A clinical 3D/4D CBCT‐based treatment dose monitoring system

Abstract To monitor delivered dose and trigger plan adaptation when deviation becomes unacceptable, a clinical treatment dose (Tx‐Dose) reconstruction system based on three‐dimensional (3D)/four‐dimensional (4D)‐cone beam computed tomograpy (CBCT) images was developed and evaluated on various treatment sites, particularly for lung cancer patient treated by stereotactic body radiation therapy (SBRT). This system integrates with our treatment planning system (TPS), Linacs recording and verification system (R&V), and CBCT imaging system, consisting of three modules: Treatment Schedule Monitoring module (TSM), pseudo‐CT Generating module (PCG), and Treatment Dose Reconstruction/evaluation module (TDR). TSM watches the treatment progress in the R&V system and triggers the PCG module when new CBCT is available. PCG retrieves the CBCTs and performs planning CT to CBCT deformable registration (DIR) to generate pseudo‐CT. The 4D‐CBCT images are taken for target localization and correction in lung cancer patient before treatment. To take full advantage of the valuable information carried by 4D‐CBCT, a novel phase‐matching DIR scheme was developed to generate 4D pseudo‐CT images for 4D dose reconstruction. Finally, TDR module creates TPS scripts to automate Tx‐Dose calculation on the pseudo‐CT images. Both initial quantitative commissioning and patient‐specific qualitative quality assurance of the DIR tool were utilized to ensure the DIR quality. The treatment doses of ten patients (six SBRT‐lung, two head and neck (HN), one breast and one prostate cancer patients) were retrospectively constructed and evaluated. The target registration error (mean ± STD: 1.05 ± 1.13 mm) of the DIR tool is comparable to the interobserver uncertainty (0.88 ± 1.31 mm) evaluated by a publically available lung‐landmarks dataset. For lung SBRT patients, the D99 of the final cumulative Tx‐Dose of GTV is 93.8 ± 5.5% (83.7–100.1%) of the originally planned D99. CTV D99 decreases by 3% and mean ipsilateral parotid dose increases by 11.5% for one of the two HN patients. In conclusion, we have demonstrated the feasibility and effectiveness of a treatment dose verification system in our clinical setting.

the treatment progress in the R&V system and triggers the PCG module when new CBCT is available. PCG retrieves the CBCTs and performs planning CT to CBCT deformable registration (DIR) to generate pseudo-CT. The 4D-CBCT images are taken for target localization and correction in lung cancer patient before treatment.
To take full advantage of the valuable information carried by 4D-CBCT, a novel phase-matching DIR scheme was developed to generate 4D pseudo-CT images for 4D dose reconstruction. Finally, TDR module creates TPS scripts to automate Tx-Dose calculation on the pseudo-CT images. Both initial quantitative commissioning and patient-specific qualitative quality assurance of the DIR tool were utilized to ensure the DIR quality. The treatment doses of ten patients (six SBRT-lung, two head and neck (HN), one breast and one prostate cancer patients) were retrospectively constructed and evaluated. The target registration error (mean ± STD: 1.05 ± 1.13 mm) of the DIR tool is comparable to the interobserver uncertainty (0.88 ± 1.31 mm) evaluated by a publically available lung-landmarks dataset. For lung SBRT patients, the D 99 of the final cumulative Tx-Dose of GTV is 93.8 ± 5.5% (83.7-100.1%) of the originally planned D 99 . CTV D 99 decreases by 3% and mean ipsilateral parotid dose increases by 11.5% for one of the two HN patients. In conclusion, we have demonstrated the feasibility and effectiveness of a treatment dose verification system in our clinical setting.

| INTRODUCTION
Treatment delivered dose in patient could vary from the planned one due to patient setup, target motion, and anatomic variation. 1,2 For lung cancer treatment, the percentage of ITV volume changes of a cohort of 40 Stereotactic Body Radiation Therapy (SBRT) lung patients were reported to range from −59.6% to 13.0% (−21.0 ± 21.4%) at the end of treatment. 3 The anatomic change caused by atelectasis, pleural effusion, and pneumonia could also significantly affect the dose distribution. 4,5 In addition, the breathing pattern variations during treatment in magnitude, period, and mean position could negatively impact the delivered dose. 6 For head and neck cancer patient, Vasquez Osorio et al. reported that the primary tumor shrunk by 25 ± 15% and the ipsilateral parotid grand by (20 ± 10%). 7 For radiotherapy in the pelvic and abdominal region, the organ filling and deformation, which may not be fully correctable by couch shifting, could lead to significant treatment dose deterioration of critical organs. [8][9][10][11][12] As a result, treatment dose assessment is a useful technique to monitor delivery accuracy for patients who may undergo notable tumor regression/progression or anatomy/motion variation.
A clinical treatment dose monitoring system can serve as an important infrastructure to support the radiotherapy treatment quality evaluation, adaptive radiation therapy decision-making, and treatment outcome modelling. 13,14 For example, the reconstructed daily treatment dose can be utilized to monitor the target coverage and organ at risk (OAR) sparing during the delivery. The cumulative treatment dose from previously delivered fractions can be used to support plan adaptation decisions. 15 Among all the factors, patient anatomic variation during radiotherapy is considered to be the number one source of uncertainty for radiobiology modeling. 16 Therefore, the reconstruction of cumulative treatment dose based on CBCT could became a prerequisite for accurate treatment outcome modeling. 2,17 Numerous studies have demonstrated the feasibility of using CBCT to construct treatment dose for photon [18][19][20][21][22][23] and recently for proton radiotherapy 24,25 on various treatment sites. However, few such systems actually are running in real clinical routine due to various practical reasons. Three major obstacles could be the limited interoperability among different software systems in a specific clinical setting, the extra clinical workload for an already busy clinic and the concern of deformable image registration (DIR) uncertainty behind dose accumulation. [26][27][28] Ideally, such a system should integrate seamlessly with the existing treatment planning system, recording and verify system, and onboard imaging system. All the necessary information such as treatment plan, delivery schedule, and CBCT image should be retrieved automatically. Also, the clinical workflow should be as intuitive and automate as possible to minimize the extra clinical workload. Furthermore, as emphasized in the recently published AAPM Task Group report 132 (TG 132), a convenient patient-specific quality assurance (QA) of DIR between CT and CBCT is essential to ensure the quality of pseudo-CT creation and treatment dose warping 29 .
Early stage non-small cell lung tumor could be difficult to identify on 3D-CBCT due to inferior image quality and motion artifact. 30,31 The 4D-CBCT technique has been widely adopted for daily imageguided alignment of moving targets. 32

2.A.2 | PCG
PCG module automatically retrieves CBCT images and couch corrections from CBCT imaging system database (XVI). A research version of a commercial DIR tool (ADMIRE 2.0, Elekta Inc.) is utilized for the DIR between CT and CBCT. Briefly, the intrapatient algorithm of ADMIRE performs a block-wise nonlinear registration to get a robust initial alignment, followed by a dense local correlation coefficient (LCC)-based deformable registration to get the final deformable vector field (DVF). This tool has been reported and evaluated in several international challenges of head and neck and lung patients DIR with high-ranking results. [35][36][37][38] It was also comprehensively evaluated in our institution for Head and Neck cancer patient with expert-delineated contours as ground truth, including seven OARs (brain stem, cord, L/R parotids, L/R submandibular gland, and mandible). 39 The

2.A.3 | TDR
This module takes care of the treatment dose recalculation. When a pseudo-CT is ready, TDR is triggered to generate a script file that automates the following steps: loading of pseudo-CT into TPS, dose A diagram illustrating the 4D-CBCT-based treatment dose reconstruction is shown in Fig. 3. After the 4D-pseudo-CT is generated, the images are load into TPS to calculate doses on all ten phases.

2.C | The clinical workflow
The clinical workflow as illustrated in Fig. 4   2.D | The initial commissioning of the DIR tool for lung patient It is difficult and time consuming to define corresponding landmarks for lung between CT and CBCT images. Instead, the landmarks on ten lung 4D-CT cases from a public dataset (www.dir-lab.com) were utilized to evaluate the accuracy of the DIR tool on lung region. 40,41 Briefly, this dataset includes large numbers of evenly distributed   Figure 5 shows the mean/STD of the TRE of our DIR tool per case, together with the best reported results and interobserver uncertainty from the DIR-Lab website. The original landmark distance before DIR is 8.52 ± 5.57 (mean ± STD) mm. After DIR, the mean TRE of our DIR tool for all ten cases is 1.05 ± 1.13 mm, as compared with the interobserver uncertainty of 0.88 ± 1.31 mm and the best and slightly lower uncertainty than expert observers. With GPU acceleration, the DIR of a typical lung CT-CBCT pair takes less than 1 min.

| RESULTS
The dose parameters representing target coverage for all ten clinical patients are listed in Table 1. It should be noted that the plan doses of lung patients were recalculated on ten phases and accumulated to the reference phase (4D plan-dose). For all lung patients, the targets were well covered by the prescription dose as indicated by the GTV D 99 . Figure 6A(1-3) show the plan dose, final cumulative 4D Tx-dose, and DVH for lung patient 1. Only a slight difference was observed in the high-dose region inside the GTV (red color wash). The planned and final cumulative treatment doses of patient 6 are shown in Fig. 6B(1-3). For this patient, the daily setup was very challenging with a small tumor and vessel bifurcations nearby.
Two out of five fractions were found to have inferior localization.
The target was covered by the prescription dose as shown in the final cumulative doses, but much different than the planned one (Fig. 6B3).
The variations of clinically relevant dose parameters of OAR are listed in Table 2  The planned and final cumulative dose are shown in Fig. 8A(1-3) for the prostate patient and Fig. 8B  Patient-specific HU tables have been proposed for dose calculation on 3D CBCT with acceptable accuracy for lung cancer patient. 23 The HU/density table is generated by manually selecting the homogeneous areas of different tissues on both the CT and the CBCT, which is time consuming and error prone. And the HU variation and potential motion artifact of the CBCT could have more significant F I G . 6. Example cases of lung SBRT patients: patient 1 (A1-A3) prescription 1200 × 5 cGy, patient 6 (B1-B3) prescription 1000 × 5 cGy.  29 The QA of the DIR algorithm, which is the under-the-hood key technique of this system, was done in two steps. The first step is the initial quantitative commission with a publicly available lung landmark data set and expert-delineated contours.
The second step is the daily patient-specific qualitative assessment by fusion image and propagated contour. Currently, we set empirical thresholds for the volume change of target and solid critical organ (>10%) and for the translation of target center of mass (>3 mm).
Any variation over the preset threshold will trigger notification to user. The user will inspect the autopropagated contour on the CBCTs as well as the fusion images to determine the DIR quality. Although the implementation of this system was based on our single institution's configuration, the system's modular architecture, generic XML file-based treatment representation and well-defined functional interfaces make the future extensions quite straightforward. For example, to take advantage of faster GPU based dose calculation, we plan to add RayStation as an optional dose reconstruction tool. This can be achieved by adding a new script generating function for the RayStation dose calculation/exportation in the TDR module without interfering with any existing functions.
Future development is also planned for adding the CBCT-based treatment dose monitoring for the spots-scanning proton therapy, which suffers more substantial impact than photon treatment from anatomical variation with its sharp distal falloff. 24,45

| CONCLUSION
The feasibility and effectiveness of a treatment dose verification system were demonstrated on various patient sites. A novel phasematching DIR between 4D-CT and CBCT has been developed to reconstruct 4D treatment dose for lung SBRT treatment. This system can serve as an infrastructure for routine treatment quality monitoring as well as for outcome modeling based on true delivered dose.

ACKNOWLEDG MENTS
This project was supported by Elekta R&D grant. The authors would like to thank Dr. Xin Wu for reviewing the DIR results of lung patients.

CONFLI CT OF INTEREST
The authors declare no conflicts of interest.