Development and clinical validation of a robust knowledge-based planning model for stereotactic body radiotherapy treatment of centrally located lung tumors.

PURPOSE
To develop a robust and adaptable knowledge-based planning (KBP) model with commercially available RapidPlanTM for early stage, centrally located non-small-cell lung tumors (NSCLC) treated with stereotactic body radiotherapy (SBRT) and improve a patient's "simulation to treatment" time.


METHODS
The KBP model was trained using 86 clinically treated high-quality non-coplanar volumetric modulated arc therapy (n-VMAT) lung SBRT plans with delivered prescriptions of 50 or 55 Gy in 5 fractions. Another 20 independent clinical n-VMAT plans were used for validation of the model. KBP and n-VMAT plans were compared via Radiation Therapy Oncology Group (RTOG)-0813 protocol compliance criteria for conformity (CI), gradient index (GI), maximal dose 2 cm away from the target in any direction (D2cm), dose to organs-at-risk (OAR), treatment delivery efficiency, and accuracy. KBP plans were re-optimized with larger calculation grid size (CGS) of 2.5 mm to assess feasibility of rapid adaptive re-planning.


RESULTS
Knowledge-based plans were similar or better than n-VMAT plans based on a range of target coverage and OAR metrics. Planning target volume (PTV) for validation cases was 30.5 ± 19.1 cc (range 7.0-71.7 cc). KBPs provided an average CI of 1.04 ± 0.04 (0.97-1.11) vs. n-VMAT plan's average CI of 1.01 ± 0.04 (0.97-1.17) (P < 0.05) with slightly improved GI with KBPs (P < 0.05). D2cm was similar between the KBPs and n-VMAT plans. KBPs provided lower lung V10Gy (P = 0.003), V20Gy (P = 0.007), and mean lung dose (P < 0.001). KBPs had overall better sparing of OAR at the minimal increased of average total monitor units and beam-on time by 460 (P < 0.05) and 19.2 s, respectively. Quality assurance phantom measurement showed similar treatment delivery accuracy. Utilizing a CGS of 2.5 mm in the final optimization improved planning time (mean, 5 min) with minimal or no cost to the plan quality.


CONCLUSION
The RTOG-compliant adaptable RapidPlan model for early stage SBRT treatment of centrally located lung tumors was developed. All plans met RTOG dosimetric requirements in less than 30 min of planning time, potentially offering shorter "simulation to treatment" times. OAR sparing via KBPs may permit tumorcidal dose escalation with minimal penalties. Same day adaptive re-planning is plausible with a 2.5-mm CGS optimizer setting.

escalation with minimal penalties. Same day adaptive re-planning is plausible with a 2.5-mm CGS optimizer setting.

K E Y W O R D S
adaptive re-planning, centrally located lung SBRT, knowledge-based planning, RapidPlan model

| INTRODUCTION
Stereotactic body radiotherapy (SBRT) for early stage localized nonsmall cell lung cancer (NSCLC) has become a significant treatment option to traditional surgical intervention providing primary tumor local control rates in excess of 97% (median, 3 year). 1,2 Historically, lung SBRT was delivered using 7-13 co/non-coplanar static beams or dynamic conformal arcs (DCA), followed by intensity modulation radiation therapy (IMRT) and more recently with volumetric modulated arc therapy (VMAT). 1,3,4 VMAT provides more conformal dose distribution to the target better sparing of organs-at-risk (OAR) and much faster treatment delivery. The dosimetric advantages of VMAT can be enhanced using 6MV-flattening filter free (6MV-FFF) beam for lung SBRT because of its higher dose rates and reduction of outof-target dose with respect to traditional flattened beams. 5 This provides clinical benefits to the patients as it improves target coverage at the lung-tumor interface and shorter treatment time; potentially improving patient convenience and reducing intrafraction motion errors. 6 In North America, the Radiation Therapy Oncology Group (RTOG) reports provides recommendations to clinicians for SBRT dosing schemata and contouring guidelines based on operable eligibility and tumor geographical location. This study concentrates on SBRT for early stage NSCLC patients with centrally located tumors following RTOG-0813 guidelines. 7 In addition to centrally located lung tumors, our clinic uses this protocol for risk-adapted prescriptions for tumors located adjacent to critical structures such as the ribs.
Generating an optimal SBRT treatment using a VMAT approach requires multiple iterations and heavily depends on a planner's skill.
This potentially results in inconsistent plan quality known as interplanner variability. 8,9 Automation of inverse planning via knowledgebased planning (KBP) aims to remove interplanner variability, improve plan quality, and decrease planning time. 10 KBP uses a model library of previously generated high-quality clinical plans to predict new treatment parameters, effectively generating new plans based on a clinic's treatment planning history. 11 A Varian RapidPlan (Varian Medical Systems, Palo Alto, CA, USA) model is a KBP engine that utilizes a knowledge-based dose-volume histogram (DVH) algorithm to estimate the DVH that can produce optimization objectives such as maximum, minimum, and new line dose constraints with associated priority values. 12 KBP has demonstrated the ability to create improved or equivalent plans for prostate, head and neck, spine, breast and thoracic sites. 8,[13][14][15][16][17][18] However, there is very limited literature available for lung SBRT treatments, 14,15,17 specifically utilizing highly conformal non-coplanar VMAT (n-VMAT) planning geometry.
In this report, a RapidPlan model is described to generate adaptable n-VMAT-based KBP treatment plans for early stage NSCLC patients with medically inoperable centrally located tumors that follows RTOG-0813 dosing schemata and contouring guidelines. Our model is exclusively trained with clinically treated high-quality n-VMAT lung SBRT plans using the advanced AcurosXB final dose calculation algorithm. We use the advanced AcurosXB algorithm for heterogeneity corrections for lung SBRT treatments as it provides a more accurate dose calculation in heterogeneous patient anatomy by better modeling secondary build-up in tissue/low-density interfaces than traditionally used superposition/convolution algorithms. 19,20 The KBP model may permit the improvement of "simulation to treatment" time from our current average 7 working days to 3 days while maintaining plan consistency and reducing interplanner variability. This may enable same or next day adaptive treatments (if needed) that aim to account for day-to-day changes in physiological characteristics or setup errors as they occur during a treatment course. A previous study using a smaller calculation grid size (CGS) of 1.25 mm vs. 2.5 mm in manually optimized VMAT lung SBRT plans with the photon optimizer (PO) algorithm demonstrated minimal dosimetric differences between the two plans but has not yet been evaluated in a lung SBRT KBP setting. 21 This led to further evaluation of the concept by generating KBPs with a CGS of 2.5 mm which drastically decreases treatment planning time (mean, 5 min) and observe if they provide similar plan quality to the KBPs plans optimized with a 1.25mm CGS.

2.A | Patient population and target definition
Following approval from our Institutional Review Board (IRB), 106 clinically treated high-quality n-VMAT lung SBRT plans generated for patients with early stage centrally located tumors as defined by RTOG-0813 were selected for training and validation. Eighty-six plans were used for training this model and the remaining 20 were  allowed. OARs such as spinal cord, ipsilateral brachial plexus, skin, esophagus, heart, trachea, total lungs minus PTV, ribs, and bronchial tree were delineated per RTOG-0813 compliance criteria for dose tracking.

2.B | Clinical n-VMAT plans
For all patients, n-VMAT SBRT plans were generated in the Eclipse treatment planning system (Varian Medical Systems, Palo Alto CA) using 3-6 (mean, 4) partial non-coplanar arcs (with ±5°-12°couch kicks) on Truebeam Linac (Varian Palo Alto, CA) consisting of standard millennium 120 MLC and 6MV-FFF (1400MU/min) beam. Jaw tracking option was enabled for each arc and optimal collimator angles were selected to minimize non-target dose and enhance plan conformity. Clinical plans were optimized using Photon Optimizer (v13.6 or v15.6) algorithm with either 1.25-mm or 2.5-mm voxel resolution. The final dose calculation was performed using the advanced AcurosXB algorithm with dose to medium reporting mode. A dose of 50 Gy or 55 Gy in five treatments was prescribed to cover at least 95% of the PTV receiving 100% of the prescribed dose ensuring that all hotspots were within the PTV. Before approval, each plan was rigorously evaluated by our treating physicians via RTOG-0813 compliance criteria and institutional guidelines including dose to OAR listed below:

2.C | KBP model input and training datasets
An extensive iterative training approach was developed to create this novel and comprehensive KBP model for SBRT of centrally located lung tumors. Eighty-six n-VMAT plans were retrospectively selected and verified to be high quality by evaluating the numbers of partial arcs and total MU consistency based on historical treatment planning practice. Original (unaltered) clinical VMAT plans were used for model training. The primary focus of this plan selection process was examination of RTOG-0813 criteria. Each plan contour was individually verified to be consistent and correct. A total lung minus PTV structure was added for each patient's plan if the structure was not previously created. Calculation models consisting of dose calculation algorithm, VMAT MLC optimizer and CGS were verified to be Acur-osXB for a 2.5-mm resolution voxel size and photon optimizer for a 1.25-mm or 2.5-mm voxel size, respectively. Optimal collimator angle and jaw tracking options were verified prior to input of the training plans. To make the model fully comprehensive for RTOG compliance, it was necessary to track and select plans of varying target size and tumor geographical locations (e.g., lower lobe vs. upper lobe, right lung vs. left lung) encompassing the both lungs (see Table 2). and residual plots were evaluated for each OAR were used for manual verification of potential outliers. 22 This approach was combined with observing the Cook's distance that indicated influential data points in a regression model and the modified Z-score, which measures the difference of an individual geometric parameter from the median value in the training set. 23 Once true outliers were identified; the entire plan or specific outlying structure was removed from the model and all data was re-extracted. A summary of the KBP model refinement process is shown in Fig. 2.
Constraints were placed on a given OAR following successful verification of the model to create a fully robust model for centrally located lung tumors and risk adapted tumor location such as those tumors abutting the rib (see Table 1). Theses constraints were chosen based on RTOG-0813 guidelines and our historical treatment planning practice.

2.E | Validation of the KBP model
A total of 20 clinical n-VMAT plans that were not used to generate the RapidPlan model were selected for final verification including recently treated lung SBRT patients where dedicated manual planning time was recorded ( Table 2). These plans were specifically selected to encompass both lungs' geometry and variable target sizes to fully test the functionally of our model's robustness. However, plan quality was not evaluated prior to selection to ensure the model could produce optimal plans for various case complexities. The overall validation set included 16 patients who received 50 Gy and 4 patients who received 55 Gy in 5 fractions, respectively. These plans were re-optimized with the RapidPlan model with identical planning geometry as the clinical n-VMAT plans. KBPs were created from a single optimization with no manual intervention. Target dose coverage for the KBPs was normalized for identical or better target coverage compared to previously used clinical n-VMAT plans.
To fully assess the performance of this new KBP model, we eval-

3.A | Dosimetric criteria
Knowledge-based plans were able to provide similar or better target coverage than clinical n-VMAT plans (Table 3). KBPs had a slightly F I G . 1. KBP-model training input data selection workflow for centrally located lung SBRT: A total of 86 high-quality clinical n-VMAT plans were selected to train this model that met RTOG-0813 requirements for contouring and OAR dose tolerances while using Acuros-based dose calculation.  These results are shown in Table 4. KBPs had an average lower V5Gy by 0.6%, (P < 0.001), V10Gy by 0.5% (P < 0.001), and MLD by 0.12 Gy (P < 0.001) suggesting a potentially lower risk of radiation-induced pneumonitis. In addition to normal lung tissue doses, all other OAR compliance criteria were assessed per RTOG-0813 ( Fig. 3). In many lung SBRT cases, risk-adapted prescription to targets adjacent to the ribs are used. The greatest sparing achieved in the KBPs was shown in the ribs (P < 0.001) for an average of 2.62 Gy (maximum up to 9.67 Gy).
Our study showed that the ipsilateral brachial plexus, esophagus, heart, trachea, and bronchial tree received an insignificant average lower dose in KBPs compared to the clinical n-VMAT plans. Additionally, KBPs on average presented an insignificant but slightly higher skin dose to spinal cord by 0.46 Gy (P = 0.32). In the patient-specific quality assurance measurements, the gamma analysis of 2%/2 mm criteria was used to assess the plan delivery accuracy differences between KBP vs. clinical n-VMAT plans. KBPs presented with a similar average pass rates of 94.4 ± 2.7% (range, 90.6-100.0%) compared to n-VMAT plans with an average pass rates of 95.4 ± 2.3% (range, 90.9-99.4%) (P = 0.11) plans suggesting that comparable treatment delivery accuracy can be achieved with KBPs.

3.C | Example validation case -Left lower lobe tumor
Dose-volume histograms of both the KBP and n-VMAT plan for a validation case with a left lower lobe tumor of a lung SBRT patient

3.D | Re-optimized KBPs with 2.5 mm CGS
The KBP calculation time was dictated by the CGS used in the optimization window. The original KBPs were calculated with a 1.25-mm CGS. However, while using a 2.5-mm CGS the treatment planning time was reduced to approximately 5 min. This setting could support even faster adaptive re-planning in emergent clinical situations. Therefore, KBPs were re-optimized with a 2.5-mm CGS for plan evaluation.  T A B L E 5 Treatment delivery efficiency and accuracy of KBP with respect to clinical n-VMAT plans. Mean value ± SD (range) and p-values were reported for both KBP and n-VMAT plans.

CONFLI CT OF INTEREST
The author have no other relevant conflicts of interest to disclose.