“Characterization of Cherenkov imaging parameters and positional constraints on an O-Ring Linear Accelerator.”

Abstract

Objective:

With the introduction of Cherenkov imaging technology on the Halcyon O-ring linear accelerator platform, we seek to demonstrate the imaging feasibility and optimize camera placement.

Approach:

Imaging parameters were probed by acquiring triggering data Cherenkov image frames for simplistic beams on the Halcyon and comparing the analyzed metrics with those from the TrueBeam platform. Camera position was analyzed by performing 3D rendering of patient treatment plans for various sites and iterating over camera positions to assess treatment area visibility.

Main results:

commercial Cherenkov imaging systems are compatible with the pulse timing of the Halcyon, and this platform design favorably impacts signal to noise in Cherenkov image frames. Additionally, ideal camera placement is treatment site dependent and is always within a biconical zone of visibility centered on the isocenter. Visibility data is provided for four treatment sites, with suggestions for camera placement based on room dimensions. Median visibility values were highest for right breast plans, with values of 80.33% and 68.49% for the front and rear views respectively. Head and neck plans presented with the lowest values at 26.44% and 38.18% respectively.

Significance:

This work presents the first formal camera positional analysis for Cherenkov imaging on any platform and serves as a template for performing similar work for other irradiation platforms. Additionally, this study confirms the Cherenkov imaging parameters do not need to be changed for optimal imaging on the Halcyon. Lastly, the presented methodology provides a framework which could be further expanded to other optical imaging systems which rely on line of sight visibility to the patient.

1. Introduction

Time-gated Cherenkov imaging, which has emerged in the last ten years as a remote imaging modality in radiation therapy for visualizing dose deposition, is a promising technique for performing machine quality assurance, treatment verification, and in vivo dosimetry^1–8. Cherenkov imaging leverages the quantitative relationship between energy deposition by high energy charged particles and the emission of Cherenkov radiation in dielectric media in order to infer dose distributions from optical images⁹. Most commonly, these images are captured by a remote intensified camera system and take advantage of the pulsed nature of radiation beams produced by linear accelerators to improve signal-to-noise by time-gating acquisition to the linac pulses. This setup provides significant flexibility when compared with radiation detectors with which the beam interacts directly, as the camera placement is only limited by the line of site visibility to the region of interest. However, since direct line-of-sight is required for any measurements, the camera position is also sensitive to occlusions from the gantry system, auxiliary treatment devices, and in the case of patient imaging, features of their body geometry outside the region of interest, necessitating interpretation of clinical images by a trained eye¹⁰.

The Halcyon system (Varian Medical Systems, Palo Alto CA) is recently developed ring-shaped, high-speed linear accelerator (linac) that is designed for high throughput image-guided radiation therapy (IGRT). In the context of Cherenkov imaging, the Halcyon system has some unique characteristics compared with conventional c-arm linacs which warrant consideration. As the first Cherenkov images of patient treatments on the Halcyon were recently acquired¹¹, this work seeks to provide insight into the technical feasibility of performing Cherenkov imaging on the Halcyon, with a focus on two topics: (1) timing and triggering characteristics for image acquisition, and (2) camera position optimization.

2. Materials and Methods

All Cherenkov imaging was performed with a C-Dose Research camera (DoseOptics, Lebanon NH), and both the Halcyon and the Truebeam linear accelerator platforms (Varian Medical Systems, Palo Alto CA) were used. The camera was outfitted with a Gen III image intensifier, which provides higher sensitivity with a red-weighted quantum efficiency curve but higher noise response when compare with other variants¹². An acrylonitrile butadiene styrene (ABS) reference board (DoseOptics LLC, Lebanon NH) measuring 40×40×2 cm was aligned at isocenter and used as a phantom for Cherenkov emission and to provide adequate scatter for triggering. The camera was mounted on a tripod at 1–1.5 m away from isocenter, and a 50 mm f/1.8 lens (Nikon, Tokyo Japan) focused and aligned to isocenter was used for all imaging.

2.1. Cherenkov Imaging Performance on Halcyon

In order to measure the camera triggering characteristics, a digital oscilloscope was used (PicoScope 3000 series, Picotech, Cambridgeshire, UK) to probe the output from the camera’s radio-optical trigger unit (RTU), the design considerations and characteristics of which have been explored in prior work¹³. The RTU consists of a high-speed scintillator coupled with a silicon photomultiplier module and converts measured linac pulses to a digital timing signal which is used to measure the camera, via detection of stray x-rays. The sensitivity of the RTU is such that it outputs saturated pulses at 5 V regardless of placement in the room for performance consistency. The camera was set up as described above, and a long coaxial cable was run out and connected to the oscilloscope in the console room along with the hybrid USB cable for data transfer and power to the camera. The oscilloscope was connected via USB to a desktop workstation running the PicoScope 6 software. Simultaneously, the camera USB cable was connected to the workstation and controlled by the accompanying software (C-Dose Research, DoseOptics LLC, Lebanon, NH). This setup is shown in Figure 1. For the triggering measurements, the Halcyon was operated in service mode using the 6MV flattening filter free (FFF) beam and the only available dose rate of 800 MU/min, with the field size set to the maximum of 28×28 cm. The PicoScope was operated in continuous mode and captured the input signal from the RTU until the memory buffer was filled. The data was exported to and analyzed in MATLAB (MathWorks, Natick MA), where the pulse widths and periods were extracted from the waveform.

Figure 1.

(A) Schematic of the Cherenkov camera, picoscope, and acquisition workstation. (B) Example breast plan rendering using an anthropomorphic female chest phantom, showing the body surface from the CT volume in gray with the surface dose sampled at 5 mm depth as a texture overlay. (C) View of 3D scene showing camera position relative to the treatment unit and the patient position; subfigures show the definition of the azimuthal angle (ϕ) and elevation angle (θ).

In addition to measuring the RTU output, various Cherenkov images were acquired with varying imaging settings to infer properties about the beam delivery timing and Characterize image quality. Cherenkov image frames were captured with fixed exposure times and contained multiple gated accumulations, each with a fixed duration and synchronized to the rising edge of each linac pulse via the RTU. The default exposure time for Cherenkov frames was 51 ms, and the default pulse duration was 4 μs. Various online image processing options could be applied for improving image quality, and by default included a 5×5-pixel spatial median filter, a 5-frame temporal median filter, and online background subtraction, which are all performed on an internal field programmable gate array (FPGA) in the camera. The camera captured time-delayed background frames in between each Cherenkov frame, with a pulse delay of 100 μs from the pulse rising edge and a default exposure time and pulse duration of 8 ms and 80 μs, respectively. The online background subtraction functioned by processing each pair of Cherenkov and background frames to produce a background-subtracted Cherenkov frame using the following equation:

$$C_{i,sub}=C_i-{\left(k_{bkg}^{Ch}\right)}_i\ast B_i$$

where $C_i$ and $B_i$ are the corresponding Cherenkov and background frames, and ${\left(k_{bkg}^{Ch}\right)}_i$ is the scaling factor which corrects for differences in exposure between the frames¹².

All Cherenkov acquisitions used the above online processing options, with the room lights turned off. Even under these conditions, there are some optical signals present in the room; for this reason, and because the background subtraction process does affect the noise characteristics of the images, background subtraction was still used to represent clinical imaging most closely. The three different delivery configurations were used for Cherenkov imaging are described in Table 1, and all beams were delivered to the reference board described above. In order to measure the width of the pulses indirectly, the Cherenkov pulse duration setting was varied to find the point at which the signal intensity stabilized. Similarly, the pulse delay setting was varied until no signal remained. Signal intensities were extracted from rectangular regions of interest (ROIs) both within and outside of the beam area for each acquisition.

Table 1.

Beam configurations used for triggering and imaging experiments.

Machine	Energy [MV]	Dose Rate [MU/min]	Monitor Units [MU]	Field Size [cm]	Gantry Angle [°]
Halcyon	6 MV FFF	800	100	10×10	0
TrueBeam	6 MV	600	100	10×10	0
TrueBeam	6 MV FFF	1400	100	10×10	0

2.2. Camera Positioning Optimization

In order to determine the optimal camera positions in the Halcyon bunker, a workflow was developed to render the camera view along with the linac and patient surface meshes in a 3D scene. The majority of this workflow was developed using pre-existing applet developed by the vendor which allowed for manual specification of the camera parameters with the same patient surface and superficial dose OpenCV-based rendering engine used in the C-Dose Research Software. This software converts a treatment plan in DICOM-RT format and converts the CT volume into a smooth patient surface defined at a threshold of −200 Hounsfield units (HU) and samples the co-registered dose volume in the first 5 mm along the direction normal to the surface. These dose values are then applied as a texture on the patient surface mesh, representing the visible treatment area, with a 10% threshold. Using the camera parameters input by the user, along with a 3D mesh of the Halcyon unit, the scene can be rendered in a way that simulates the imaging environment, as shown in Figure 1. The percentage of the treated area which is visible from any given camera view can be calculated, and this value was used as a quality metric for camera position, given by the following equation,

$$V_i=\frac{{\left(A_{vis}\right)}_i}{A}$$

Where $A$ is the total surface area on the CT surface with dose texture, and ${\left(A_{vis}\right)}_i$ is the portion of this surface area with is directly visible by the camera as any given position.

Since this metric is highly dependent not only on camera position but also on patient body shape and treatment site, multiple treatment plans for various sites were used, and the visibility metric was calculated on a square grid of camera positions with a spacing of 20 cm in each dimension. The extents along each axis were based on the dimensions of a Halcyon treatment bunker at our site: along the x (lateral) axis, the bunker measured 7.6 m wall-to-wall, with each wall equidistant to isocenter; the distance from isocenter to the back wall along the y (longitudinal) axis was 2.4 m, and the same measurement from the front wall to the isocenter was 4.6 m; lastly, the floor-to-isocenter and isocenter-to-ceiling measurements were 1.1 m and 1.6 m, respectively. To obtain a reasonable spread of treatment area variability, the treatment sites chosen were head and neck (HN), right breast, left breast, and prostate. Five randomly chosen treatment plans from our clinic for each of these sites were anonymized and exported from the treatment planning system (Aria, Varian Medical Systems, Palo Alto CA) as DICOM-RT files and imported to the applet for analysis. Once complete, the resultant visibility metric values at each camera position were exported as tabular text files and analyzed in MATLAB. Lastly, for each site, the visibility metrics at each camera location were averaged to produce a mean visibility at each position for that given site. The grid of camera positions used remained constant for all plans and treatment sites analyzed. In order to simplify the presentation of results, the 3D grid of visibility metric values was analyzed using two angles. These angles $\varphi$ and $\theta$ represent the azimuth and elevation are calculated as follows:

$$\varphi ={\tan }^{-1}\left(\frac{y}{x}\right)$$

$$\theta ={\tan }^{-1}\left(\frac{z}{\sqrt{x^2+y^2}}\right)$$

This reduced spherical representation was chosen as the visibility metric is independent of the radial coordinate$r$, given that the metric is measured as the visible percentage of the total treatment area, which is not a function of the projected size of the visible treatment area. The x, y, and z positions are defined in Figure 1 and are relative to isocenter.

3. Results and discussion

3.1. Imaging and Triggering Measurements

The triggering measurement using the oscilloscope captured 16 pulses over the course of around 90 ms, and this waveform can be found in Figure 2. The pulse frequency detected by the RTU was calculated by inverting the time between consecutive pulses (5.99 ms on average), measured at the rising edge at a threshold of 2V, yielding a result of 166.979 ± 0.001 Hz. At 800 MU/min, this corresponds to about 12.5 pulses/MU. Additionally, the width of each RTU pulse was measured as the difference between the rising and falling edges, again at a threshold of 2V, yielding a mean pulse width of 11.610 ± 0.004 μs.

Figure 2.

(A) Measured waveform containing pulses detected by the RTU at 166 Hz. (B) Zoomed-in view of one single pulse, showing a width of about 11 μs.

Since the width of the RTU pulse is not necessarily reflective of the true pulse width, as it is a saturated signal which includes the decay time from the scintillator element, the pulse width was inferred by varying the imaging settings. These results are shown in Figure 3, where the signal stabilizes at 3 μs for both the Halcyon and TrueBeam deliveries, as measured using variations in both pulse duration and delay. This indicates that the true pulse width for the Halcyon in the tested configuration is approximately 3μs, similar to the TrueBeam. Additionally, the number of pulses per frame was analyzed for the acquisitions, yielding 9.0 ± 0.1 pulses/frame for the Halcyon images, and 18.9 ± 0.5 pulses/frame for the TrueBeam images. These results indicate that the default imaging settings in the vendor software (51 ms exposure time and 4 μs pulse duration) are compatible with the Halcyon, as the pulse width was found to be very similar to that of the TrueBeam, the platform on which the settings were optimized. This also suggests there is no additional benefit to increasing the pulse duration setting on this platform relative to others. Additionally, the number of pulses captured per frame on the Halcyon compared with the TrueBeam is consistent with the ratio of the measured 167 Hz repetition frequency of the Halcyon pulses and the 360 Hz repetition frequency of the TrueBeam pulses, and this measured frequency agreed with the value published by the manufacturer^13–15.

Figure 3.

(A) Mean signal intensity inside of the beam in each Cherenkov image with varying pulse duration settings from 0.5 to 4 μs, showing a stabilization around 3 μs. (B) Mean signal intensity inside of the beam in each Cherenkov image with varying pulse delay settings from 2 to 3.5 μs, confirming a drop to zero at 3 μs.

When comparing the frames between the Halcyon and TrueBeam acquisitions, it was found that the Halcyon images had significantly higher signal to noise (SNR) per frame than the TrueBeam images (174.7 ± 89.1 and 11.9 ± 1.5 respectively), where signal to noise was calculated in each frame as the ratio of the mean intensity within an ROI inside the beam area to the standard deviation in an ROI outside the beam area. When inspecting the images, as shown in Figure 4, it was noted that the Halcyon acquisition showed far less impulse noise than the TrueBeam image, which is usually related to the level of scattered X-rays hitting the camera.

Figure 4.

(A) Single frame of default Halcyon acquisition (6MV FFF, 800 MU/min, 10×10 cm field). (B) Single frame of default TrueBeam acquisition (6MV, 600 MU/min, 10´10 cm field). Both images were acquired with 51 ms exposure time and 4 μs pulse duration.

To confirm that substantial differences in X-ray scatter at the camera was the cause of the SNR discrepancy between the Halcyon and TrueBeam images, a separate ratio SNR_B was calculated. This was defined as of signal within the beam area of the Cherenkov image to standard deviation within the same ROI in the background image, yielding 3.8 ± 1.2 and 2.6 ± 0.3 for the Halcyon and TrueBeam images, respectively. While the scaling of SNR_B was drastically changed relative to SNR due to the exposure discrepancy between the Cherenkov and background images, the SNR_B values between the Halcyon and TrueBeam are far more comparable to one another indicating that lower impulse noise due to reduced scattered X-rays in the Halcyon bunker reaching the camera is the cause of the increased SNR. It is worth noting that in general Cherenkov emission intensity per frame varies both with dose rate and with energy⁹; in this case, the lower mean energy of the 6MV FFF Halcyon beam would yield lower Cherenkov emission than the 6MV beam from the TrueBeam¹⁶, while the higher dose rate of the Halcyon delivery has an opposing effect. Furthermore, this effect in flattened beams is not uniform, as the beam hardening at the center of the field is increased compared to the periphery; these effects have been discussed in prior work, relating to the increased Cherenkov emission per unit dose from higher energy particles^9,17. In the portions of the beam which have increased hardening, namely the center of the beam, Cherenkov emission is increased. Therefore, only the relative variation in signal to noise metrics between the Halcyon and TrueBeam images are relevant, and absolute comparisons are not appropriate in this context. The lower levels of impulse noise in the Halcyon images correspond with lower documented head leakage from the Halcyon as a result of its dual layer MLC design and presence of a beam stopper^18,19, and may also result from dosimetric differences of a flattening-filter free beam, such as reduced penumbra. This benefits Cherenkov imaging on the Halcyon platform compared with other linac designs with increased single frame signal to noise, which is highly beneficial to image temporally varying treatment plans, especially those with small complex apertures such as VMAT. Future work will involve examining the per frame image quality of IMRT plans on the Halcyon platform relative to previous work^10,20.

3.2. Camera Position Analysis

Figure 5 shows visibility metric contour plots of each plan tested in terms of camera view angles. The positions with the best visibility metric for each plan, as well as the mean of all plans, are shown in Table 2, and correspond to the regions with the best visibility in Figure 5. Because of the bimodal nature of the visibility distribution, corresponding to the front and rear of the bore, both the front $(-180<\varphi <0)$ and rear $(0<\varphi <180)$ regions were considered separately.

Figure 5.

Contour plots for the visibility metric in the camera view angle parameter space for all plans tested, as well as the mean for each site.

Table 2.

Positions and view angles corresponding to the best visibility for each plan tested.

Site	Plan	Front $(-180<\varphi <0)$						Rear $(0<\varphi <180)$
		x (mm)	y (mm)	z (mm)	$\varphi$(°)	$\theta$(°)	Vis. (%)	x (mm)	y (mm)	z (mm)	$\varphi$(°)	$\theta$(°)	Vis. (%)
L Breast	A	1200	−3000	1400	−68.2	23.4	88.5	1600	2800	1200	60.3	20.4	60.8
	B	1400	−3000	1800	−65.0	28.5	59.1	1400	2800	1800	63.4	29.9	65.9
	C	1200	−3000	1800	−68.2	29.1	100.0	1800	2800	2000	57.3	31.0	99.2
	D	1000	−3000	2200	−71.6	34.8	56.9	2000	2800	1000	54.5	16.2	57.4
	E	1200	−3000	1800	−68.2	29.1	77.1	1400	2800	2000	63.4	32.6	75.3
	Mean	1200	−3000	1600	−68.2	26.3	74.3	1600	2800	1600	60.3	26.4	68.5
R Breast	A	−1000	−3000	1800	−108.4	29.6	88.4	−1800	2800	1800	122.7	28.4	73.2
	B	−600	−3000	2000	−101.3	33.2	80.3	−1800	2800	1400	122.7	22.8	68.5
	C	−600	−3000	2000	−101.3	33.2	63.5	−1800	2800	1800	122.7	28.4	59.8
	D	−800	−3000	2200	−104.9	35.3	65.0	−1800	2800	1200	122.7	19.8	56.7
	E	−1200	−3000	1800	−111.8	29.1	82.3	−1800	2800	2000	122.7	31.0	90.7
	Mean	−1000	−3000	1800	−108.4	29.6	74.5	−1800	2800	1400	122.7	22.8	67.7
HN	A	−200	−3000	2200	−93.8	36.2	24.5	−2400	2800	800	130.6	12.2	35.9
	B	−600	−3000	2000	−101.3	33.2	30.7	2600	2800	600	47.1	8.9	42.5
	C	0	−3000	2000	−90.0	33.7	26.4	−2400	2800	200	130.6	3.1	35.8
	D	0	−1800	1400	−90.0	37.9	25.9	−2200	2800	1200	128.2	18.6	38.2
	E	200	−3000	2000	−86.2	33.6	27.8	−2600	2800	600	132.9	8.9	54.2
	Mean	−400	−3000	2000	−97.6	33.5	26.0	−2400	2800	600	130.6	9.2	40.2
Prostate	A	800	−3000	1600	−75.1	27.3	53.9	1400	2800	1600	63.4	27.1	51.9
	B	600	−3000	600	−78.7	11.1	60.4	1400	2800	2000	63.4	32.6	55.0
	C	1000	−3000	1600	−71.6	26.8	55.6	−600	2800	2400	102.1	40.0	55.2
	D	800	−3000	1800	−75.1	30.1	56.1	−1000	2800	2200	109.7	36.5	61.1
	E	−800	−3000	1800	−104.9	30.1	54.7	1000	2800	2200	70.3	36.5	57.0
	Mean	1000	−3000	1600	−71.6	26.8	55.1	1200	2800	2000	66.8	33.3	54.7

Figure 6 illustrates the result of the visibility analysis for right breast plan E, for both the optimal front and rear positions. For each plan, the bore geometry constrains the camera positions with line of site to the treatment area to a biconical distribution on either side of the machine, which corresponds to the pair of semicircular contours shown in Figure 5. Lastly, Figure 7 displays the relative visibility between views from the front and rear of the Halcyon, indicating that of the four sites investigated, only HN plans had a substantial benefit for one side over the other, with the camera behind the machine yielding improved visibility on average. The median front-view visibility metric values for the four sites tested were 77.13%, 80.33%, 26.44%, and 55.64% for left breast, right breast, head and neck, and prostate, respectively. The corresponding values for the rear view were 65.93%, 68.49%, 38.18%, and 55.25%. These results also indicate that for breast treatments, on both the left and right side, camera placement on the front side of the machine is likely favorable.

Figure 6.

(A) 3D visibility metric distribution, showing the front side of the Halcyon bore (𝜙<0). (B) The rendering of the treatment plan (R Breast E) from the position in (A) with the highest visibility. (C) 3D visibility metric distribution, showing the rear side of the Halcyon bore (𝜙>0). (D) The rendering of the treatment plan (R Breast E) from the position in (C) with the highest visibility.

Figure 7.

Distribution of visibility metrics at optimal camera positions for each site, from both the front (−180<𝜙<0) and rear (0<𝜙<180) of the machine.

The camera positioning results portrayed in the contour plots in Figure 5 show strong laterality dependence for the breast treatments, which is expected given the laterality of the treatments themselves. However, with only 5 plans per site, the laterality of the optimal azimuthal view angle for HN and prostate treatments should not be generalized, as those sites are usually treated with arcs which spread the dose across the patient surface in a semi-unpredictable manner. It is also worth noting that an azimuthal angle of −90 or 90 degrees, or along the patient axis, is likely not the ideal placement for HN or prostate treatments due to the geometry of the neck, skull, and belly, and this is readily seen in the contour plots for the prostate plans in Figure 5, which have local minima at those angles. If two or more cameras were used, these results would imply that the imaging of treatment plans with no expected laterality bias in the surface dose distribution would benefit from complementary views at the same azimuthal offset from the patient axis (x = 0). Given the optimal view angles presented in Table 2, and room constraints based on a chosen fixed mounting method on the wall or ceiling, one can determine the ideal x, y, and z location of the camera relative to isocenter by inverting the equations above.

The substantially lower visibility metric values for HN plans comes from the concave surfaces where the superficial dose is deposited, between the chin, neck, and shoulders. This leads to more significant body occlusion when compared with prostate and breast treatments, and it is predicted that Cherenkov imaging of HN sites would benefit from multiple camera views. One other noteworthy limitation of the 3D rendering method presented here is that only body surface features which are present in the CT scan are included in the visibility analysis. Namely for breast treatments, the arms are typically raised above the head to remove them from the beam path, which may block the camera’s view of the treatment area; however, the CT scan does not extend far enough superiorly to include the arms and they are therefore not considered here. Additionally, treatment accessories and other items such as masks, bolus, and clothing may interfere with Cherenkov imaging even though they are not considered in this analysis¹⁰. Furthermore, the visibility metric as defined here does not account for differences in captured Cherenkov emission as a function of camera angle due to radiological and/or optical phenomena, such as the angular distribution of Cherenkov photons at the surface, nor does it account for the intensity of the Cherenkov images at proposed locations¹⁷.

Lastly, these camera positioning results may be extended beyond Cherenkov imaging to other optical imaging or surface tracking devices which rely on line-of-sight detection of the patient surface, and optimal ($\varphi$,$\theta$) combinations from this analysis can be used in combination with the distance constraints of an imaging device to determine the ideal mounting position in 3D space. This technique may also be extended to c-arm linac platforms, with the added complexity of gantry rotation introducing a temporally varying occlusive surface. While the applet used for the visibility analyses was developed within the vendor software platform and is therefore not openly available, an open-source MATLAB-based application for performing this workflow is being developed as an ongoing part of this project.

4. Conclusions

This work presents a characterization of Cherenkov imaging performance on the Halcyon, as well as the first comprehensive camera positioning study for Cherenkov imaging. The results presented are useful for future installations of Cherenkov camera in Halcyon bunkers, and also confirm the imaging system is well-suited to perform imaging on the Halcyon platform in its default state.