone. The largest value was observed when the turntable rotated −80° which corresponded to target inserts overlapping with the spine. The smallest weighting factor was observed when tumor was overlapping only the ribs.

Figure 6.

Demonstration of the weighting factors as a function of imaging angle. Near‐periodic change is observed as x‐rays penetrate various thicknesses of material. [Color figure can be viewed at wileyonlinelibrary.com]

3.B. Impact of Slice thickness on tracking

The resolution of a template depends on the CT slice thickness at which it is reconstructed. To optimize template quality, several experiments were performed by varying the CT slice thickness used to generate templates. These templates were subsequently used in MTT software to track individual targets. To avoid the overwhelming amount of data due to combinations of variables, only tracking success rate and tracking errors are presented. Table 1 depicts tracking MAE and success rates for SE and DE images, where the columns represent CT slice thicknesses and rows correspond to tumor diameter.

Table 1.

Comparison of markerless tumor tracking tracking results between single enerfy (SE) and dual energy (DE) imaging for various computed tomography (CT) slice thicknesses used to reconstruct templates. The tracking success rate is listed in parenthesis. Tumor diameters are listed in rows, whereas the corresponding slice thickness is tabulated in columns. Note: The template for the 5 mm target that was generated using a 3 mm slice thickness failed to track on the SE images, and is denoted by “N/A.”

Tumor Diameter (mm)	CT slice thickness (mm)
	0.75	1.0	1.5	2.0	3.0
Mean absolute error (mm) for SE imaging
5	5.83 ± 0.98 (54%)	6.04 ± 0.99 (46%)	4.47 ± 1.61 (18%)	0.62 ± 0.14 (6%)	N/A (0%)
10	0.40 ± 0.06 (84%)	0.39 ± 0.06 (85%)	0.40 ± 0.06 (84%)	0.42 ± 0.06 (85%)	0.53 ± 0.06 (100%)
15	0.12 ± 0.02 (100%)	0.12 ± 0.02 (100%)	0.12 ± 0.02 (100%)	0.12 ± 0.01 (100%)	0.18 ± 0.03 (100%)
20	0.14 ± 0.02 (100%)	0.14 ± 0.02 (100%)	0.14 ± 0.02 (100%)	0.15 ± 0.02 (100%)	0.18 ± 0.03 (100%)
25	0.15 ± 0.02 (100%)	0.15 ± 0.02 (100%)	0.17 ± 0.02 (100%)	0.17 ± 0.02 (100%)	0.17 ± 0.03 (100%)
Mean absolute error (mm) for DE imaging
5	0.18 ± 0.03 (100%)	0.18 ± 0.03 (100%)	0.18 ± 0.03 (100%)	0.21 ± 0.03 (100%)	0.98 ± 0.30 (90%)
10	0.13 ± 0.02 (100%)	0.13 ± 0.02 (100%)	0.13 ± 0.02 (100%)	0.13 ± 0.02 (100%)	0.16 ± 0.02 (100%)
15	0.16 ± 0.02 (100%)	0.17 ± 0.02 (100%)	0.17 ± 0.02 (100%)	0.17 ± 0.02 (100%)	0.18 ± 0.02 (100%)
20	0.14 ± 0.02 (100%)	0.14 ± 0.02 (100%)	0.14 ± 0.02 (100%)	0.14 ± 0.02 (100%)	0.17 ± 0.02 (100%)
25	0.09 ± 0.01 (100%)	0.09 ± 0.01 (100%)	0.09 ± 0.01 (100%)	0.09 ± 0.01 (100%)	0.09 ± 0.01 (100%)

For SE imaging, the 5 mm target had the largest tracking error regardless of CT slice thickness. The tracking success rates of this target were 54% (N = 43), 46% (N = 37), 18% (N = 14), 6% (N = 5), and 0% (N = 0) for 0.75, 1.0, 1.5, 2.0, and 3.0 mm slice thicknesses, respectively. For the larger targets (≥10 mm) increasing the slice thickness from 0.75 to 2.0 mm had only a minor effect on the tracking accuracy, whereas the 3.0 mm slice thickness resulted in the largest MAE values. For these targets, the tracking success rate was nearly 100%.

On DE images, the MAE values were comparable, with the exception of the 5 mm target with the 3 mm CT slice thickness. The smallest MAE values for the 5 mm target were obtained using slice thicknesses of ≤1.5 mm. For 10, 15, and 20 mm targets tracking errors were comparable for all slice thicknesses ≤2 mm, whereas the 3 mm slice thickness resulted in highest tracking error. For the largest target, the tracking errors were nearly the same regardless of CT slice thickness.

Generally DE imaging performed better than SE for smaller targets. For the 5 mm target, DE tracking accuracy was an order of magnitude better than the SE imaging regardless of CT slice thickness used to generate the template. For the 10 mm target, the DE errors were approximately three times less than the SE errors. The DE and SE tracking MAE were comparable for the 15 mm target across all slice thicknesses. For larger targets (≥20 mm) DE tracking errors were comparable or better than the SE imaging. Moreover, the tracking success rates were > 90% for all combinations of tumor size and slice thickness on DE images, whereas the success rates for SE images were relatively lower for the smaller targets.

3.C. Tracking on rotational acquisitions

This section highlights results of MTT for SE and DE imaging using fast‐kV switching on full 360° rotational acquisition. As described in Section 3.2.4.2A DE images were produced for each gantry angle by registering images and optimizing weighting factor before subtraction. DE, along with SE images, were processed using MTT software. All tumor templates were generated using a 0.75 mm CT slice thickness.

To quantify the localization accuracy of MTT, the MAE of the tracked target position was computed. Due to variations of tracking performance as a function of the varying degrees of bone overlap, MAE was further split into two categories: overlap of tumor with rib vs spine. Figure 7 depicts the tracking results of the 20 mm tumor for SE and DE images over one complete rotation (60 s). The white areas are where the tumor overlaps rib, whereas the cross‐hatched areas are where the tumor overlaps of spine. Table 2 depicts results of this study for SE and DE images where the three motion categories (depicted in columns) are considered for all tumor sizes (rows).

Figure 7.

Figure depicting an example of tumor tracking across various bone overlap regions of phantom for SE (top) and DE (bottom). Green triangles and red lines represent measured and expected relative tumor positions (y‐axis), respectively, as gantry rotates (x‐axis). Shaded and nonshaded regions represent tumor overlapping spine and ribs, respectively. [Color figure can be viewed at wileyonlinelibrary.com]

Table 2.

Tracking accuracy comparison between single energy (SE) and dual energy (DE) images for various tumor motions and various tumor sizes. The tracking success rate is listed in parenthesis. Tumor diameters are listed in rows, whereas the corresponding tumor motion is tabulated in columns. Left and right three columns represent results SE and DE images, respectively. A greater improvement in MAE values is observed using DE imaging for cases where tumor overlapped the spine.

Tumor diameter (mm)	SE tracking errors (mm) (Tracking success rate)			DE tracking errors (mm) (Tracking success rate)
	Static	Slow motion	Fast motion	Static	Slow motion	Fast motion
Tumor overlapping ribs
5	N/A (0%)	0.18 ± 0.12 (19%)	0.17 ± 0.10 (23%)	0.51 ± 0.21 (33%)	0.32 ± 0.12 (59%)	0.44 ± 0.21 (64%)
10	0.84 ± 0.29 (81%)	1.31 ± 0.38 (79%)	0.78 ± 0.24 (74%)	0.21 ± 0.12 (88.3%)	0.35 ± 0.12 (91%)	0.25 ± 0.08 (90%)
15	0.21 ± 0.06 (88%)	0.27 ± 0.08 (90%)	0.24 ± 0.05 (90%)	0.15 ± 0.03 (85%)	0.15 ± 0.02 (93%)	0.17 ± 0.02 (92%)
20	0.13 ± 0.02 (96%)	0.18 ± 0.02 (96%)	0.19 ± 0.02 (95%)	0.11 ± 0.02 (90%)	0.25 ± 0.08 (93%)	0.18 ± 0.02 (94%)
25	0.21 ± 0.02 (96%)	0.17 ± 0.02 (97%)	0.18 ± 0.02 (96%)	0.09 ± 0.01 (95%)	0.10 ± 0.01 (95%)	0.15 ± 0.02 (97%)
Tumor overlapping spine
5	N/A (0%)	0.21 ± 0.14 (3%)	0.17 ± 0.07 (3%)	0.54 ± 0.17 (24%)	2.12 ± 0.67 (38%)	0.65 ± 0.25 (54%)
10	0.33 ± 0.04 (50%)	2.81 ± 0.70 (40%)	2.07 ± 0.59 (38%)	0.38 ± 0.17 (79%)	0.18 ± 0.02 (83%)	0.21 ± 0.02 (86%)
15	0.69 ± 0.17 (66%)	1.30 ± 0.29 (77%)	2.22 ± 0.40 (73%)	0.20 ± 0.03 (86%)	0.19 ± 0.02 (89%)	0.20 ± 0.02 (87%)
20	0.52 ± 0.06 (87%)	0.93 ± 0.20 (85%)	1.01 ± 0.19 (88%)	0.15 ± 0.02 (92%)	0.16 ± 0.02 (95%)	0.29 ± 0.11 (97%)
25	0.19 ± 0.02 (90%)	0.53 ± 0.13 (81%)	0.39 ± 0.07 (79%)	0.13 ± 0.02 (88%)	0.17 ± 0.02 (95%)	0.17 ± 0.02 (95%)

Generally, DE imaging had improved tracking accuracy in all cases, except the 5 mm target. On SE images, when tumor was overlapping ribs, the tracking success rate for the 5 mm target (0%, 19%, and 23% in static, slow, and fast motions) were lower than DE images (33%, 59%, and 64% in static, slow, and fast motions). Lower tracking success rates were also observed for the same target on SE images (0%, 3%, and 3% in static, slow, and fast motions) versus DE images (24%, 38%, and 54% in static, slow, and fast motions) when tumor overlapped the simulated spine of the phantom, producing an artificially better tracking accuracy. This effect is the result of the few data points perfectly matching the expected trajectory.

For the 10 mm target, when it was overlapping ribs, DE imaging increased the tracking success rate from 81% to 88%, 79% to 91%, and 74% to 90% for stationary, slow, and fast targets, respectively. For the same target when it was overlapping spine, greater improvements were observed with DE imaging (50% vs 79%, 40% vs 83%, and 38% vs 86%, or stationary, slow, and fast targets, respectively). For the larger targets (≥15 mm), the success rate using SE imaging ranged from 66% to 97%, with the lowest success rate occurring for the stationary 15 mm target overlapping spine. Similarly, using DE imaging, the success rate for the larger targets ranges from 85% to 97%. In general, larger improvements were observed for tumors overlapping spine.

The match score PSR values were calculated for all five different target inserts and three categories of motion (Fig. 8). For all tumor sizes and motion categories, a minimum improvement of 15% was observed with DE imaging. In both SE and DE imaging PSR values were >3 for targets having diameter greater than 10 mm.

Figure 8.

PSR values for dual energy (DE) and single energy (SE) images for various tumor inserts. Gray and green bars represent SE and DE peak‐to‐side lobe ratio values. Static, slow, and fast moving targets are shown in top, middle, and bottom figures, respectively. [Color figure can be viewed at wileyonlinelibrary.com]

4. Discussion

In this study, a bench top DE fluoroscopy system using fast‐kV switching was implemented and evaluated for MTT. Although similar studies were conducted to assess MTT using kV fluoroscopy images^{19, 20, 21, 22} and sequential DE kV images,^{28, 29} to our knowledge, MTT has not been studied using a fast‐kV switching system.

The effect of CT slice thickness on template generation and its impact on MTT was evaluated. Using the smallest available slice thickness on the CT scanner may be the simplest approach, however, the results of our study indicate that there may be a range of optimal slice thicknesses for each tumor size. For SE imaging, the smallest CT slice thickness may be preferable for tumor diameters ranging between 5 and 10 mm, whereas for larger tumors (15–25 mm), a CT slice thickness <3 mm is desirable. For DE imaging, a slice thickness of <3 mm would result in comparable tracking error for all targets <25 mm in diameter. For the 25 mm target, and potentially larger, the tracking errors are comparable regardless of slice thickness on both SE and DE imaging.

Tracking accuracy was compared between SE and DE images using inserts of different sizes over a full rotation. For small targets, DE subtraction improved tracking accuracy. In particular, the largest gains were noted for the 5 and 10 mm spherical inserts. However, it is important to note that even for the smallest target (5 mm), the gains associated with DE imaging resulted in less than satisfactory tracking results. In these cases, other modalities may have advantage of being combined with DE tracking such as digital tomosynthesis.³⁶ For targets ≥15 mm, the tracking accuracy seems to plateau with little difference observed between SE and DE tracking. However, the higher PSR and improved tracking accuracy of DE over SE may offer advantages in the clinical setting when tumors may have less than ideal shapes and varying densities. Additionally, in the clinical setting, the ground truth will not be readily available, and hence a surrogate such as PSR may be used to indicate the accuracy and success of tracking.³⁷

A number of studies have been conducted on MTT using both SE and DE imaging. Lewis et al.¹⁹ tracked tumors using digitally reconstructed radiographs (DRR) from four‐dimensional computed tomography and searched for those templates in the cone‐beam plane using the MI metric. The authors reported the 95th percentile of absolute errors to be <1.7 mm and <3.3 mm for phantom and patient studies, respectively. A similar method was used by Yang et al.²¹ to track tumors using the DRR from CT plans and achieved a tracking error of <1 mm for the majority of points on the phantom. In a retrospective study of four patients treated with stereotactic body radiation therapy, the differences in average tumor location between the conventional techniques and their methods were <2.2 mm. Dhont et al.²⁸ used a sequential acquisition of two orthogonal kV imagers for real‐time tumor tracking. DE fluoroscopy improved the contrast ratios for most projections, but full rotational tracking was not discussed. Patel et al.²⁹ used sequential acquisitions to track tumors using both a modified phantom and data for a single patient. They were able to track tumors on DE images with an accuracy of 1.4 ± 1.1 and 1.2 ± 0.6 mm in phantom and patient studies, respectively.

Although, previous studies showed improved MTT with DE imaging, these methods were not feasible for real‐time applications. The main disadvantage of these methods was sequential acquisitions that required additional postprocessing to find a matching low kVp image for each high kVp image. This procedure can be demanding in some applications and may take up to 6 s to match a single image.²⁸ Fast‐kV switching to perform DE imaging can address these issues. In our study, an improvement of subpixel tracking accuracy for the experimental setups was observed. Moreover, in this study, we used successive images to perform the weighted logarithmic subtraction. In theory, it is possible to interleave these images to double the effective frame rate.

There are several limitations of this study. X‐ray imaging for MTT results in additional dose to the patient. For the exposures considered, the DE imaging dose at the skin surface for the patient, assuming a source‐to‐skin distance (SSD) = 90 cm and 7.5 Hz frame rate, is approximately 1.022 mGy/s.³⁷ The impact of this imaging dose needs to be discussed with the radiation oncologist to determine if the gains associated with MTT outweigh the risk of additional dose. This study also used an idealized model of respiratory motion. However, tracking is performed on frame‐by‐frame basis by scanning the template across the image. Hence, tracking should not be sensitive to the underlying motion pattern. On the other hand, the relative speed of the tumor is a factor that could affect tracking results since fast moving objects may result in motion artifacts on the DE images. This interplay will need to be further investigated. Moreover, the optimal imaging parameters may depend on patient size. In particular, for the low kVp settings the mAs may need to be adjusted for larger patients. These changes may affect the MTT success rate, MAE values, and the gain of DE over SE imaging. It is also important to remember that DE subtraction introduces additional noise into the resultant images. In our prior study, a noise reduction algorithm was applied and shown to increase the SDNR by a factor of two.³⁰ The impact of noise reduction on template tracking is an area that requires further investigation. Another limitation is that all experiments were conducted on idealized spherical targets. Therefore, when the algorithm was able to detect the target, it was usually accurate. However, in patient cases the tumors are irregularly shaped with complex geometries which could reduce the accuracy of the tracking algorithm. In addition, changes in tumor shape can occur between the time of simulation and time of treatment. Future enhancements include the use of 3D printed tumors based on patient data to provide realistic simulation of clinical setup.^{38, 39}

5. Conclusion

In this study, MTT was investigated using fast‐kV switching DE fluoroscopy on a bench top system. Our results suggest that DE imaging can improve the performance of MTT. These improvements were largely observed in small targets (≤10 mm) with significant improved rate of successful template matches. Although for larger targets (>10 mm) DE imaging moderately impacted successful template matches 2D localization errors were much smaller than using SE imaging alone. The findings of our study suggest that DE imaging using fast‐kV switching on the OBI of a linac can enhance MTT in image‐guided radiotherapy applications.