An Automated knowledge‐based planning routine for stereotactic body radiotherapy of peripheral lung tumors via DCA‐based volumetric modulated arc therapy

Abstract Purpose To develop a knowledge‐based planning (KBP) routine for stereotactic body radiotherapy (SBRT) of peripherally located early‐stage non‐small‐cell lung cancer (NSCLC) tumors via dynamic conformal arc (DCA)‐based volumetric modulated arc therapy (VMAT) using the commercially available RapidPlanTM software. This proposed technique potentially improves plan quality, reduces complexity, and minimizes interplay effect and small‐field dosimetry errors associated with treatment delivery. Methods KBP model was developed and validated using 70 clinically treated high quality non‐coplanar VMAT lung SBRT plans for training and 20 independent plans for validation. All patients were treated with 54 Gy in three treatments. Additionally, a novel k‐DCA planning routine was deployed to create plans incorporating historical three‐dimensional‐conformal SBRT planning practices via DCA‐based approach prior to VMAT optimization in an automated planning engine. Conventional KBPs and k‐DCA plans were compared with clinically treated plans per RTOG‐0618 requirements for target conformity, tumor dose heterogeneity, intermediate dose fall‐off and organs‐at‐risk (OAR) sparing. Treatment planning time, treatment delivery efficiency, and accuracy were recorded. Results KBPs and k‐DCA plans were similar or better than clinical plans. Average planning target volume for validation was 22.4 ± 14.1 cc (7.1–62.3 cc). KBPs and k‐DCA plans provided similar conformity to clinical plans with average absolute differences of 0.01 and 0.01, respectively. Maximal doses to OAR were lowered in both KBPs and k‐DCA plans. KBPs increased monitor units (MU) on average 1316 (P < 0.001) while k‐DCA reduced total MU on average by 1114 (P < 0.001). This routine can create k‐DCA plan in less than 30 min. Independent Monte Carlo calculation demonstrated that k‐DCA plans showed better agreement with planned dose distribution. Conclusion A k‐DCA planning routine was developed in concurrence with a knowledge‐based approach for the treatment of peripherally located lung tumors. This method minimizes plan complexity associated with model‐based KBP techniques and improve plan quality and treatment planning efficiency.


| INTRODUCTION
Stereotactic body radiation therapy (SBRT) of lung tumors is an alternative treatment modality to surgery for early stage non-smallcell lung cancer (NSCLC) patients, boasting local control rates greater than 97% at 3-yr. [1][2][3] These outstanding clinical outcomes were predominantly based on traditional lung SBRT treatments performed with 7 to 13 coplanar/non-coplanar three-dimensional (3D)-conformal static beams or with a few dynamic conformal arcs (DCA). 4,5 With modern advances in technology, lung SBRT can be delivered using intensity modulated radiation therapy (IMRT) or volumetric modulated arc therapy (VMAT). VMAT offers the most conformal dose distribution with higher chances of sparing organs-at-risk (OAR). 6 When coupled with a 6 MV flattening filter free (FFF) beam, VMAT benefits are enhanced by providing higher dose rates, reduction in out of field dose, and improved coverage at the lung-tumor interface when compared to traditionally flattened beams. 7,8 The generation of a high quality VMAT lung SBRT plan can require multiple iterations of optimization due to difficult patient geometry, tumor size, or location. 9 In general, inverse planning heavily depends on a planner's experience, treatment planning time, and planner's skill.
Inter-planner variability potentially leads to inconsistent plan quality and reduced patient safety. 10 Efforts have been made to increase treatment planning efficiency and plan quality using a form of inverse planning automation known as knowledge-based planning (KBP). 11 Modelbased KBP is a commonly utilized automatic planning strategy that gathers historical treatment planning data to predict achievable OAR doses for a prospective patient. 12 This form of KBP has demonstrated success in creating dosimetrically similar or better plans when compared to manual planning across many treatment sites with limited but recently increasing literature for lung SBRT. [12][13][14][15][16][17][18] However, a major concern with using KBP for lung SBRT is its tendency to increase total monitor units (MU) and overall plan complexity 12,18 which can increase treatment delivery complexity leading to unintended consequences particularly with VMAT plans. This includes the interplay effect between MLC motion and the tumor motion due to breathing cycle. 19

2.A | Patient population and clinical plans
Approval from our institutional review board was obtained to utilize 90 clinically treated patients' treatment plans for peripherally located early stage, NSCLC tumors that met the criteria set forth by RTOG-0618. Motion management for these patients was primarily performed using abdominal compression unless the patient presented with a comorbidity that did not allow for compression, in these cases a 4D-CT scan was performed. A gross tumor volume (GTV) was delineated in a lung window and a planning target volume (PTV) was created with added margins of 1.0 cm superior/inferior and 0.5 cm laterally per protocol guidelines. With the 4D-CT scan, the PTV was generated using a 0.5 cm isotropic margin around the internal target volume (ITV). OARs were contoured per RTOG-0618 guidelines.
Clinical non-coplanar VMAT plans were created in Varian's Eclipse

2.B | Development and validation of KBP model
The new KBP model was trained and validated using 90 previously treated high quality non-coplanar VMAT lung SBRT plans. Seventy plans were selected for training and the remaining 20 plans were used for validation. Prior to input, training and validation datasets were manually verified to have correct calculation models and grid sizes (e.g., PO MLC algorithm and Acuros-XB algorithm enabled).
Training contours and overall plan quality were then evaluated for compliance per RTOG-0618 guidelines. Once the KBP model was verified, normal tissue constraints, and dose objectives were iteratively selected.
To ensure the KBP model was fully functional and robust, 20 independent patients were specifically selected to include both Meaning, they were normalized so that at least 95% of the PTV volume received full prescription dose.

2.C | Dosimetric comparison criteria
Re-optimized plans were evaluated for target conformity, dose gradient and intermediate dose spillage as described by RTOG-0618. Target conformity was assessed using the conformity index defined as the ratio of the 100% isodose line volume to PTV volume. Dose gradient was assessed using the RTOG recommended gradient index (GI) defined as the ratio of the 50% isodose line volume to the PTV volume. The maximum dose 2 cm away from the PTV (D2cm) in any direction and the gradient distance (GD), defined as the average distance between the 100% and 50% isodose lines, were used to quan-

2.D | A novel k-DCA planning routine
To integrate the benefits of both traditional planning techniques and modern lung SBRT treatment practices using VMAT optimization, a routine was developed to improve the plan quality and patient safety in prospective treatments. This routine creates a k-DCA plan using a combination of manual and automated planning approaches with minimal deviation from traditional planning workflow. To begin, planning geometry is manually determined by deploying dynamic conformal arcs and collimator angles. An MLC is then added to each field and is fit with a 2-mm margin around the PTV on each DCA. Within the PO algorithm (v15.0 or higher) exists the new MLC aperture shaper controller (ACS). Following creation of planning geometry, the ACS is adjusted from its default setting of 'low' to 'very high.' This is modified to aid in the reduction of MLC modulation in the final plan.
Once this aperture setting is applied, a 3D DCA-based dose is calculated and field weights are adjusted to give a practical starting point and a base dose for the future VMAT optimization. Following the DCA-based dose calculation, the VMAT optimizer is launched and DVH estimates are automatically generated by enabling the KBP model (see Section 2.B) within the VMAT optimization window.
VMAT optimization is performed using the newly and automatically generated dose optimization objectives and priorities via the KBP model. Fig. 1 outlines this process.

2.E | Independent dose verification
To verify knowledge-based plans independently, patient-specific quality assurance was performed using an in-house Monte Carlo (MC) program. 24,25 This was performed in lieu of traditional based quality assurance measurements as recent technological advancements in online/offline-adaptive re-planning strategies may not allow enough time to perform a conventional physical measurement. 26 The in-house MC code uses a vendor provided phase space file to base its functionality off the PENELEOPE MC code. 27 Rather than physical measurement of multi-leaf collimators at the machine, a vendor provided schematic was used to model in the MC code and independent dose verification. More details of this algorithm used for this physics second check tool can be found in the cited literature above.

3.A | Clinical plans vs KBPs
Knowledge-based plans produced similar or better target coverage than clinical plans (Table 1) (Table 1).

3.B | Clinical plans vs k-DCA plans
When the proposed automatic planning routine to create a k-DCA plan was deployed, a higher target dose was achieved at minimal costs to plan quality when compared to clinical plans. The GTV minimum dose was escalated on average 3.7 Gy in k-DCA plans. This is due to the increased average MLC aperture size and less MLC modulation through the target. PTV target metrics showed higher dose with an increase in mean dose by an average of 2.9 Gy (P < 0.001) with no clinically significant differences in PTV minimum coverage.
Despite the higher delivered GTV dose, the CI differences between the k-DCA plans and clinical plans were insignificant. As expected, and following the trend of KBPs, k-DCA plans were more homogenous indicated by the lower GI (P = 0.005) and delivered less intermediate dose spillage reflected in a lower values of GD (P = 0.004).
D2cm was slightly increased in k-DCA plans with respect to clinical plans but this increase was not statistically significant (Table 1). Normal lung tissue sparing was tracked for V5Gy, V10Gy and MLD because literature suggested these better predict radiation-induced pneumonitis than V20Gy [28][29][30] (Table 2). For V5Gy, V10Gy, V20Gy and MLD, KBPs were able to significantly reduce (all P < 0.001) the dose to normal lung when compared to clinical plans. This suggests that in most cases KBPs show reduced normal lung doses and could potentially allow for re-treatment in future as needed. Clinical plans delivered higher doses to normal lung tissue across all metrics when compared to k-DCA plans, however only V5Gy (P = 0.006) and V20Gy (P < 0.001) were statistically significant. This could correlate to a potential lower risk of radiation-induced pneumonitis via k-DCA plans.

3.C | OAR sparing
A dose color wash distribution with both the axial and coronal views of an example validation case is shown (Fig. 3). Corresponding dose-volume histogram is shown in Fig. 4. Select OAR are also shown for reference to the tumor location. Highly conformal radiosurgical dose distribution with a tighter 50% isodose colorwash (blue) can be observed in both clinical and KBPs, however, there was a reduced central hotspot in both plans when compared to the k-DCA plan. This reflects our findings that k-DCA routine was able to increase minimum dose to GTV. This larger central hotspot displayed in the k-DCA was achieved with minimal to no additional costs in OAR dosing. It is shown that the 50% isodose color wash was slightly larger in the k-DCA axial slice but still easily met RTOG-0618 criteria.

3.D | Planning efficiency and deliverability
The k-DCA plans were generated using plan geometry identical to previously treated plans in less than 30 min. Table 3

| DISCUSSION
A novel automatic planning routine was developed to generate a non-coplanar VMAT lung SBRT k-DCA plans in less than 30 min.
Both conventional and the k-DCA planning routine generates a similar or better-quality plan than manually planning. This method reduces inter-planner variability and lowers the plan complexity when compared to original clinical and conventional knowledge-  12 This reflects similar findings in our study but using our automated k-DCA planning routine we were able significantly reduce the total MU (see Table 3 Mean ± SD and P-values were reported. n. s. = not significant. SD = standard deviation. Significant values are highlighted in bold.
online adaptive re-planning or same/next day offline adaptive replanning of lung SBRT treatment. It has previously been shown that 30 Gy in a single fraction can be delivered to the lung lesion in a 15-min time slot. 32 Delivering a single fraction treatment subjects the plan to delivery potential errors that could greatly enhanced the interplay effect, so our k-DCA routine could potentially limit this effect by providing less MLC modulation across the target and improve small-field dosimetry. 33 Further validation and clinical implementation of this KBP model and automated k-DCA routine for SBRT patient treatment is underway.

CONF LICT OF I NTEREST
None.