ESTRO 36 Abstract Book

S237 ESTRO 36 _______________________________________________________________________________________________

Conclusion A fully automated planning system has been developed that allows configuration by expert treatment planners and oncologists. The evaluative study presented shows high quality plans can be produced with no user input, following the initial site-specific configuration process. This simple process allows high-quality automated plans to be produced for new treatment sites in an efficient manner. OC-0447 CyberArc: a 4π-arc optimization algorithm for CyberKnife V. Kearney 1 , J. Cheung 1 , T. Solberg 1 , C. McGuinness 1 1 University of California UCSF, department of radiation oncology, San Francisco CA, USA Purpose or Objective To demonstrate the feasibility of 4π-arc radiotherapy using CyberKnife for decreased treatment delivery times. Material and Methods A novel 4π-arc optimization algorithm (CyberArc) was developed and evaluated in 4 prostate and 2 brain cancer patients previously treated with CyberKnife using Iris collimation. CyberArc was designed for continuous radiation delivery between beam and node positions using 4π treatment geometry. During beam delivery, the isocenter and Iris collimator diameter are allowed to freely move within machine tolerances. For comparison purposes, new plans were generated using the same total number of beams and range of Iris collimation. Dose calculation was based on the MatRad pencil beam algorithm, modified using the machine commissioning data to fit the CyberKnife flattening filter free beam profiles and percent depth doses. An initial 4π library of beam coordinates is cast over the allowed delivery space. A constrained subplex-based optimization algorithm then selects from an initial library of 6 node positions for each beam coordinate using a 5mm x 5mm fluence map resolution to obtain the first set of beam/node/collimator configurations. A preliminary monitor unit calculation is performed, and beam/node/collimator positions that fall under a threshold are discarded. A 3D traveling salesman problem is solved using a genetic algorithm to obtain the paths between beams ( Figure 1) . From the second set of beam/node/collimator positions, intermediate beam/node/collimator coordinates are calculated along the path between neighboring coordinates using cubic interpolation. A third set of continuous intermediate beam/node/collimator doses are calculated every 2° along the arc path with a 2mm x 2mm fluence map resolution. MUs are calculated for each beam/node/collimagor position using an L-BFGS-B optimization engine. All plans were normalized to the 70% dose volume of the PTV for comparison.

Vesicles (PSV), Prostate & Pelvic Nodes (PPN) and Head & Neck (HN). Material and Methods A fully automated VMAT planning system has been developed using the scripting functionality of RayStation (RaySearch Laboratories, Stockholm, Sweden). For each treatment site, a set of clinical priorities is determined as a list of constraints and ‘tradeoffs’. The system is designed to ensure constraints are met while optimization of the prioritized tradeoffs is guided by an a priori calibration process. This process involves optimizing for the first priority trade-off with a range of objective weights. A GUI allows the user to navigate through these plans using the DVH and dose-distribution to determine the optimal weight. This weight is then stored and the process is repeated for the next priority. When all trade-offs have been optimized, the whole process is repeated to refine the set of objective weights. As the underlying automated-planning algorithm tailors the base plan to individual patient anatomy, a single set of configuration data can be used for all patients of a given site and plan can be generated with no user interaction. Ten patients of each configured site were planned using the automated system and compared against the clinically approved manual plan. Quantitative comparisons were made using relevant DVH metrics and qualitative comparisons of dose distributions. Results A selection of the DVH metrics, averaged across all patients in each site, is listed in Table 1. A set of representative dose distributions is provided in Figure 1. The tabulated data shows that the automated plans tend towards improved OAR sparing. For PSV and PPN, this was at the expense of target coverage. For HN plans, the automated plans improved coverage while also reducing OAR doses. Visual comparisons of dose distributions showed that, for all three sites, the automated plans were of equal or better quality relative to manually optimized plans. All plans met local clinical DVH constraints and were deemed to be clinically acceptable.

Figure 1. The set of final beam positions and their corresponding paths for prostate patient 3. Results Among the six patients analyzed, the average difference in PTV min dose, max dose, and V95 was 2.47% ± 2.13%, 4.11% ± 2.62%, and 1.63% ± 3.01% respectively. The average conformity index (CI) was 1.09 ± 0.07 for the brain patients and 1.12 ± 0.09 for the prostate patients. Figure 2 shows the plan comparison DVHs for a prostate and brain