Multicriteria optimization: Site-specific class solutions for VMAT plans

ARTICLE IN PRESS

JID: MDO

[mUS5Gb;May 15, 2019;4:20]

Medical Dosimetry xxx (xxxx) xxx

Medical Dosimetry journal homepage: www.meddos.org

Multicriteria optimization: Site-specific class solutions for VMAT plans Mariana Guerrero, PhD a,∗, Zachary Fellows, Bachelor of Professional Studies in Radiation Therapy a, Pranshu Mohindra, MD a, Shahed Badiyan, MD a,b, Narottam Lamichhane, PhD a, James W. Snider, MD a, Shifeng Chen, PhD a a b

Department of Radiation Oncology, University of Maryland School of Medicine, Baltimore, MD 21201, USA Department of Radiation Oncology, Washington University, St Louis, MO

a r t i c l e

i n f o

Article history: Received 20 October 2018 Revised 16 February 2019 Accepted 11 April 2019 Available online xxx Keywords: Radiation Therapy Optimization Treatment Planning IMRT VMAT

a b s t r a c t Multicriteria optimization (MCO), a novel commercially available optimization method for IMRT and VMAT has the potential to improve treatment planning techniques and workflows. MCO allows planners and physicians to assess in real time the impact and tradeoffs between all clinical goals and organ constraints. We investigate the feasibility of a universal set of objectives and constraints for VMAT plans in different anatomical sites and the impact of involving the Physician in the navigation of the generated Pareto plans. We randomly selected 20 prostate only, 14 whole pelvis, 10 advanced lung, 15 pancreas, and 7 head and neck plans planned with a VMAT technique. Using the clinically delivered isocenter and beam set-up, we retrospectively generated MCO plans with a universal set of constraints and objectives for each anatomical site. The MCO plan scores were compared with clinical plans or an independent plan generated with DMPO. For prostate only plans the TCP values for the clinical and MCO plans were similar and the rectum NTCP values and overall P+ were slightly better for the MCO plans. For whole pelvis, the resulting MCO plans were comparable in all the dosimetric measures to the clinical plans. For lung, the MCO dosimetric comparison also yielded comparable plans but when evaluating individual patients, there were 5 patients for which MCO plans had a clear advantage in reducing dose to lung and/or esophagus while improving/maintaining target coverage, 4 patients with comparable plans and 1 patient where MCO was worse. Allowing the physician to navigate independently produced a different selection of dosimetric trade-offs. Comparable MCO plans were also obtained for pancreas and head and neck. Based on our experience with many anatomical sites and a large number of patient plans, we have found that VMAT MCO plans are comparable to the clinical plans and can be produced with a universal set of objectives and constraints, even for a wide range of geometries and anatomies. © 2019 American Association of Medical Dosimetrists. Published by Elsevier Inc. All rights reserved.

Introduction The advent of intensity modulated radiation therapy (IMRT) marked a stark change in the field of dosimetry, demanding that planners became skilled in the art of inverse planning, which operates differently in each of the available planning systems in the market. In principle one would expect that inverse planning would help streamline the planning process, but the reality is that no universal set of objectives can be used due to the considerable individual variations in the inverse planning solutions for different patients. A common problem is that local minima may be achieved before the true minima of a good plan, especially if all constraints ∗ Reprint request to M. Guerrero, Ph.D., Department of Radiation Oncology, University of Maryland School of Medicine, 22 South Greene St, Baltimore, MD 21201, USA. E-mail address: [email protected] (M. Guerrero).

are met and the plan is not pushed further. Hence, an approach is needed to make plan development more consistent and additionally allow opportunities to explore trade-offs in a dynamic manner. Multicriteria optimization (MCO) is a sophisticated approach developed decades ago in other fields. It has only been recently applied to treatment planning in radiation oncology.1,2 The basic concept involved in MCO is that of Pareto optimized solutions, which are defined as solutions where no objective can be improved without worsening another objective. MCO does not find a unique optimized plan but rather a set of plans that are part of the so-called Pareto surface. The user is then able to “navigate” from one plan to another while visualizing the effect in the dose distributions while changing plans. There have been a number of articles describing the use of MCO for specific sites.3-12 Some studies were limited to IMRT static beams.3-7 Others discussed VMAT for specific sites.9-12 None

https://doi.org/10.1016/j.meddos.2019.04.003 0958-3947/© 2019 American Association of Medical Dosimetrists. Published by Elsevier Inc. All rights reserved.

JID: MDO 2

ARTICLE IN PRESS

[mUS5Gb;May 15, 2019;4:20]

M. Guerrero, Z. Fellows and P. Mohindra et al. / Medical Dosimetry xxx (xxxx) xxx

of these articles focused on the standardization of treatment planning. The goal of our study is to develop a universal set of objectives and constraints for each individual anatomical site to streamline the planning process. We focus exclusively on VMAT plans given that this technique is prevalent in our clinic and many others due to its efficient delivery time and excellent plan quality. Our hypothesis is that the specific anatomy and geometry of each patient will come into play during the navigation process. MCO is an important tool that will help us continue to standardize our clinical practice, not just with a specific set of clinical goals but in terms of the produced plans, regardless of the background and experience of the dosimetrists generating the plans. Furthermore, this study explores the possibility of Physician navigation of the plans, which can lead to a higher impact of clinician input in the resulting plans. Materials and Methods

of the prescription isodose inside the target to the total volume of the prescription isodose. This same definition of CI index was used for all anatomical sites. P+ was defined as the difference between the probability of benefit minus the probability of injury, PB -PI , with PB = Prostate TCP and PI = 1-(1-NTCPBladder )(1-NTCPRectum ). This definition assumes total correlation between PB and PI, which is the worstcase scenario. This assumption is not critical for our work since we are using P+ to compare 2 different plans rather than in absolute values. For prostate we used a Poisson model for TCP = exp(-N0 exp(-n(α d-β d2 ))) where N0 is the initial clonogenic cell number, d is the dose per fraction and n is the number of fractions with linearquadratic (LQ) model parameters parameters α = 0.15Gy−1 , α /β = 3.1Gy, Tpot = 42 days and with N0 fixed at 3 × 106 for intermediate-risk.13 For bladder and rectum NTCP, we used the Lyman-Kutcher-Burma model with the QUANTEC parameters for rectum (n = 0.09 m = 0.13 TD50 = 76.9Gy)14 and the Emami-Burman 1991 parameters for bladder (n = 0.5 m = 0.11 TD50 = 80Gy).15 Whole pelvis In order to investigate plans with more irregular shaped targets we considered 14 high-risk prostate cancer patients that received 45 Gy in 1.8 Gy fractions to the whole pelvis as their initial treatment. As before, we considered the OAR overlap with the PTV to characterize the different patient anatomies. Table 1 first 3 columns represent the complete information for each patient, showing the significant anatomical variation among patients. We used our recently developed institutional bladder and rectum constraints for initial pelvis plans (Bladder V45 < 15% and V40 < 35%, Rectum V43 < 15%, and V37 < 35%,) as well as the CI and PTV Dmax parameters.

We considered treatment plans for: prostate and seminal vesicles alone (as way to start with the simplest case and get familiar with the software), whole pelvis (to test more irregular shapes and larger volumes) and advanced lung patients (to test behavior when heterogeneity corrections are involved). Head and neck is an example of irregular shapes as well but with a larger number of OARs and higher complexity. Pancreas on the other hand is usually a simpler shape and a case where almost all clinical plans fulfill all the clinical goals but with regular planning it is hard to know how far any individual goal can be pushed without affecting the others unless a significant amount of time is invested, a question that MCO is designed to address. We used a commercial planning system (Raystation version 4.5-6), which has been implemented in our clinic for 3 years, although at the beginning of our project only the Direct Machine Parameter Optimization (DMPO) algorithm was used clinically for IMRT and VMAT plans. For each anatomical site we developed a universal set of objectives and constraints based on the vendor’s recommendations and our own extensive tests. We compared the resulting MCO plans with the clinical plans using a combination of dosimetric and biological metrics, except for pancreas, where the planner was a dosimetry student who created both a DMPO and an MCO plan independently to the best of his ability and then asked a physician to blindly select one of the two plans. We explicitly considered patients with very different patient characteristics to demonstrate that the universal set of constraints can be personalized and adapted to different circumstances doing the Pareto surface navigation. For prostate and SVs alone and whole pelvis plans the plan difficulty was measured by the percentage of the organs at risks (OARs) that overlap with the planning target volume (PTV). For advanced lung, head and neck and pancreas patients plans we considered a range of gross tumor volume (GTV) sizes and also total lung volumes for lung cases. The planning process for Raystation MCO VMAT plans can be described in 3 steps. The first step is the generation of the Pareto plans based on the desired constraints and objectives. Once these plans are computed by the planning system, the planner or the physician “navigate” from one plan to another while viewing the changes of the navigation in the DVHs and isodose distribution of targets and OARs and deciding the most clinically desirable compromise. Since the VMAT navigated plans are optimized based on the photon fluence, the second step consists of converting the desired fluence into a deliverable plan. After this step, it is likely that some degradation of the plan can occur. In order to “fix” the degradation of the plan the third step is a special use of the Direct Machine Parameter Optimization (DMPO) called “postprocessing” where the MCO plan acts as one of the objectives for the DMPO optimization. The user can add other objectives with smaller weight as needed. Typically the objectives required in this step are conformality and dose uniformity in the targets and OARs.

The clinical treatment plans were designed using a single isocenter placed in the center of the PTV. All plans were generated using 2 coplanar full arcs, with a fixed collimator angle of 10 and 350 degrees. All 15 patients received 45 Gy in 25 fractions to the PTV. Each patient was blindly planned with both (MCO) and Direct Machine Parameters Optimization (DMPO). The plans were optimized until meeting a list of standardized dose constraints deemed the golden standard of the institution. For all 15 MCO-VMAT plans, the planner used a preconstructed list of planning objectives and constraints to derive the Pareto plans from. The navigation was driven by the specific need of the plan and after conversion to deliverable it was postprocessed to achieve a desirable plan. PTV volumes were highly variable ranging from 214 cc to 934 cc.

Prostate ± proximal seminal vesicles (SVs)

Head and neck

Twenty intermediate-risk prostate cancer patient’s VMAT plans were randomly selected from our clinical database. The prostate and proximal seminal vesicles were treated to either 75.6 or 79.2 Gy in 1.8 Gy fractions or 78 Gy in 2 Gy fractions with an appropriate planning target volume (PTV) margin. The plan complexity was assessed by the percentage of overlap of the clinical structures (bladder and rectum) with the PTV, which ranged from 1% to 16% for both OARs. A set of objectives and constraints was developed and the resulting MCO plans were compared to clinical plans (generated with DMPO optimization). All plans had 2 complementary full arcs although some occasionally had 3 full arcs, one with the collimator rotated 90 degrees. All the plans follow the clinical plan beam arc selection and settings. Uncomplicated tumor control probability model (UTCP), also called P+, which combines Tumor Control Probabilities and Normal Tissue Complications Probabilities (TCP and NTCP) models together to express the treatment plan quality in a single value, was used in combination with the conformity index (CI) and the PTV max dose. For simplicity the CI used is the one defined by Raystaion: the volume

Seven head and neck initial plans were selected, all with bilateral neck nodes treatments. The patients’ doses of the initial plan were 50.4 Gy for 4 patients and 54 Gy for 3 patients all in 1.8 Gy fractions. The variability of geometry in this set is given by a variation of PTV volume, from 612 cc to 999 cc. Two coplanar arcs were used for all 7 patients.

Advanced lung In order to evaluate MCO performance with significant heterogeneity corrections, we randomly selected 10 lung cancer patients (Stage II-IV) planned with a 2 arcs VMAT technique. Prescription doses ranged from 50.4 Gy to 66 Gy in 1.8 or 2 Gy per fraction, with one case of dose painting. Target volumes were generated per clinical standard of care. The anatomical variation in this case is given by the range of the tumor volumes (the PTVs ranged from 86 cc to 719 cc) and the sizes of the total lungs (1888 cc to 4626 cc) with PTV to total lung volume ratios of 4% to 26%. Using the clinically delivered isocenter and beam set-up, we retrospectively generated the MCO plans with a universal set of constraints and objectives that was developed by trying different combinations in a few patients as before. Two versions of the MCO plan scores (one navigated by the dosimetrist [MCOd] and one by the physician [MCOp]) were compared with clinical plans generated with DMPO, the standard Raystation optimization technique. Dosimetric parameters evaluated were PTV V95% , (CI) and PTV Dmax . The biological score considered was the uncomplicated probability (UP) similar to the UP defined for the prostate patients but in this case representing the absence of lung and esophagus complications. The cord dose was not evaluated because almost all plans had Dmax doses of less than 45 Gy, with negligible probability of complications. The NTCP parameters for esophagus and lung complications were from the QUANTEC values.16 Pancreas

Results Prostate and proximal SVs The universal set of constraints and objectives for prostate were as follows (in terms of percentage of prescription dose): we always used 3 constraints (PTV Min dose 100%, PTV max dose 103%, and GTV min dose 100%) and 5 objectives (PTV uniform dose 100%, external dose fall-off from 95% to 50% in 0.6 cm, external dose falloff from 50% to 0% in 10 cm, bladder EUD = 0 with a = 1 and rectum EUD = 0 with

JID: MDO

ARTICLE IN PRESS

[mUS5Gb;May 15, 2019;4:20]

M. Guerrero, Z. Fellows and P. Mohindra et al. / Medical Dosimetry xxx (xxxx) xxx

3

Table 1 Selected dosimetric parameters for whole pelvis plans. The first three columns describe the overlap of rectum, bladder and small bowel value with PTV. The plans were normalized to have the same target coverage. All the parameters studied show no statistical significant difference between the MCO and clinical plans

Table 2 Biological and dosimetric scores of the 20 prostate plans. The top 5 were prescribed to 75.6 Gy and the bottom 15 to 78 or 79.2 Gy. The rectum-PTV overlap is shown for reference. MCO plans have a slightly better P+ (mean 85.1% vs 83.5%, p value=0.02). The CI (mean 0.94 for MCO and 0.92 for clinical plans, p value = 0.2), and PTV Dmax (MCO mean 105.4% vs 105.0%; clinical relative to prescription dose, p = 0.2) were comparable and below the 107% clinical goal.

difference with the vendor’s recommendations is the constraints of minimum and maximum dose for the PTV (100% to 103%) is a narrower range than recommended. The resulting plans however have a much wider range of PTV min-max dose given the navigation process and the conversion to deliverable degradation. Table 2 shows the dosimetric and biological scores of the MCO and clinical plans with the amount of rectum-PTV overlap as a reference. While the TCP values for the clinical and MCO plans were similar (mean 95.4% for both, p value = 0.97) (not shown), the rectum NTCP values were slightly better for the MCO plans (mean 8.9% vs 10.3%, p value = 0.03) (not shown), resulting in a better P+ for MCO plans (mean 85.1% vs 83.5%, p value = 0.02). The CI (mean 0.94 for MCO and 0.92 for clinical plans, p value = 0.2) and PTV Dmax (MCO mean 105.4% vs 105.0% Clinical relative to prescription dose, p = 0.2) were comparable and below the 107% clinical goal. However, there were a few patients with low CI in the clinical plans (displayed in red font in Table 2 when below 0.9) that were significantly improved in the MCO plan (currently there is no quantitative clinical goal in terms of CI in our clinic, but rather the qualitative analysis of the isodose lines). Three patients showed a significant improvement in rectum NTCP and consequently 5% or more gain in P+ with MCO. The bladder NTCP turned out negligible for most plans, probably due to the large TD50 and the treatments with full bladder for prostate and proximal seminal vesicles only. This TD50 value was based on Emami’s article since no new value was published in QUANTEC. Figure 1 shows the correlation of the rectum NTCP with the rectum PTV overlap for the MCO plans with a correlation coefficient of 0.84, demonstrating that while using a universal set of objectives and constraints one can still deliver a personalized plan, where the outcome correlates with the anatomy of each patient. The dose to femoral heads, penile bulb and small bowel was the same or lower for the MCO plans even though those structures were not included in the optimization. The dose-fall off function was enough to limit those doses to acceptable levels, within the institutional limits and often better than the clinical plans. Whole pelvis

a = 1). These constraints were a modified version of the one suggested by the vendor. The main modifications were that we reduced as much as possible the number of objectives and constraints: for example we did not include the femoral heads or penile bulb as objectives since we found no need for it given that the dose falloff objective is sufficient to take care of those structures. Reducing the number of objectives and constraints helped reduce the planning time and simplified the navigation process. We also reduced the distance for the dose-fall off between the 95% and 50% to accommodate rapid fall-off in difficult plans and we standardized the dose fall-off for the 50% to 0% to 10 cm as opposed to a variable quantity recommended by the vendor based on the distance between the isocenter and the patient body edge. Notice that the EUD = 0 objective is not a realistic goal since clearly all the OARs will have EUD larger than zero, but it is the way the Raystation MCO algorithm is set up so we follow the approach as recommended by Raystation. Another

The universal set of constraints and objectives for whole pelvis were the same as before with an additional objective for the small bowel EUD = 0 with a = 1. All the plans were navigated to match the target coverage for the clinical plan and therefore the target coverage was essentially the same for both MCO and clinical plans and the OARs dosimetric parameters as well as the PTV Dmax and the conformity index CI. The resulting MCO plans were comparable in all the dosimetric measures to the clinical plans except for the bladder V40 and rectum V37 , where MCO had a significantly smaller value (average bladder V40 was 28.4% ± 2.9% for MCO vs 31.7% ± 3.7% clinical, p value = 0.02; rectum V37 was 25.9 ± 2.3 for MCO vs 29.9 ± 3.7 for clinical, p value = 0.03). Bladder V45 , however were similar (average 14.7% ± 2.0% for MCO vs 14.5 ± 2.2% for clinical, p value = 0.84) and the same is true for Rectum V43 (14.4% ± 1.9 for MCO vs 15.5 ± 1.9 for clinical, p value = 0.18). PTV Dmax were comparable (108.6% ± 0.4% for MCO vs 107.7% ± 0.4% clinical, p value = 0.17) as well as CI (0.96 ± 0.01 for MCO vs 0.95 ± 0.01 for clinical, p value = 0.38). Details of selected dosimetric parameters can be found on Table 1, where the overlap of OARs with PTVs is also noted, showing large variations of patients anatomies in the selected group.

JID: MDO 4

ARTICLE IN PRESS

[mUS5Gb;May 15, 2019;4:20]

M. Guerrero, Z. Fellows and P. Mohindra et al. / Medical Dosimetry xxx (xxxx) xxx

Fig. 1. Rectum NTCP correlation with Rectum-PTV Overlap in terms of percent of rectum volume. The large correlation shows that the universal set of constraints produced personalized plans.

Fig. 2. Example of a physician-navigated plan dose distribution (upper-left, solid DVH) compared to the clinical plan (lower-left, dotted DVH). Both plans fulfill the clinical goals as seen by the green checkmarks in the right bottom panel. The physician would rather spare the esophagus more and reduce the conformity by increasing the dose anteriorly. (Color version of figure is available online.)

Advanced lung The universal set of constraints and objectives for the lung patients were an extension to those from the prostate alone patients: we used 3 constraints (PTV Min dose 100%, PTV max dose 105% and ITV min dose 100%) and the objectives were (PTV uniform dose 100%, external dose fall-off from 95% to 50% in 0.6 cm, external dose fall-off from 50% to 0% in 10 cm, ipsilateral lung EUD = 0 with a = 1, contralateral lung EUD = 0 with a = 1, esophagus EUD = 0 with a = 1 and spinal cord

EUD = 0 with a = 5 to 10). In the case of the lung patients, the conversion from the navigated plan to the deliverable plan showed the largest differences, probably due to heterogeneity corrections not properly taken into account by the pencil beam algorithm used to generate the Pareto plans. Some plans were developed without the ipsilateral lung as an objective but for the most part using both lungs as independent objectives allowed for better plans and the possibility of navigating each lung independently, even though our clinical goal referred to the total lung DVH. Using a = 5 or larger for the spinal cord allows to put more weight

ARTICLE IN PRESS

JID: MDO

[mUS5Gb;May 15, 2019;4:20]

M. Guerrero, Z. Fellows and P. Mohindra et al. / Medical Dosimetry xxx (xxxx) xxx

5

Table 3 Biological and dosimetric score comparison of planner navigated MCO and clinical plans. The blue font represents plans where MCO showed a small advantage. The green font plans showed cases with different trade-offs. The black font plans are about the same and the red font MCO plan showed a small disadvantage to the clinical plan. PTV coverage, UP and CI p-values showed equivalency but Dmax was smaller in the clinical plans (p value = 0.004)

Lung-Esophagus Uncomplicated Probability

Paent 1 2 3 4 5 6 7 8 9 10

PTV Coverage V95%

Clinical

MCO Planner navigated

Clinical

81% 87% 83% 55% 80% 82% 67% 72% 64% 68%

84% 88% 89% 60% 84% 82% 65% 72% 62% 69%

94% 100% 100% 97% 99% 99% 95% 96% 99% 99%

Conformality Index

MCO Planner Clinical navigated

in the high dose which is desirable for a serial organ like the cord or the brain stem. Dosimetry-navigated MCOd vs clinical comparison: the values for PTV V95 were MCOd: 99% (95 to 100) vs clinical plans: 98% (94 to 100), p value = 0.2, the CI were similar in both cases (p value = 0.5) but the Dmax was higher for the MCO plans (average 109% vs 107%, p = 0.004), the lung-esophagus UP was similar between MCOd (Mean: 74%, range: 55% to 87%) and the clinical plan (Mean:75%, range: 60% to 89%), p = 0.1. Corresponding MCOd vs MCOp comparisons were similar for the parameters tested. However, using real time assessment of dosimetric trade-offs, in each individual plan the physician was able to actively choose a preferred combination of dose to target vs organs. When evaluating individual patients, there were 5 patients for which MCO plans had a clear advantage in reducing dose to lung and/or esophagus while improving/maintaining target coverage, 4 patients with comparable plans and one patient where MCO was worse. The detailed scores are presented in Table 3 and an example where the physician-navigated dose distribution showed a very different trade-off than the clinical plan is shown in Fig. 2. Pancreas The universal set of constraints and objectives for the pancreas patients were also based on the those from the prostate alone patients: we used 3 constraints (PTV min dose 100%, PTV max dose 103% and GTV/CTV min dose 100%) and the objectives were (PTV uniform dose 100%, external dose fall-off from 95% to 50% in 0.6 cm, external dose fall-off from 50% to 0% in 10 cm, and OARs EUD = 0 with a = 1 for liver, kidneys and stomach, a = 2 for small bowel and a = 5 for spinal cord. The main changes during postprocessing were hot-spot reduction and improvement of conformality. In this particular site, the planner used the uniformity constraint during the postprocessing which significantly reduced the hot spots in the PTV for MCO plans. If any clinical goal was not met by the MCO plan, it was also added to the postprocessing as needed. The plans were normalized to match PTV coverage. A reduction of OAR doses was found across the board when using MCO as seen in Fig. 3 (bottom panel) while the conformality indices were similar among both groups (Fig. 3 top panel). Two physicians were presented with the 2 plans (double-blinded study) and asked to select one for potential treatment. Both physicians chose the MCO plan in all but one patient. Head and neck The universal set of constraints and objectives for head and neck was similar to those of other sites: we used 3 constraints (PTV min dose 100%, PTV max dose 103% and GT V/CT V min dose 100%) and the objectives were (PTV uniform dose 100%, external dose fall-off from 95% to 50% in 0.6 cm, external dose fall-off from 50% to 0% in 10 cm, for OARs EUD = 0 with a = 1 for most OARs except spinal cord where a larger value of the parameter “a” was used as before. For head and neck the planner included a PTV minimum dose as an objective. This objective can help to fix

97% 100% 100% 95% 99% 100% 100% 100% 97% 100%

0.97 0.99 0.91 0.76 0.70 0.95 0.97 0.99 0.95 0.94

MCO Planner navigated 0.95 0.98 0.98 0.76 0.91 0.95 0.94 0.97 0.93 0.92

Dmax (as percent of prescripon dose) Clinical 108% 106% 104% 108% 110% 107% 106% 109% 108% 105%

MCO Planner navigated 107% 108% 106% 113% 111% 110% 108% 112% 108% 107%

a minimum dose to the PTV during navigation. However, based on the experience with other sites we find that it does not make a significant difference in the resulting plan, although it may simplify the navigation for beginner planners. There was also a maximum dose objective for the body structure. This is a precaution to make sure there are not hot spots outside the PTV. The plans were normalized to have the same coverage and the OAR doses, conformality and PTV heterogeneity were compared between the MCO and DMPO plans. For all patients, all critical structures constraints were met with both techniques, but the low dose portion of the DVH for most of the critical structures were much lower for MCO plans compared to DMPO. Small differences between PTV hot spots and CI for MCO and DMPO based VMAT plans were not statistically significant. The only significant difference was seen in the mean dose of both parotids, where MCO generated significantly lesser doses as compared to DMPO (p value ≤ 0.05, Fig. 4). Mean dose for MCO generated VMAT plans for the Rt. parotids were 32.9 Gy ± 8.6 Gy versus 35.5 Gy ± 6.6 Gy for DMPO (p value = 0.04). Mean doses for MCO generated VMAT plans for the Lt. parotids were 29.7 Gy ± 6.7 Gy while DMPO were: 33.5 Gy ± 6.7 Gy (p value = 0.004).

Discussion In the past few years, there have been a series of studies investigating MCO for treatment planning. Several studies addressed fixed-beam IMRT plans and limited anatomical sites.3-7 IMRT MCO planning in Raystation is significantly different than VMAT MCO planning because the IMRT algorithm for the Pareto surface creation is a segment-based optimization technique, while the equivalent VMAT algorithm is a fluence-based optimization. However, it is encouraging that all these studies have found MCO as a promising and valid optimization technique. In terms of MCO and VMAT, Buschmann et al.9 compared prostate + pelvic lymph nodes in 2 dose levels to tomotherapy plans and found that MCO produced comparable plans with reduced planning hands-on time and similar difficulty in the challenging SIB plans. Other MCO VMAT studies were limited to one body site and a small number of patients (Young et al.11 , Ghandour et al.12 ) but found encouraging results for MCO VMAT planning. Our study on MCO VMAT plans is comprehensive in terms of the number of sites (5 total) and number of patients (66 plans). Our emphasis is the standardization of objectives and constraints for each anatomical site for VMAT plans. For the first site stud-

JID: MDO 6

ARTICLE IN PRESS

[mUS5Gb;May 15, 2019;4:20]

M. Guerrero, Z. Fellows and P. Mohindra et al. / Medical Dosimetry xxx (xxxx) xxx

Fig. 3. Pancreas plans dosimetric comparison between MCO plans (blue) and DMPO plans (green). Top panel: Conformity indexes: Conformity indexes are comparable. Bottom panel: OARs mean doses (in Gy) or volumetric constraints (in percentage). Mean doses and volumetric constraints are either comparable or better for the MCO plans. (Color version of figure is available online.)

ied (prostate plus proximal seminal vesicles), we considered 20 plans, which allowed us to get familiar with the particulars of the software. Initially, several variations in sets of objectives and constraints were tried before coming to a universal set. For the additional anatomical sites, we considered 10 to 15 patients as a sufficient quantity, given that the universal set was an extension of the one used for prostate. For all anatomical sites the patients were selected randomly from our database. However, variations in anatomy and geometry were verified by calculating the overlap between targets and OARs in prostate plus seminal vesicles and whole pelvis, target size and lung volume in the lung cases and PTV volumes for pancreas and head and neck. In order to derive the standard set of objectives and constraints we used the following process: we started with the vendor’s recommendations and tried different variations. We first noticed that it is not necessary to include OARs that are reasonably far from the target because the dose fall-off constraint. This observation reduces the number of objectives and therefore the time needed for the Pareto plan generation and simplifies the navigation process. We used a standard distance for the dose fall-off, so that we do not need to determine it on a patient by patient basis as the vendor recommends. We tried different values of the parameter “a” in EUD for the OAR ob-

jectives. We increased the minimum dose and reduced the maximum dose in the target constraints to reduce the hot spots and improve the target coverage. The ultimate test was to use the same set and the same philosophy for all the sites. While it is possible that these sets of objectives and constraints are not unique and that other variations of objectives and parameters can be used with similar results, our study demonstrates that for a variety of anatomies and patient geometries it is possible to use a class solution for VMAT plans and obtain plans that are comparable or better than the clinical plans in terms of biological, dosimetric, conformality, and homogeneity measures. During the postprocessing phase, we made sure that the MCO plan had as little modifications as possible, by minimizing the weights of the additional objectives (typically additional objectives were reduction of hot spots and improvement conformality). The development of a universal set of objectives and constraints can significantly streamline the planning process and potentially reduce the variation in the plan quality among dosimetrists or planners with different experience levels. Currently, there are several efforts to generate automatic plans in commercial planning systems, with different approaches. For example, Pinnacle-based Autoplanning has been recently reported for head and neck pa-

JID: MDO

ARTICLE IN PRESS M. Guerrero, Z. Fellows and P. Mohindra et al. / Medical Dosimetry xxx (xxxx) xxx

[mUS5Gb;May 15, 2019;4:20] 7

One limitation of our study is that the plans were developed by only 2 planners, and therefore the use of the universal set of constraints was not tested for many planners with different levels of experience. We also did not systematically investigate plans with more than one level of prescription doses (dose painting), which is a topic of future testing. Applications of MCO to stereotactic body radiation therapy (SBRT) were not included given that we do not typically use inverse planning for our SBRT plans. Other anatomical sites and simultaneous integrated boosts will be topics of future investigation. Conclusions Based on our experience with many anatomical sites and a large number of patient plans, we have found that VMAT MCO plans are comparable to the clinical plans and can be produced with a universal set of objectives and constraints, even for a wide range of geometries and anatomies. Conflicts of Interest None. Fig. 4. Mean parotid dose for MCO (blue) and DMPO (green) for head and neck plans. Top panel: Right parotid. Bottom panel: Left parotid. A small reduction of parotid doses was seen with MCO plans. Mean dose for MCO generated VMAT plans for the rt. parotids were 32.9 Gy ± 8.6 Gy vs 35.5 Gy ± 6.6 Gy for DMPO (p value = 0.04). Mean doses for MCO generated VMAT plans for the lt. parotids were 29.7 Gy ± 6.7 Gy while DMPO were: 33.5 Gy ± 6.7 Gy (p value = 0.004). (Color version of figure is available online.)

tients.17 Pinnacle autoplanning does not guarantee treatment plans that are part of the Pareto surface. In an effort to overcome this limitation, a combination of Pinnacle Autoplanning and Raystation MCO was recently reported for nasopharyngeal carcinoma.18 Eclipse Rapid plan is based on dosimetric parameters generated in a past patients database. MCO does not rely on past patient information and provides a powerful tool to develop patient specific optimal plans using standardized objectives. It has the inherent advantage of individualizing the plan based on the Pareto plan navigation, given that the Pareto surface is specific to each patient anatomy and to the beam set selected. Our work provides strong evidence to support this statement and shows that MCO can improve the planning efficiency. RayStation version7.0 or later will provide the capability to run MCO Pareto plans for a group of patients with the same objectives and constraints as a batch job that can be run overnight. This is a good example of a significant potential improvement to planning efficiency and our work supports the basis for such implementation. Our study also addresses the possibility of physician involvement in the treatment process. Mueller et al.3 performed a study in which a template of planning objectives was created, and physicians were invited to navigate the Pareto plans, and then the dosimetrists finalized the plan conversion and postprocessing. They found that for the prostate and brain cases studied, the physician navigated plans were comparable to the clinical ones overall but often with different dosimetric trade-offs. Our study shows a similar conclusion for VMAT plans of advanced lung cases. The physicians involved in the study were enthusiastic and felt empowered by the ability to control the trade-offs of the individual plans. As MCO becomes more available in our field, it is possible that physician involvement in the treatment planning process can increase and become an asset to improve patient outcomes.

References 1. Halabi, T.; Craft, D.; Bortfeld, T. Dose-volume objectives in multi-criteria optimization. Phys. Med. Biol. 51:3809–18; 2006. 2. Bokrantz, R. Multicriteria optimization for volumetric-modulated arc therapy by decomposition into a fluence-based relaxation and a segment weight-based restriction. Med. Phys. 11:6712–25; 2012. 3. Mueller, B.S.; Shih, H.A.; Efstathiou, J.A.; et al. Multicriteria plan optimization in the hands of physicians: A pilot study in prostate cancer and brain tumors. Radiat. Oncol. 12:168–78; 2017. 4. Kamran, S.C.; Mueller, B.S.; Paetzold, P.; et al. Multi-criteria optimization achieves superior normal tissue sparing in a planning study of intensity-modulated radiation therapy for RTOG 1308-eligible non-small cell lung cancer patients. Radiother. Oncol. 118:515–20; 2016. 5. Kierkels, R.G.; Visser, R.; Bijl, H.P.; et al. Multicriteria optimization enables less experienced planners to efficiently produce high quality treatment plans in head and neck cancer radiotherapy. Radiat. Oncol. 10:87–94; 2015. 6. McGarry, C.K.; Bokrantz, R.; O’Sullivan, J.M.; et al. Advantages and limitations of navigation-based multicriteria optimization (MCO) for localized prostate cancer IMRT planning. Med. Dosim. 39:205–11; 2014. 7. Wala, J.; Craft, D.; Paly, J.; et al. Maximizing dosimetric benefits of IMRT in the treatment of localized prostate cancer through multicriteria optimization planning. Med. Dosim. 38:298–303; 2013. 8. Ghandour, S.; Cosinschi, A.; Mazouni, Z.; et al. Optimization of stereotactic body radiotherapy treatment planning using a multicriteria optimization algorithm. Zeitschrift für Medizinische Physik 26:362–70; 2016. 9. Buschmann, M.; Seppenwoolde, Y.; Wiezorek, T.; et al. Advanced optimization methods for whole pelvic and local prostate external beam therapy. Phys. Med. 32:465–73; 2016. 10. Chen, H.X.; Craft, D.L.; Gierga, D.P. Multicriteria optimization informed VMAT planning. Med. Dosim. 39:64–73; 2014. 11. Young M.R., Craft D.L., Colbert C.M. et al. Volumetric-modulated arc therapy using multicriteria optimization for body and extremity sarcoma J. Appl. Clin. Med. Phys. 17 283–91. 12. Ghandour, S.; Matzinger, O.; Pachoud, M. Volumetric-modulated arc therapy planning using multicriteria optimization for localized prostate cancer. J. Appl. Clin. Med. Phys. 16:258–69; 2015. 13. Wang, J.Z.; Guerrero, M.; Li, X.A. How low is the alpha/beta ratio for prostate cancer? Int. J. Radiat. Oncol. Biol. Phys. 55:190–203; 2003. 14. Michalski, J.M.; Gay, H.; Jackson, A.; et al. Radiation dose-volume effects in radiation-induced rectal injury. Int. J. Radiat. Oncol. Biol. Phys. 73:S123–9; 2010. 15. Burman, C.; Kutcher, G.J.; Emami, B.; et al. Fitting of normal tissue tolerance data to an analytic function. Int. J. Radiat. Oncol. Biol. Phys. 21:123–35; 1991. 16. Marks, L.B.; Bentzen, S.M.; Deasy, J.O.; et al. Radiation dose-volume effects in the lung. Int. J. Radiat. Oncol. Biol. Phys. 76:S70–6; 2010. 17. Gintz, D.; Latifi, K.; Caudell, I.; et al. Initial evaluation of automated treatment planning software. J. Appl. Clin. Med. Phys. 17(3):331–46; 2016. 18. Wang, J.; Chen, Z.; Li, W.; et al. “A new strategy for volumetric-modulated arc therapy planning using AutoPlanning based multicriteria optimization for nasopharyngeal carcinoma. Radiat. Oncol. 13(94); 2018.