Abstract Book

S1025

ESTRO 37

Figure 1. The comparison of dose distribution of CP with RP for a patient, who has 3 different dose levels (PTV65, PTV60, PTV54). Conclusion An RP model using generated objectives was able to meet all clinical constraints and improve parotid sparing but at the expense of target homogeneity. Therefore, further work is required to modify the model in order to improve the PTVs homogeneity. EP-1893 Automated planning through explicit optimization of plan quality L. Engberg 1 , K. Eriksson 1 , A. Forsgren 2 1 Raysearch Laboratories AB, Research department, Stockholm, Sweden 2 KTH Royal Institute of Technology, Department of Mathematics, Stockholm, Sweden Purpose or Objective In this study, we address methodological shortcomings in the conventional penalty-based objective functions, and argue that resolving these issues has a similar potential as other automating methods to streamline the planning process. Our hypothesis is that any tractable formulation of objective functions with a clear connection to plan quality indices qualifies as being trial-and-error eliminating and thereby automating. The potential then depends on the ability to generate plans of at least equivalent quality as the conventional formulation. Initial investigation of such potential of a specific formulation was the subject of our recent preliminary study. The purpose of the current study is to examine this formulation in a more clinical setting, where, e.g., deliverability of plans is accounted for. Material and Methods The plan quality indices involved are dose levels in the dose-volume histogram (doses-at-volume). To achieve a clear connection to plan quality, the proposed objective functions abandon the penalty-function paradigm. The intractable dose-at-volume functions are explicitly approximated using mean-tail-dose objective functions, resulting in a convex optimization problem. The formulation is extended with deliverability constraints for sliding window (DMLC) delivery, which preserves convexity. Results We examined the plan quality obtained among Pareto optimal DMLC treatment plans, which result from using the proposed objective functions in a multicriteria optimization (MCO) framework. As a reference, DMLC treatment plans were generated using the penalty- function based MCO module in RayStation (RaySearch Laboratories). Two patient cases were considered. For each case, three plan quality indices were chosen as tradeoff objectives, and the remaining clinical criteria were imposed using hard constraints. After a final accurate dose computation step, all generated plans were evaluated based on the three plan quality indices. The distribution of these indices is graphically visualized in a three-dimensional coordinate system.

Conclusion This study allowed us to simplify our practices with the use of the NTO as a replacement for all optimization volumes (rings). Treatment plans obtained with PO13 allow a better saving of OAR with an equal or a better coverage of PTV. EP-1892 Clinical validation of Rapid Plan Head and Neck Model S.C. Lee 1 , C.P. South 1 , E.J. Adams 1 1 St. Luke's Cancer Centre Royal Surrey County Hosp, Medical Physics, Guildford, United Kingdom Purpose or Objective Treatment planning for head & neck (H&N) intensity modulated radiotherapy (IMRT) or volumetric modulated arc therapy (VMAT) is time consuming and requires significant resource, particularly when a significant number of patients require adaptive planning during treatment. Therefore, developing automated or knowledge-based planning techniques for H&N cancer patients is important. The purpose of this study was to validate the initial RapidPlan (RP) H&N model at our institution. Material and Methods Varian Eclipse v13.7 was used for this study. An RP model was created using 81 randomly-selected H&N patients initially planned with 1-3 dose levels using IMRT or VMAT, and including both unilateral and bilateral tumours for a variety of diagnoses. Upper and lower objectives were added for targets, and line objectives for organs at risk (OARs). In both cases the default RP settings were used, allowing RP to automatically generate priorities. Local settings were used for the Normal Tissue Objective (NTO). Once the model had been created, it was tested for 11 previously planned patients, covering the same range of treatment techniques and diagnoses as were included in the model. MU constraints were used as they would be clinically but there was no additional user input for the optimizer when using RP. Plans were evaluated against local plan acceptance criteria. Results RP showed a statistically significant reduction in mean parotid doses. All plans met local criteria for PTV coverage and homogeneity, but RP showed statistically significant reduced homogeneity compared to the clinical plans (CP), as shown in Fig. 1. Spinal cord doses were similar between RP and CP and were within tolerance in all cases. Generally, Monitor Units (MU) of RP were higher than the CP, but they were all within the acceptable range. Further modifications were made to the RP model to improve target homogeneity and initial results indicate that it is possible to achieve this whilst still retaining much of the improved parotid sparing.