Non-coplanar lung SABR treatments delivered with a gantry-mounted x-ray tube

Objective. To create two non-coplanar, stereotactic ablative radiotherapy (SABR) lung patient treatment plans compliant with the radiation therapy oncology group (RTOG) 0813 dosimetric criteria using a simple, isocentric, therapy with kilovoltage arcs (SITKA) system designed to provide low cost external radiotherapy treatments for low- and middle-income countries (LMICs). Approach. A treatment machine design has been proposed featuring a 320 kVp x-ray tube mounted on a gantry. A deep learning cone-beam CT (CBCT) to synthetic CT (sCT) method was employed to remove the additional cost of planning CTs. A novel inverse treatment planning approach using GPU backprojection was used to create a highly non-coplanar treatment plan with circular beam shapes generated by an iris collimator. Treatments were planned and simulated using the TOPAS Monte Carlo (MC) code for two lung patients. Dose distributions were compared to 6 MV volumetric modulated arc therapy (VMAT) planned in Eclipse on the same cases for a Truebeam linac as well as obeying the RTOG 0813 protocols for lung SABR treatments with a prescribed dose of 50 Gy. Main results. The low-cost SITKA treatments were compliant with all RTOG 0813 dosimetric criteria. SITKA treatments showed, on average, a 6.7 and 4.9 Gy reduction of the maximum dose in soft tissue organs at risk (OARs) as compared to VMAT, for the two patients respectively. This was accompanied by a small increase in the mean dose of 0.17 and 0.30 Gy in soft tissue OARs. Significance. The proposed SITKA system offers a maximally low-cost, effective alternative to conventional radiotherapy systems for lung cancer patients, particularly in low-income countries. The system’s non-coplanar, isocentric approach, coupled with the deep learning CBCT to sCT and GPU backprojection-based inverse treatment planning, offers lower maximum doses in OARs and comparable conformity to VMAT plans at a fraction of the cost of conventional radiotherapy.


Introduction
Radiotherapy (RT) research often focuses on state-of-the-art methods to provide small improvements in the treatment of cancer patients in high-income countries, while less focus is put on providing low-cost treatments suitable for the majority of people in the world who have little or no access to radiotherapy.According to a study conducted by Abdel-Wahab et al, optimal access to linear accelerators in low-and middle-income countries (LMICs) could save a million lives annually by 2035 (Abdel-Wahab et al 2021).Additionally, the conventional model of highly centralized radiotherapy networks, even in developed countries like Canada, Australia, and the UK, has reduced access to care and utilization rates due to the geographically dispersed patient populations, which makes the distance from a treatment center a crucial factor (Heathcote and Armstrong 2007, Baade et al 2011, Markossian et al 2014).Thus, there is a potential requirement for more economical external beam radiation therapy systems in cancer care worldwide.One solution to this issue could be the implementation of lower energy kilovoltage x-ray technology, which is relatively inexpensive and requires less shielding and infrastructure.
Grid therapy was one of the earliest attempts to use kilovoltage photons for non-superficial lesion treatment in the early 1900s (Laissue et al 2012).This method involved passing the primary photon beam through a metal 'grid' to create an array of parallel small photon beams.The purpose was to fractionate the radiation delivery spatially, thereby reducing skin complications.In recent years, various research groups have explored the potential of using kilovoltage photons to treat deep-seated lesions.For example, Loughery et al aimed to develop and assess the feasibility of a compact kilovoltage intensity modulated radiotherapy platform for contrastenhanced radiotherapy, which was found to be clinically feasible in both simulation and measurement (Loughery et al 2019).Rose et al used a modified CT scanner to treat brain metastases and reported a reduction in metastasis size while maintaining safe healthy brain doses (Rose et al 1999).Prionas et al demonstrated the feasibility of using a dedicated breast CT scanner for the treatment of breast lesions (Prionas et al 2012), while Abbas et al used a polycapillary optical cable to produce converging kilovoltage x-rays with a focal spot size of 0.2 mm, resulting in a broadening of the entrance dose and resulting in a skin-sparing effect (Abbas et al 2014).Likewise, Breitkreutz et al developed a Monte Carlo (MC) model of a simple isocentric treatment kilovoltage arc therapy (kVAT) system that utilizes a customized collimator to produce a linear array of converging beamlets for treating deep-seated lesions with minimal damage to skin and organs at risk (OARs) (Breitkreutz et al 2017(Breitkreutz et al , 2018)).
As we look at opportunities to introduce kV photon beams as a treatment modality for deep malignant lesions in geographic regions where radiotherapy is currently unavailable, one of the main challenges is that there is a lack of necessary tools and equipment for planning the treatment, such as computed tomography (CT) scanners and non-coplanar planning algorithms.The growing use of machine learning in combination with novel non-coplanar treatment planning strategies has the potential to address these issues.
Adaptive radiotherapy techniques often focus on using machine-learning augmented cone-beam CT (CBCT) to plan treatments (van de Schoot et al 2021).By using the same advanced techniques and treatment planning tools, adaptive radiotherapy workflows can be repurposed to plan treatments in places without access to planning CTs, which is one of the obstacles to the implementation of low-cost radiotherapy.Various methods have been utilized to increase the use of CBCT images in RT, including model-based (Zhu et al 2009, Wu et al 2014, Xu et al 2015) and deep learning-based techniques (Xu et al 2018, Zhang andYu 2018).While some methods aim to correct specific types of CBCT artifacts such as beam hardening or cupping, a recent approach using deep learning methods generates higher-quality synthetic CT (sCT) directly from CBCT images.Dahiya et al utilize a supervised image-to-image translation technique based on conditional generative adversarial networks (cGANs) to translate CBCT images to sCT images while also performing OAR segmentation driven by a novel physics-based artifact/noise-induction data augmentation pipeline (Dahiya et al 2021).The use of such algorithms has the potential to remove the necessity of a standalone planning CT from the radiotherapy workflow, lowering radiotherapy costs and thus be more accessible to LMICs.
Another large issue that arises when planning kV treatments is that the equispaced coplanar angles and trajectories used in clinical intensity modulated radiotherapy (IMRT) and volumetric modulated arc therapy (VMAT) plans, when used with kV beams, have exceptionally high skin dose.We address this issue by adopting more intelligent angle sampling schemes in non-coplanar directions.Some non-coplanar treatment planning algorithms have been developed for MV radiotherapy such as the station parameter optimized radiation therapy (SPORT) suite of algorithms.This includes segmentally boosted VMAT, dense angularly sampled and sparse intensity-modulated (DASSIM) radiation therapy, and an algorithm for fully automated Pareto optimal and clinically acceptable treatment planning (Li and Xing 2013, Xing and Li 2014, Smyth et al 2019).In the work of Huang et al, the non-coplanar Pareto optimal projection search (NC-POPS) algorithm is proposed for fully automated non-coplanar treatment planning to address these limitations (Huang et al 2021).However, to properly plan a kV isocentric treatment at depth, care must be taken to ensure adequate dose is delivered to the tumour as the source output is low compared to that of a linac and the dose calculation methods must be adapted to a kV beam which delivers dose more heterogeneously due to greater photoelectric interaction.Thus, a novel, dose-aware, non-coplanar beam optimizer is developed in this work specifically for kV planning.
While the use of modulated kV beams for external beams is not new, and much great work has been done in this area, there is currently no truly low-cost solution for external beam lung cancer treatment.Clinically admissible lung and breast dose distributions have been achieved using synchrotron sources (Buonanno et al 2019, Sarno et al 2020) as well as specialized converging treatment heads (Breitkreutz et al 2018(Breitkreutz et al , 2019)).However, synchrotron sources are relatively expensive, and the manufacturing of converging treatment heads proved infeasible.Although simpler approaches have shown good dose distributions for breast treatments using polyenergetic x-ray tubes, it has never been demonstrated that one can deliver SABR treatments for lung.As meeting the stringent RTOG 0813 requirements (Bezjak et al 2019) is challenging using a simple gantrymounted x-ray tube.Furthermore, a complete kV treatment workflow that eliminates the need for costly planning CT scans has not been shown in combination with these kV treatments.In this work, our aim is to demonstrate the proof of concept for a gantry-mounted kV treatment method using a polyenergetic x-ray tube for both planning and treatment delivery.
In our work, we explored novel design parameters for a treatment system using a kV x-ray tube and an articulating, 4-geometry capable, low-cost gantry.This system leverages the decreased cost of a kV x-ray tube compared to the cost of an MV linac Such a system could cost an-order-of-magnitude less than an MV system, with low operating costs, few specialised components, and low shielding requirements.The SITKA machine would avoid the estimated 4.1M price of a new linac and 1.2M additional cost of constructing the 2 m thick cement bunker needed for safe operation of a linac (Breitkreutz et al 2020).

System overview
The SITKA machine design simulated, shown in figure 1 and whose parameters are tabulated in table 1, incorporated a commercially available, 14 mA, 320 kVp x-ray tube and an articulating, 4-π geometry capable, low-cost gantry with a source-axis distance of 40 cm.Mounted on a C-arm opposite the source is a commercial CsI kV flat panel detector, which provided cone-beam CT (CBCT) imaging.A deep learning CBCT to sCT method is employed to provide planning sCT.Machine learning OAR autosegmentation was used in combination with planning target volume (PTV) contours drawn by remote radiation oncologists.A novel inverse treatment planning approach was employed, which used GPU backprojection to create a highly noncoplanar treatment plan with circular beams generated by a commercial iris collimator.All hardware used was from commercial sources although the exact models have been made confidential at the request of our industry  partner.MC methods were used to evaluate different aspects of this machine, e.g. to test the treatment efficacy of the device and workflow, and to check compliance of two lung cancer pateint plans against the radiation therapy oncology group (RTOG) 0813 protocols (Bezjak et al 2019) for lung SABR treatments with a prescribed dose of 50 Gy.

Patients
Two patients with lung cancer were treated with stereotactic ablative radiotherapy (SABR) using VMAT on a 6 MV Truebeam linear accelerator delivering 50 Gy over 5 fractions.The patient's planning CT images, treatment plans were anonymized and used to create SITKA plans.CT data was composed of 512 × 512 × 163 voxels with dimensions of 0.977 × 0.977 × 2.00 mm.Patient 1ʼs PTV was 49.5 cc located in the right lung while patient 2ʼs PTV was 27.7 cc PTV located in the left lung.
Patient selection for this study aimed to identify candidates suitable for treatment with the system under investigation.We specifically targeted two key patient parameters.First, we included cases with tumours of smaller sizes to assess the system's utility in reducing treatment times, which is of particular significance when utilizing cost-effective treatment devices.Secondly, patients with tumours located in close proximity to the ribs were considered to evaluate the system's impact on bone dose.This selection strategy forms an integral part of our methods, providing a clear rationale for patient inclusion and serving as the basis for the subsequent analysis and discussion.

Synthetic CT generation
The treatment planning optimization used a modification of Dahiya et alʼs sCT and auto-contouring generative machine learning models (Dahiya et al 2021).Due to memory limitations on most commercial GPUs, only 128 voxel cubes could be converted from CBCT to sCT.Therefore, a novel low-high-resolution stitching method was developed to convert the large patient volumes in this study.To achieve this, original dimension (highresolution) data, down-sampled (low-resolution) data, and a normalization method were used to ensure consistent model outputs and avoid stitching artifacts.
First, overlapping low-resolution images were stitched together, with images cropped to a volume of interest of 256 × 400 × 128.This resulting image was downsampled by half in the x and y directions to create an image of 128 × 200 × 128.A uniform kernel with strides of 10 voxels in the y direction was used to create images for stitching, and these images were saved and used as input for the model.The output of the model was normalized by ensuring that the overlap between consecutive windows had the same mean and standard deviation.The mean of all outputs for a given region of the images was taken to yield a low-resolution image.
Second, images were stitched together at the CBCT image resolution.Again, the volume of interest was cropped to 256 × 400 × 128, and a uniform kernel with strides of 16 in the x and y directions was used to create images for stitching.The images were stitched by matching the mean and standard deviation of each highresolution image to the same region of the low-resolution image.This resulted in a high-resolution 512 × 512 × 128 image.

Non-coplanar treatment planning
A novel high-speed GPU-based non-coplanar treatment planning algorithm, described in figure 2 was employed to generate optimized SITKA treatment plans.Firstly, the tumor volume was converted into a point cloud and then fitted with a Gaussian mixture model (GMM) to determine the isocenter and dimensions of the treatment beam circular collimation.In order to measure the radiation dose accurately in the tumor and OARs, the CT volume was adjusted from 100 kVp attenuation coefficients to 320 kVp attenuation coefficients using the Fastcat Python package (O 'Connell and Bazalova-Carter 2021).A fifth-order polynomial curve was then generated from the reconstructed image values, which allowed the original 100 kVp CT's attenuation values to be converted to the 320 kVp attenuation values relevant for treatment beam dose calculations.
Next, attenuation through the modified CT volume was ray-traced using the TIGRE Python package (Biguri et al 2016) at equally spaced angles around a sphere centered at the isocenter.Three thousand angles were calculated using the method described in appendix A, angles that entered on the boundary of the CT volume rejected.A 320 kVp x-ray beam was then ray-traced through in both the tumor and the OARs to determine the beam attenuation, using cylindrical collimation determined from the GMM dimensions.To obtain the dose in the tumor, the attenuation was divided by the relative electron density, which was acquired using the HU to relative electron density conversion tables of Bazalova et al (2008).
Angles were then selected based on a cost function that rewarded dose in the tumor and penalized dose in each OAR according to an individual weight.These beam angles were further weighted by the inverse of their dose to the tumor to produce uniform dose in the PTV from all angles.However, to spread out the skin dose, a minimum weighting of one third was given to the highest dose angles.The weighting was determined as follows where W i is the weighting for beam i, D i is the dose estimate for beam i and D max is the maximum dose estimate.
The selected beam angles were then used to backproject and reconstruct doses using the FDK algorithm (Feldkamp et al 1984), providing an initial estimate of radiation dose in the PTV.Subsequently, areas that received insufficient dose were identified in the reconstruction.To address these underdosed regions, additional GMMs were fitted, and collimation adjustments to provide additional dose to these regions were calculated based on a the shape of the boundary of each regions in each projection image.The additional projected images through each underdosed region were added to the original projected images with a preliminary weighting of W r

W
, 2 where μ PTV is the mean of the PTV and μ r is the mean of the region.This process was continued iteratively until the plan was deemed to be sufficiently conformal in the back projection by the treatment planner.On a Linux workstation equipped with an Nvidia RTX 2070 GPU, treatment plans were completed in 28 and 26 seconds, for the patient 1 and 2, respectively.Patients 1 and 2 required 6 and 5 iterations, respectively, to achieve sufficient PTV uniformity.

Monte Carlo dose calculations
In our study, we employed analytical methods for dose estimation, as previously described.However, for the final dose calculation, we opted for a Monte Carlo approach due to its well-established accuracy and precision in radiation therapy dosimetry.While the analytical methods were innovative, they lacked the established validation and robustness that Monte Carlo simulations offer in terms of calculating dose distributions.As a result, our choice to use Monte Carlo provided a more definitive and reliable metric, making SITKA dose distributions directly comparable to those calculated in Eclipse for the VMAT plans.
The dose calculations on the planning CT images were subsequently performed using the TOPAS MC (Perl et al 2012) code.A total of 2 × 10 9 total particles were simulated for each collimation iteration recommended by the backprojection-based dose calculation algorithm, using the angles, weightings, and collimation provided by the treatment planning algorithm.This amounted to 6 and 5 times 2 × 10 9 particles for the first and second patient, respectively.The kV Fastcat simulation utilized an analytical x-ray source with a focal spot diameter of 3 mm generated in Spekpy with an anode angle of 20 • and a tube voltage of 320 kVp.The source was subject to additional analytical filtration of 1.5 mm aluminum, 0.75 mm tin, and 0.25 mm copper.The simulations used the Geant4 Penelope physics list with a particle range cutoff of 0.5 mm and were run on a Linux desktop computer equipped with 8 Intel Skylake CPUs, no variance reduction techniques were used.Dose was scored in the CT volume and compared to the doses calculated in the Eclipse treatment planning system for the VMAT plans for each patient, respectively.Each patient was converted from HU values to materials using the method of Schneider et al (2000) using the default settings in Topas for this method.All materials were defined using cross section from the NIST database.

Collimation setting weighting
In order to ensure proper coverage of the PTV and conformality to the prescribed dose, as outlined in RTOG 0813 section 6.4 (as reproduced in table 1), a novel cost function was developed to weight MC for each collimation setting.To reduce noise in the MC dose distributions for each collimation setting, a Gaussian filter with a standard deviation of 0.6 was applied.The initial cost function was based on the coefficient of variation, which is a measure of the standard deviation relative to the mean of the PTV.

D
xD 3 where x i are the weights for each collimation setting, D PTV,i is the part of the dose distribution inside the PTV for each collimation setting, and D PTV is the weighted dose distribution.The parameters D PTV s and D PTV m represent the mean and standard deviation of the dose in the weighted PTV, respectively.This metric was selected because its minimization corresponds to achieving homogeneous coverage of the PTV.To minimize the cost, a limitedmemory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) algorithm was utilized in Python, with initial weights set at 1 and bounds between 0.5 and 10 inclusive.These initial weights were then employed as the basis for an additional cost function centered around the RTOG 0813 constraints.Specifically, according to RTOG 0813, 90% of the prescribed dose must be delivered to 99% of the PTV volume, while 95% of the prescribed dose must be administered to 100% of the PTV.To ensure that the isodose surface selected satisfied these criteria, the dose (D) was scaled such that neither of the criteria were violated: , 5 , 1 0.9 percentile , 1 5 where D s is the dose scaling factor, 50 Gy is the prescribed dose, and D¢ is the scaled dose.The next step involved maximizing dose conformity by minimizing the number of voxels above the prescription dose in the region 2 cm outside the PTV, which is referred to as the 'z ring'.This was achieved by using Powell's method (Powell 1977) to generate an isodose surface that conforms to the PTV and meets the prescribed dose criteria.The cost function for this optimization process was defined as follows: An additional term was incorporated to penalize the voxels outside the PTV that exceeded the 95th percentile of the PTV dose (D high ).This was deemed necessary to address cases where bony structures surrounding the PTV received high doses from the 320 kVp x-ray beam.

Treatment time calculation
In order to calculate the treatment time for the SITKA machine the x-ray tube model was simulated with the BEAMnrc package of EGSnrc (Rogers et al 2005).The x-ray anode, a tungsten block at an angle of 20 • , received incident fluence of 320 keV electrons.Below the anode was the beam filtration of 4 modules made of 0.8 mm beryllium for the exit window and 1.5, 0.75, and 0.25 mm of aluminum, tin, and copper filtration, respectively.The beam was then collimated to a diameter of 3 cm at isocenter through a 2 mm thick circular lead collimator and a phase space was captured at the SAD of 40 cm.The conversion factor between the number of original electrons to the number of photons in the phase space, combined with the absolute dose calculated in Topas for a given treatment was used to calculate the treatment time needed to deliver one 10 Gy fraction from the 14 mA, 320 kVp x-ray tube.

Synthetic CT generation
The sCT generation and auto-contouring (figure 3) simultaneously lessened artifacts, improved contrast, and corrected CT numbers of the CBCT data, providing a step toward CBCT-based treatment planning.Pronounced CBCT motion artifacts that can be seen in the heart and diaphragm of figure 3 Although the workflow in figure 2 shows planning based on sCT, CT images instead were chosen for the dose calculation in this work to enable direct comparison to the VMAT plan.This approach was taken as DVH calculations based on the sCT would still contain artifacts from motion and the imperfect registration process that would introduce uncertainty in the results.Moving forward with the prototype machine, planning with the sCT images will need to be evaluated directly using dosimetry in phantoms.

Non-coplanar treatment planning
The novel treatment planning algorithm executed quickly to produce highly non-coplanar treatment plans, which avoided OARs, maximized tumour dose and only took 28 and 26 s to complete on an RTX 2070 GPU, for the two plans respectively.Optimal angles were seen to be two opposing bands of helical angles in both lung patients (figure 4 which curve with the shape of the spine while avoiding the trachea, esophagus, and bronchial tree).

Treatment plan evaluation
SITKA treatment plans, shown in figure 5 showed conformal dose distributions and on average 6.7 and 4.9 Gy reduction of the maximum dose in soft OARs for patients 1 and 2 respectively.This was accompanied by a small increase in mean dose of 0.17 and 0.30 Gy in soft tissue OARs for patients 1 and 2, respectively.This is reflected in the SITKA plan DVHs (figure 6, table 2) which showed an increase in the lower dose volume in most OARs, but a reduction of the maximum dose to the aorta, spinal cord, and esophagus due to the use of noncoplanar beams.A larger increase to mean rib dose (4.2 and 2.1 Gy, respectively) as compared to soft-tissue dose was observed due to the high photoelectric absorption of the 320 kVp beam in bone.However, the absolute mean doses to ribs were still modest at 9.15 and 4.5 Gy for the SITKA plans.The SITKA dose falloff according to the 50% and 20% isodose lines (figure 5) were of similar conformality to the VMAT plan.Overall, the plan was deemed acceptable for clinical use by a radiation oncologist.
Lung SABR was chosen as the site for this study for a variety of reasons.Lung cancer is a common cancer, even in low-income countries.Additionally, it has the highest rate of cancer mortality of any cancer.Due to the low density of lung, it is also regularly in reach of the limited penetration depth of kV x-ray beams and the relatively few OARs in the region allow for non-coplanar treatments.While we hope to treat diverse sites in the future, initial treatments have focussed on lung cancer as an initial test case.Additionally, RTOG 0813 was chosen as a guideline as it was most consistent with the planning objectives for the lung SABR VMAT treatments discussed in the study.In different areas there exist other guidelines like the UK SABR Consortium Guidelines (UK SABR Consortium 20192019).However, while some specifics are different, most concepts and clinical objectives are similar.
Regarding RTOG 0813, both treatment plans met the dosimetric constraints for the PTV and the OARs without any violations (refer to table 2).The most demanding constraints in both plans were achieving the ratio of 100% and 50% of the prescription isodose volume to the PTV volume (R 100% and R 50% , respectively).This was challenging due to the Iris collimator's inability to provide arbitrary beam shapes, unlike the higher cost, multi-leaf collimator (MLC).With the limitation of using only cylindrical collimation produced by  shifting a circular beam shape, SITKA isodose surfaces were characterized as more spherical than the arbitrary shapes provided by the MLC equipped VMAT plan.Consequently, these more spherical SITKA isodose surfaces provided less conformal coverage of the PTV than the VMAT surfaces.This was particularly evident in patient 1, where more homogenous PTV DVHs tended to include the rib bones in the prescription isodose and violate the R 50 % constraint, as discussed in appendix B, forcing the PTV DVH to be inhomogenous as seen in figure 6.
To meet the R 100% and R 50% constraints in both plans, the allowances of 99% of the PTV volume receiving only 90% of the dose and 100% of the PTV receiving only 95% of the prescribed dose were used to compensate for areas of the PTV with irregular shapes that were challenging to cover with a spherical prescription isodose volume.It is important to note that the areas receiving less than 100% of the prescribed dose in the SITKA plan were strictly in the PTV margin and not in the gross tumor volume (GTV).
Integral doses were calculated for the PTV and the whole body contours using the method of Aoyama et al (2006) for the PTV and the whole body.VMAT integral doses for the patient 1 was seen to be 3.09 GyL for the PTV and 34.0 GyL for the body, while SITKA integral doses were 3.15 GyL for the PTV and 62.6 GyL for the body.For the patient 2 VMAT integral doses were 1.62 GyL for the PTV and 24.8 GyL for the body, while SITKA integral doses were 1.64 GyL for the PTV and 52.1 GyL for the body.

Limitations
A more sophisticated approach to treatment planning, involving optimized beam weighting by angle and beamletbased inverse dose calculations, has the potential to create SITKA radiation treatment plans of greater conformality and target homogeneity than those discussed herein.Due to the fact that constraints on maximum and mean doses were well below RTOG 0813 limits in SITKA OARs we can see some latitude for such an optimization strategy to improve tumor doses conformity while still being in compliance with RTOG dose limits.This study employed MC treatment planning, recognized as the standard of excellence for dose calculations, as a step in the optimization in the absence of established, open-source analytical treatment planning tools.However, a key drawback of MC treatment planning is its speed, as demonstrated by the fact that a single simulation in this case took 8 h to complete.Additionally, the backprojection-based algorithm employed in this study was inadequate in accounting for the inhomogeneities present in the lung, which is challenging to simulate through simple ray tracing techniques.Further work seeks to employ analytical inverse treatment planning, using beamlet methods or GPU MC techniques, to enable more sophisticated beam weighting.
The SITKA treatments for patients 1 and 2 were observed to have a estimated duration of 22 min and 15 s, and 27 min and 25 s, respectively, for each 10 Gy fraction based on the BEAMnrc models.While not as fast as conventional radiotherapy, this study aimed to show that adequate treatment plans could be achieved using a simple gantry mounted x-ray tube setup.It is important to note that the parameters employed in this work, such as beam weighting and beam filtration, were included some trade-offs that reduced the treatment dose rates.To increase the dose rates, one may replace the inverse dose weighting approach used in this study, which assigned lower weights to high dose beam angles, with a more sophisticated beamlet inverse planning technique that could perhaps yield both conformal PTVs and high dose rates.Longer treatment times inherently introduce a higher risk of patient motion and physiological variations, which can compromise the accuracy of radiation delivery.Therefore, the inclusion of cost-effective gating and motion tracking solutions is vital.Additionally, the inherent imaging capabilities of the machine would allow kV guidance in treatment delivery, which could be used to augment the systems motion management capabilities.
While the SITKA system does not have an MLC one can still create arbitrary dose distributions using the iris collimator.Examples of this optimization can be seen in the work of Jang et al (2016) in which iris collimated CyberKnife plans are compared to MLC VMAT plans with little difference in DVHs even in convex geometries.The downside of treating in this way is that with the introduction of small subfield the dose rate delivered to the tumour is reduced resulting in longer treatment times.Since the treatments on the SITKA machine are already time consuming this presents an obstacle in the optimization.
In the context of lung stereotactic ablative radiotherapy (SABR) treatments using a medical linear accelerator, the typical beam on time with a flattening filter is generally less than 4 min, while without the flattening filter it is approximately half that time for a 6 MV beam (Jaruthien et al 2021).These treatment durations are associated with maximum dose rates of 4.0Gy min −1 and 9.0 Gy min −1 , respectively, at a depth of 10 cm (Jaruthien et al 2021).In contrast, the SITKA system has a maximum dose rate of 0.6 Gy min −1 at 10 cm depth, indicating longer treatment times.However, when comparing treatment times, it is worth noting that the SABR treatments using a medical linear accelerator still demonstrate more favourable durations than the cyberknife system, which typically requires between 60 and 90 min to treat a lung SABR patient (Gibbs and Loo 2010).Given this comparison, it would be valuable to explore additional techniques for increasing the dose rate on the SITKA system.This could potentially involve incorporating additional sources mounted on the gantry or utilising higher powered x-ray tubes to enhance treatment efficiency.
However, a conventional fraction of 2 Gy could be delivered by the SITKA system in 5 min.So, while the stringent dosimetric criteria for lung SABR have been followed here, as the fewer patient visits and commonality of lung lesions in LMICs make it an impactful treatment in the LMIC setting, the machine could also be used for more conventional treatments.
Additionally, the beam utilized in this study was hard for a kV beam, having been filtered by 0.75 mm tin.Employing a beam with less filtration would increase the dose rate while likely still maintaining the skin dose within the prescribed RTOG 0813 limit, as neither patientʼs skin dose was near the limit.In addition, higher current x-ray tubes are available than the 14 mA model employed in this study, which could further reduce the treatment times.
The SITKA treatments incurred significantly more integral dose to the body than the VMAT treatments.SITKA treatments integral body doses were 62.6 and 52.1 GyL as compared to 34.0 and 24.8 GyL for the VMAT treatments.This shows the low dose bath created by the non-coplanar treatment plans.This presents a significant drawback of the SITKA method of treatment as this higher integral dose may have radiobiological consequences in terms of immune response and secondary malignancy which would have to monitored if SITKA plans were delivered clinically.
Overall, we present a straightforward radiotherapy system design that can efficiently deliver targeted doses to lung lesions while ensuring sufficient avoidance of OARs.The focus of our future work will be to build and characterize a prototype SITKA system for the delivery of veterinary treatments.This will include development of the necessary robotic and collimator controls to deliver the plans postulated herein and dosimetric evaluation of proposed plans prior to veterinary and patient treatments.

Conclusion
In conclusion, this study successfully demonstrated the feasibility and effectiveness of the SITKA system for delivering SABR lung treatments.The primary objective of this work was to provide a cost-effective alternative for radiotherapy, particularly targeted at LMICs.
The SITKA system, featuring a 320 kVp x-ray tube mounted on a gantry and incorporating innovative techniques such as deep learning CBCT to sCT conversion and GPU-based inverse treatment planning, proved to be compliant with all dosimetric criteria set forth by RTOG 0813.This compliance indicates SITKAʼs potential for safe and effective treatment delivery in line with established standards.
Notably, the SITKA treatments showcased a substantial reduction in the maximum dose to soft tissue OARs when compared to the conventional 6 MV VMAT approach, with an average reduction of 13.4% and 9.8% of the prescription dose for the two patients, respectively.This reduction in maximum dose comes with a minor increase in the mean dose to soft tissue OARs, emphasizing the system's ability to optimize treatment plans for improved OAR sparing while maintaining overall treatment conformity.
The significance of this work lies in its potential to address the challenges of providing high-quality radiotherapy in resource-constrained settings, particularly for lung cancer patients.The SITKA system's noncoplanar, isocentric approach, coupled with advanced image-guidance and treatment planning techniques, offers a promising solution to minimize radiation-related side effects and improve treatment outcomes.Furthermore, the cost-effectiveness of the SITKA system makes it a viable option for LMICs, where access to advanced radiotherapy is often limited.
In summary, this study demonstrates the potential of the SITKA system as a maximally low-cost yet effective alternative to conventional radiotherapy systems for lung cancer patients, with a particular focus on addressing the healthcare disparities in LMICs.The results presented here underscore the importance of innovation in radiation therapy technology to broaden access to high-quality cancer care, ultimately contributing to improved patient outcomes and global health equity.optimization process was that the beam's falloff was too sharp, resulting in certain areas of the tumor receiving doses below the prescription isodose line.The inclusion of a penumbra, which was not directly addressed in our simulations, would have helped in the optimization process by providing a smoother dose distribution.However, we used Gaussian filtration of the beam to remove noise, which had a similar smoothing effect as a penumbra.The omission of real collimation from the simulation was done to maintain computational efficiency.
To demonstrate the impact of the penumbra on arc kV beams, we modelled two arcs: one with lead collimation of 2 mm placed 10 cm away from the source, and another with the idealised collimation as used in our work.Both beams were collimated to 3 cm at the isocenter, and the dose for 5 × 10 7 photons was measured in a 10 cm diameter water cylinder for a single angle.The dose distribution was then radially convolved around the cylinder to create an arc. Figure D.1 presents the dose distributions for the two beams.It can be observed that while the realistic collimation results in a slightly smoother dose falloff, the deviation from the idealised dose distribution does not exceed 4%.

Figure 1 .
Figure 1.Schematic of the SITKA machine.A sagittal and axial view of a patient on the treatment couch is shown.

Figure 2 .
Figure 2. Overview of the treatment planning workflow starting with a CBCT acquisition and ending in beam delivery.
(a) were removed in the sCT allowing accurate dose calculation in these regions.Additionally, in figure 3(b) the autocontouring correctly identifies the OARs in the volume, while figure 3(c) shows the correction of the HU values from CBCT to sCT, a property essential for CBCT-based treatment planning.

Figure 3 .
Figure 3. (a) A planning CT is compared to a registered CBCT and the stitched, full-volume sCT.(b) The auto-contouring capability of the model is demonstrated.(c) The conversion from CBCT HU values to sCT HU values is seen.

Figure 4 .
Figure 4. (a) The raytraced dose in the tumour and dose in the OARs is shown sampled at 6000 equally spaced angles around a sphere, a hammer projection was used to map the sphere to a two dimensional plot.(b) An example of the cost function as a function of helical and non-coplanar angle, darkened areas denoting the angles selected.

Figure 5 .
Figure 5. Lung dose distributions for bose patients are shown for a lung SABR patient using either SITKA (a) or VMAT (b) treatments, the PTV is denoted as the red dotted line, 100%, 50%, and 20% prescription isodose lines are shown in red, green, and blue.

Figure 6 .
Figure 6.DVH curves computed for the SITKA and VMAT plans.are shown for both patient 1 (a) and patient 2 (b).