Abstract
Keywords
Introduction
Modern radiation therapy requires a high-quality CT simulation to generate a radiation treatment plan that can successfully treat diseased tissue while simultaneously sparing the healthy surrounding tissue. 1 However, the CT obtained at simulation only represents the patient's anatomy at the time of imaging and cannot account for anatomical changes or setup differences on the day of treatment.2–4 More importantly, it cannot be used to address whether a tumor is growing or shrinking throughout treatment. 4 Given that some treatments take longer than a month to complete, this may result in suboptimal treatment outcomes. 5
Adaptive radiotherapy (ART) was designed to account for changes between the time of simulation and treatment. ART covers a wide scope of clinical practices and can rely on different imaging such as MRI or CT to determine whether patients need to be adapted. ART is often separated into two main branches: online and offline. Online ART systems are normally automated commercially designed systems while offline ART currently relies on in-house clinical reassessment and manual replanning. 6 This manuscript is focused on testing a tool for offline ART using CT images. In ART, patients are re-simulated during the course of treatment to allow replanning according to changes in anatomy or other clinical factors.4,5,7 While this allows for improved dose delivery, patients must undergo additional imaging sessions. Alternatively, to make the ART process more efficient, methods have been developed to generate new planning quality CT images without requiring re-simulation, and some promising methods rely on improved CBCT.8–16 CBCTs are acquired daily for patient setup, but are normally of a lower quality, contain artifacts, and a more limited field of view.
In this project, a phantom and a sample of previously adapted patients were used to test if a CBCT correction algorithm can improve CBCT image quality enough for an accurate dose calculation. While various examinations using the algorithm exist in the literature,9–17 none examine how the algorithm performs in comparison and in conjunction with new HyperSight(Varian, Palo Alto, CA) equipped machines. HyperSight combines hardware and software upgrades to produce higher quality CBCTs.18,19 This project includes conventional CBCTs along with CBCTs taken with HyperSight to improve the assessment of how the algorithm performs as CBCT quality improves. Furthermore, only four papers tested the algorithm with photon therapy.9–12 All of these use a Collapsed Cone (CC) dose algorithm as opposed to Monte Carlo (MC). Of these, only two reported using any specific deformation techniques for performing deformations used by the algorithm.9,11 Both of these also performed extra deformations to compare final dose distribution. Only one of these two included patients that had same day CTs for comparing with the corrected CBCT. 11 Authors all noted awareness of these various limitations as well.9–12 None of those studies included patients with tumors in the extremities or reported issues related to target volumes extending to the edge of the CBCT's field of view (FOV). For these reasons, more testing on the algorithm was desired before making an assessment on a corrected CBCT's potential use for planning.
Methods and Materials
Data Curation
A CT simulation of a lung phantom, CIRS Thoracic Phantom (CIRS Inc, Norfolk, VA), was acquired followed by two CBCTs on two different clinical linacs, one TrueBeam and one Ethos (Varian, Palo Alto, CA). TrueBeam had a conventional CBCT imaging system with a small imaging panel and traditional image reconstruction software. The Ethos used in the study had been upgraded with HyperSight technology.
Two CT simulations were acquired according to a head and neck (H&N) and a chest protocol using a Catphan 604 (The Phantom Laboratory, Greenwich, NY). Corresponding CBCTs were acquired on TrueBeam and Ethos.
The study was conducted on anonymized retrospective data under two IRB-approved institutional retrospective review protocols. Seven previously adapted H&N patients and seven previously adapted sarcoma patients were randomly selected from a larger pool of patients.
All assessments were done on a single beam set for any given patient even if the whole plan consisted of multiple beam sets. Dose grid voxels sizes were 0.25×0.25×0.25cm3 or smaller. H&N plans had primary prescription doses ranging from 16.8 Gy to 69.6 Gy to 95% of high dose PTV. Primary tumor locations ranged from frontal sinuses to larynx and elective regions included CTVs extending to supraclavicular lymph nodes. All H&N plans were treated on TrueBeam linacs. Time between initial CT simulation (initial CT) and adaptive CT re-simulation (adaptive CT) ranged between 27 to 44 days with a median of 32 days. Four extremity sarcoma plans contained 70Gy, 60Gy, and 50Gy PTVs based on a clinical trial. One thoracic sarcoma plan contained 60Gy and 46Gy PTVs. The final two abdominal sarcoma plans contained 70Gy and 50Gy PTVs. All sarcoma plans were treated on an Ethos linac. Time between initial CT and adaptive CT ranged between 20 to 36 days with a median of 22 days.
Anonymized files stored in RayStation 2023B (RaySearch Laboratories, Stockholm, Sweden) contained each patient's initial CT, adaptive CT, the corresponding adaptive plan, and a CBCT from the same date as the adaptive CT. No identifiable patient information was stored in the files.
All three images of the lung phantom were placed into a test patient in RayStation. A treatment plan was generated to deliver 60Gy to GTV (gross tumor volume) based on a lung SBRT protocol on the full CT scan. Then, external (Ext) and field of view minus one centimeter (FOV-1 cm) contours were generated for the CBCTs. The CT was then deformably registered to each CBCT scan using the ANACONDA algorithm 20 v3.2 in RayStation.
After the previous steps, the CBCTs were processed using the RayStation CBCT correction algorithm. This step requires selecting the initial CT that the scan was registered to along with the CBCT that needs to be corrected. This step also allows for the selection of the desired FOV. All corrections in this paper were made using the CBCT's FOV-1 cm contour. Once this was performed, the correction algorithm corrects the CBCT by running an iterative optimization that raises the average HU of the CBCT to that of the reference CT and identifies and eliminates artifacts. A corrected CBCT was generated for both TrueBeam and Ethos CBCTs of the lung phantom. Then, the plan initially made on the CT scan was recalculated on both of the corrected CBCTs.
The same procedure was followed using the Catphan CTs and CBCTs without generating or calculating a plan. The H&N protocol CT was used with the TrueBeam CBCT. The chest protocol was used with the Ethos CBCT. The Catphan CBCTs and their corrections were processed according to a monthly image testing protocol and basic image quality metrics were extracted.
A similar procedure was followed for the 14 previously adapted patients. CBCTs, taken the same date as an adaptive CT, were corrected using the patient's initial CT. Then the adaptive CT was registered to the corrected CBCT. The corresponding adaptive plan was then recalculated on the corrected CBCT in Ray Station with Monte Carlo v3.0 algorithm. Figure 1 depicts the workflow.
This details the general process used for data creation. The first three steps are all that is required for generating a corrected CBCT for planning.
All recalculated doses were resampled to match the original dose file using Mirada RTx (Mirada Medical, Oxford, UK) for proper alignment and comparison of the doses. For the phantom, both corrected CBCT plans were compared to the initial plan using a 3%/2 mm gamma analysis. The phantom study independently assessed the algorithm's ability to correct CBCTs without variation due to the anatomical changes that could occur in a patient over time. For patients, the adaptive plans were compared to the recalculated adaptive plan using the 3%/2 mm gamma analysis.
Upon obtaining the initial results, it was determined that the process would result in multiple patients failing gamma analysis. In the failing cases, the deformable registrations used in creating the corrected CBCTs were redone using the FOV-1 cm as a Focus ROI. Rigid registrations between the corrected CBCTs and the adaptive re-simulations were also redone prioritizing the alignment of target volumes. The gamma analysis for each of these was then repeated.
Statistical Analysis
Catphan CBCTs were loaded into SunCheck (Sun Nuclear Corp., Melbourne, FL) for image analysis. Spatial resolution, contrast, and noise were measured before and after correction. Spatial resolution was determined by detecting line pairs per millimeter. Contrast was the difference between the HU of Teflon and background material normalized by the HU of Teflon. Noise was the standard deviation of HU in a region of water-equivalent material.
All 3%/2 mm gamma analyses were performed in the external minus 7 mm region as well as in the GTV region. This allowed for evaluation of the overall plan and verification of differences inside the primary target. Threshold of the analyses was set to 10% of max dose. All analyses were performed in 3DVH (Sun Nuclear Corp., Melbourne, FL).
Mann-Whitney U-tests were performed in Microsoft Excel (Microsoft, Redmond, WA) comparing the percentage of passing voxels in the gamma analyses between patient CBCTs from TrueBeam and Ethos.
Results
Table 1 shows results for the image analysis performed on the Catphan before and after correction. The most significant difference between TrueBeam and Ethos was noise. Ethos also had lower spatial resolution.
CBCT Correction Imaging Data.
Figure 2 contains images of the CBCTs analyzed in Table 2. Images are all set to RayStations liver window level. Decreased noise is visible between TrueBeam and Ethos images.
Catphan CBCTs. (A) TrueBeam. (B) Corrected TrueBeam. (C) Ethos. (D) Corrected Ethos.
Gamma Analyses Results (3%/2 mm) for the Lung Phantom.
Table 2 shows the results from the gamma analyses for the lung phantom. Both corrected CBCTs had 100% passing rates inside the GTV and the External-7 mm contour. The plan was also calculated on the raw CBCT using the same CT to mass density curve of the initial CT's scanner.
Table 3 shows the results for all of the H&N patients’ adaptive plans recalculated on their respective corrected CBCTs. Six out of seven patients had over a 90% passing rate inside the External-7 mm contour. All seven patients had over a 95% passing rate inside of the GTV.
Gamma Analyses Results (3%/2 mm) for H&N Patients.
Table 4 shows the results for all of the Sarcoma patients’ adaptive plans recalculated on their respective corrected CBCTs. Five out of seven had over a 90% passing rate in both the External-7 mm and the GTV contour.
Gamma Analyses Results (3%/2 mm) for Sarcoma Patients.
From Figure 3, the raw TrueBeam CBCT had the worst image quality, resulted in target dose errors exceeding 5% and minimum errors greater than 1%. After correction, the maximum target dose error decreased to 4% with the majority of voxels having less than 1% error. The raw Ethos CBCT had better performance with the majority of errors between 1% and 2% within the target. The corrected Ethos CBCT had excellent performance with the vast majority of the voxels having less than 1% error and a maximum of 2% error for a few voxels in the periphery of the target. For all CBCTs, errors more than 2 cm outside the target were generally less than 1%. Visually, the corrected CBCTs (Figures 3C and E) more closely resemble the base plan (Figure 3A) than their counterparts (Figures 3B and D).
Comparison of dose on phantom dose calculations. The red contour indicates the GTV. Color washes represent 110% (light blue), 105% (pink), 100% (red), 95% (dark blue), 50% (green) and 25% (cyan) of the prescription dose. (A) Standard CT. (B) TrueBeam CBCT. (C) Corrected TrueBeam CBCT. (D) Ethos CBCT. (E) Corrected Ethos CBCT.
H&N_1, H&N_4, Sarcoma_1, and Sarcoma_3 had less than 95% passing rates in the External-7 mm. Figures 4 through 11 shows where the failing voxels are located for each patient.
Images of H&N_1. (A) Initial CT with external contour (green) and CBCT FOV-1 cm (red). (B) Raw CBCT. (C) Corrected CBCT. (D) Adaptive re-simulation with 16.8 Gy adaptive boost. Cyan contour is primary PTV. Red contour is primary GTV. Blue contour is low dose PTV. (E) Corrected CBCT with dose. F) 2%/4%/5% dose difference.
Dose differences in H&N_1 increased near the edges of the CBCT's FOV-1 cm. This is likely due to a combination of the poor CBCT reconstruction at the edges and suboptimal deformation of the initial CT in this region due to the sudden start and stop of the CBCT acquisition. For H&N patients, significant amounts of tissue were excluded posteriorly, as seen in Figure 4C. The combination of the missing tissue and noise in the CBCT resulted in the lower lymph nodes receiving excess dose compared to the adaptive plan.
Figure 5B confirmed the majority of failing points in the gamma analysis occur further from the GTV where the corrected CBCT has worse image quality due to the limited FOV of the CBCT. Small deviations in positioning of structures such as tongue, chin, and tilt of neck may have also contributed to the dose differences.
3DVH 3%/2 mm gamma analysis results for H&N_1. Red contour is GTV. Cyan contour is primary PTV. Dark blue contour is low dose PTV. Brown contour is External-7 mm within which the gamma analysis was performed. (A) Sagittal view. (B) Axial view. (C) Histogram of passing and failing voxels. (D) Coronal view.
While this patient had a 92% gamma pass rate, we again see the CBCT correction algorithm carved tissue out of the back of the corrected CBCT in Figures 6C and 6D. This results in hotter dose to the inferior lymph nodes.
Images of H&N_4. (A) Initial CT. Green contour is external. Brown contour is CBCT FOV-1 cm. (B) CBCT. (C) Corrected CBCT. (D) Adaptive re-simulation with dose. 68 Gy adaptive plan. Cyan contour is primary PTV. Red contour is primary GTV. (E) Corrected CBCT with dose. F) 5%/4%/2% dose difference.
The gamma analysis also revealed that failing voxels occur primarily at the edge of the FOV of the CBCT where the corrected CBCT has high noise and missing tissue due to the deformation of the CT outside of the CBCT FOV. This was most apparent in Figure 7D.
3DVH 3%/2 mm gamma analysis results for H&N_4. Red contour is GTV. Cyan contour is primary PTV. Purple contour is low dose PTV. Yellow contour is External-7 mm within which the gamma analysis was performed. (A) Sagittal view. (B) Axial view. (C) Histogram of passing and failing voxels. (D) Coronal view.
Several problems were seen after inspection. Towards the superior region of the arm the GTV crosses the CBCT FOV and thus the region has a discontinuity in the corrected CBCT as seen in Figure 8C. This likely is responsible for the dose difference in the superior portion of the plan in Figure 8F. When comparing Figures 8D and 8E, it should also be noted the right side of the images where the tumor is located do not align at the surface. Approximately, 1 cm of tissue is missing. After inspecting the CBCT compared to the corrected CBCT. This did not appear to be a reconstruction error, but an actual difference in how the tissue rested during the treatment. This suggests that the dose error in the low dose inferior regions of the plan was due to actual positioning differences between the CBCT and adaptive resimulation.
Images of Sarcoma_1. (A) Initial CT. Green contour is external. Brown contour is CBCT FOV-1 cm. (B) CBCT. (C) Corrected CBCT. (D) Adaptive re-simulation with dose. 70 Gy adaptive plan. Only one of two beam sets was used. Dark red contour is low dose PTV. Light green contour is primary GTV. Red dose cloud is 70 Gy. (E) Corrected CBCT with dose. (F) 5%/4%/2% dose difference.
Failing points in the gamma analysis for Sarcoma_1 largely agreed with dose differences present in RayStation. However, the median dose difference given in Figure 9C is less than 10cGy. Combined with the fact that 77% of voxels were still passing the overall gamma analysis, more voxels had close agreement than not.
3 DVH 3%/2 mm gamma analysis results for Sarcoma_1. Light green contour is GTV. Dark red contour is low dose PTV. Red contour is External-7 mm within which the gamma analysis was performed. (A) Sagittal view. (B) Axial view. (C) Histogram of passing and failing voxels. (D) Coronal view.
Primary issues with Sarcoma_3 were similar to that of Sarcoma_1. Positioning of the arm had minor but notable inconsistencies. Figures 10D and 10E represented the best registration that was achievable. Changes in tilt of shoulder and elbow likely contributed to the dose differences. A unique issue appeared to be large low-density artifacts surrounding the humerus. This was present in the CBCT itself and likely means there is a limit to how low the image quality can be for the correction algorithm to still be effective. This may have contributed to the dose differences above the tumor along with the differing elbow angle. A final issue of note is the external contour in the adaptive CT included large extra external volumes of air. This likely has a low impact on the difference as the External-7 mm contour trims most of this away, and the extra regions only contain air and do not significantly attenuate the beam.
Images of Sarcoma_3. (A) Initial CT. Green contour is external. Cyan contour is CBCT FOV-1 cm. (B) CBCT. (C) Corrected CBCT. (D) Adaptive re-simulation with dose. 70 Gy adaptive plan. Dark red contour is low dose PTV. Light green contour is primary GTV. Dark green contour is high dose PTV. (E) Corrected CBCT with dose. (F) 5%/4%/2% dose difference.
The central hot differences seen in Figure 11B correspond to the low-density artifacts surrounding the humerus. The cold failing regions seen inferior to the tumor likely correlate to inconsistent shoulder placement making tissue rest somewhat differently on the treatment table.
3 DVH 3%/2 mm gamma analysis results for Sarcoma_3. Light green contour is GTV. Dark red contour is low dose PTV. Dark green contour is high dose PTV. Brown contour is External-7 mm within which the gamma analysis was performed. (A) Sagittal view. (B) Axial view. (C) Histogram of passing and failing voxels. (D) Coronal view.
In an attempt to correct the issues found the failing patients, two methods were simultaneously employed. First, the initial deformable registration used in creating the corrected CBCT was made again using the FOV-1 cm as a Focus ROI in RayStation. This significantly decreased the amount of missing tissue at the CBCT FOV-1 cm to initial CT boundary. Second, rigid registration between the corrected CBCT and adaptive re-simulation was redone prioritizing the alignment of GTVs between the scans. The analyses were then redone and Table 5 shows the results.
Improved Gamma Analyses Results (3%/2 mm) for the Failing Patients.
The improved methods successfully fixed all failing patients. Although using the FOV-1 cm as a Focus ROI was not mandatory in RayStation, it was required to prevent poor corrected CBCT generation near the edge of the CBCT FOV.
Table 6 shows the final median passing rates of patient tests and p-values for a Mann-Whitney U-test performed between them. The tests showed that corrected TrueBeam images outperformed Ethos at a statistically significant level, but all median pass rates were above 95%.
Comparisons of Final Patient Data with Mann-Whitney U Test.
Discussion
Phantom analyses revealed the correction algorithm can perform accurate CBCT corrections that reduce the average dose differences between a CT and a corrected CBCT to less than 1% for the vast majority of voxels. Both TrueBeam and Ethos CBCTs were significantly improved by the correction algorithm as seen in Figure 3. The uncorrected Ethos CBCT and the Ethos corrected CBCT outperformed the TrueBeam corrected CBCT which suggests that initial CBCT image quality impacts how well the correction algorithm can perform. Further analysis using the Catphan suggests that noise was the primary factor causing dose error in the CBCTs.
As seen in Figures 4C, 6C, 8C, and 10C, the correction algorithm generated noisy regions towards the end of the CBCT's field of view and incorrectly removed tissue right outside of the CBCT's FOV. This was due to the image beginning to transition to a deformed version of the initial planning CT outside of the CBCT's field of view and the edges of the CBCT having more noise to begin with. Thus, closer to the target there was less than 2% error per voxel while further away notable increases in error were observed. A critical factor needed in this case would have been to increase the field of view of the CBCT to include all of the targets plus a small margin to ensure any erroneous deformations and noisy regions don’t occur in the path of a treatment beam. Alternatively, if the anatomy is suspected to have not changed in those regions, the deformation of the initial CT should use the FOV-1 cm as a Focus ROI. This will significantly decrease deformations occurring outside the FOV-1 cm and stitch in the initial CT in a more rigid manner. Table 4 demonstrates this significantly improved the results in the failing H&N cases and helped the edges of the failing Sarcoma cases. Careful inspection should still follow the generation of the corrected CBCT to prevent the chance that this method fails unexpectedly. Some studies reported using the Focus ROI with the FOV contour resulting in great success.9,11 Others did not, but still had high levels of success.10,12 Our results indicated that the Focus ROI feature should be used to prevent the generation of poor corrected CBCTs.
Another factor to consider is how minor variations in positioning like tilt of the chin or shifts in the shoulder impact the change in dose distribution. This could explain deviations towards the surface of the patient as well. If it does, it would suggest that the difference in dose calculations is actually due to how patient position changed between the CT simulator and the linac treatment table. In this case, it may be preferable to use the corrected CBCT as it better represents the position the patient has on the treatment table. Refocusing the registration of the failing Sarcoma cases to align GTVs as opposed to boney anatomy significantly improved the dose distribution as seen in Table 4. However, achieving this registration required ignoring the rest of the anatomy. This further indicates that there was notable change in patient setup throughout the course of treatment.
Overall, 10/14 patients had over a 95% passing rate in the External-7 mm contour. After adjustment of registration errors, all patients had over a 95% passing rate, which would pass clinical QA and be deliverable. However, there were issues that should be addressed before implementing the algorithm on a patient. The three primary issues were the CBCT FOV not being large enough, the positioning of the patient not allowing for an ideal rigid registration of the patient, and large artifacts existing in the CBCT.
Some dose differences can be attributed to internal shifts in the body such as gas in the intestines, as found in sarcoma cases. This error is unavoidable and occurs on a daily basis. However, by using the corrected CBCT, the images are taken closer to the actual treatment as in the case of online adaptive treatments. This could result in the corrected CBCT being the preferred image over a re-simulation CT. Similarly, since patients are in the exact position as in the treatment, it minimizes the potential concerns about the reproducibility of the position.
Similar evaluations were performed previously.9–17 These studies largely used the correction algorithm as a way to evaluate if the patient anatomy had changed sufficiently to trigger an adaptation. They relied on older versions of the algorithm and frequently included deformed doses that were calculated with CC algorithms as opposed to MC. In addition to this, some studies exclusively analyzed proton therapy dose and did not look at photon therapy.13,15–17
Kurz et al tested a similar algorithm that did not possess the ability to correct artifacts. Their results indicated that this algorithm was clinically sufficient for replanning an intensity modulated radiation therapy (IMRT) H&N plan. This test relied on 3%/3 mm gamma analysis instead of the stricter 3%/2 mm standard used in this paper. Also, the analysis was only for IMRT and not VMAT and was calculated with a CC algorithm. 14 The combination of these factors along with the lack of artifact correction indicated a need for further testing.
In the following years, several studies were performed on the initial versions of the algorithm for photon-based dose calculations.10–12 These covered a wide variety of tumor sites including lung, rectal, breast, and H&N cancers. All of the studies concluded that the algorithm had clinically acceptable performance at the time. These studies performed gamma analyses based on deforming all the doses to a common image. This was frequently necessary due to lack of adaptive CTs being taken the same day as a CBCT.
Lechner et al tested a more recent version of the algorithm on H&N, gynecological, and lung cancer patients. The study found high levels of agreement with gamma analysis even at a 1%/1 mm passing criteria. However, a major limitation of the study was that the corrected CBCT had no same day CTs for comparison. Instead, planning CT dose and corrected CBCT dose were both deformed to a version of the CT deformed to the CBCT. The multiple layers of deformation place limitations on how confidently one can actually compare the doses. On top of this, the doses were computed using a CC algorithm as opposed to MC. 9 These limitations made it desirable to reassess the most recent version of the algorithm by comparing the dose computed on a corrected CBCT to that of a same day adaptive CT with a MC dose algorithm.
Previous publications have also tested other techniques for CBCT correction and synthetic generation.8,11,12,14,21 These range from basic deformations of the CT to more complex deep learning-based solutions. Many of these papers report similar success rates in using other techniques. However, all of these studies fall into the same set of limitations previously addressed in the introduction. The CBCT correction algorithms biggest advantage over the other algorithms is that its design preserves the anatomy contained in the CBCT. Methods that rely on deformed images and deep learning-based techniques are likely more susceptible to generating large anatomical changes that would not be easy to verify. In addition, the CBCT correction algorithm makes it possible to accurately use the mass density curve of the CT scanner used in the correction. This makes it so that a high-quality mass density curve does not need to be maintained for the CBCT imager itself.
Limitations of this study include small patient samples and the use of a CT to mass density curve for calculation of dose on the uncorrected phantom CBCTs. Large samples are difficult to obtain for testing algorithms such as this because they require a large number of adaptive patients and a large amount of time to process each patient. Given that the algorithm was able to succeed for all patients, it is suspected that failures are unlikely and should be caught when inspecting a corrected CBCT upon generation. At the time of study, there was also no machine specific CT to mass density curve generated for the two linacs. This may have given the phantom corrected CBCTs an advantage over the corresponding CBCTs. However, the overall success of the algorithm on the patient cases is not diminished by that fact.
This study confirmed the accuracy of the correction algorithm using a lung phantom on various machines. It also demonstrated that the higher quality the initial CBCT is the better the quality of the resultant corrected CBCT will be. Mann-Whitney U-Tests demonstrated the opposite in patient cases. However, this is likely due to underlying differences between tumor sizes and location as all H&N cases were on TrueBeam and all Sarcomas were on Ethos. Regardless, all cases ultimately reached an over 95% passing rate in the gamma analyses. Reasons for gamma analysis failures were identified so that if the algorithm was implemented clinically, the most recent CBCT could be reviewed to determine its suitability for dose calculation. If the image is insufficient due to CBCT FOV, a focused deformation can be used to improve the quality of the corrected CBCT which resulted in 100% of patients in this study passing. This would allow many patients to avoid additional CT simulations, as plans could instead be created using corrected CBCTs. This approach reduces the patient burden and streamlines clinic workflow.
Conclusions
Given the success rate of the correction algorithm and predictability of failures, the algorithm offers an opportunity to not only eliminate some adaptive re-simulations but even improve treatment quality by enabling better dose monitoring between fractions. The algorithm strictly improves accuracy of dose calculations on both traditional TrueBeam CBCTs and higher quality HyperSight equipped Ethos CBCTs as demonstrated in the phantom analysis. While the corrected TrueBeam CBCTs still contain some errors which need further improvement, they are still of sufficient quality for assessing larger deviations in dose delivery that would trigger adaptation. HyperSight equipped Ethos corrected CBCTs produce nearly identical dose distributions to that of a planning CT in phantom testing. At that level of performance, the corrected Ethos CBCT could be used for full adaptive offline replanning. As more TrueBeams are upgraded with HyperSight, further work must be done to evaluate the algorithm on this new technology.
Footnotes
Ethics
This retrospective study is approved by the Institutional Review Board (IRB) of the H. Lee Moffitt Cancer Center and Research Institute in Tampa, FL under IRB no. Pro00035642 for the sarcoma cases and IRB no. Pro00009323 for the H&N cases. No further approval or consent was required since the study uses retrospective anonymized data.
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Disclosure
Preliminary results of the project were presented as a poster at ASTRO 2025. These results were only preliminary, and the full project goes beyond what was presented there.
Data Availability
Patient data cannot be shared at this time. All other data is included in the paper.