ESTRO 36 Abstract Book

S126 ESTRO 36 2017 _______________________________________________________________________________________________

Purpose or Objective In order to develop a more efficient workflow and to achieve RTT-independent plans for a portion of breast cancer treatments (whole breast radiotherapy with optional boost irradiation), a fully automated planning process has been clinically introduced. The process offers a number of alternative treatment plans to the RTT, who can choose to either select the optimal clinical plan, or improve one of the candidate plans to a clinically acceptable level. We investigated the acceptance rate of automatically created plans and the motivations for rejection and/or adaptation of these plans. Material and Methods From November 2015 to September 2016 657 treatments have been planned using the automatic procedure. The plans consist of medial and lateral tangential beams, with an optional IMRT beam to deliver a boost dose to the primary tumour (bed) (Fig.1). The tangential beams consist of an open segment, delivering 75% of the dose, and a limited number of IMRT segments, delivering 25% of the dose. The open segments target the PTV (blocks on the heart when applicable), but are open outside the patient contour to allow for anatomical changes. 6MV and 10MV medial and lateral beams are offered. A heart clearance choice of 0 mm or 5 mm is also offered. This results in a total of 4 (right-sided breasts) or 8 (left-sided breasts) candidate plans. The in-house automatic planning software (FAST), controlling the Pinnacle 3 TPS, generates the plans and corresponding dose distributions automatically without any intervention from the RTT. This procedure commences as soon as the radiation oncologist has delineated the target volume.

Conclusion Considering that in close to 50% of all cases one of these plans was accepted for clinical use, a significant time saving is apparent (pre-clinical evaluation predicted an acceptance rate of 60%). This saving is estimated to be 1000 hours/year based on the projected 800 patients/year. In 45% of the cases in which an automatic plan was not chosen, only minor modifications were made to the plan, still resulting in a time reduction close to 450 hours/year. OC-0253 Machine Learning-Based Enables Data-driven Radiotherapy Treatment Planning Decision Support. G. Valdes 1 , L. Wojtowicz 2 , A.J. Pattison 3 , C. Carpenter 4 , C. Simone 2 , A. Lin 2 , T. Solberg 1 1 University of California UCSF, Radiation Oncology, San Francisco CA, USA 2 University of Pennsylvania, Radiation Oncology, Philadelphia, USA Purpose or Objective Due to the complexity of dose deposition and variety of treatment delivery technology, plan outcomes remain non-intuitive. The ability to predict radiotherapy treatment discrete plan outcomes before planning enables the clinician to more accurately guide therapy decisions before engaging in the time-consuming plan creation process. We demonstrate the ability to accurately predict plans for lung photon and for head and neck proton and photon therapy using machine learning. Material and Methods 100 patients with early stage lung cancer who received stereotactic body radiation therapy (SBRT) and 36 patients with head and neck cancer who received postoperative proton radiotherapy were identified. Each head and neck patient also had corresponding photon-based volumetric modulated arc therapy plan (VMAT). DICOM-RT datasets were processed using commercial plan-prediction software (QuickMatch™, Siris Medical Inc.) to predict dose to Planning Target Volumes (PTVs) and Organs at Risk (OARs). For lung SBRT plans, we predicted doses to the lung, cord, brachial plexus, skin, esophagus, heart, great vessels, trachea, rib, chestwall, and PTV. For H&N plans, we predicted doses to the: parotid, submandibular[AL1] , brain stem, cord, optic nerve, mandible, constrictors, esophagus, oral cavity, and larynx. We computed error metrics and established guidelines for dataset size. In addition, several deliverable plans were created to demonstrate the advantages of predictive modeling. Results We were able to effectively predict dose distributions and dataset sizes required for desired accuracy errors for the OARs and PTVs for both photon and proton treatments. For lung SBRT plans, a dataset size of at least 69 plans resulted in all mean errors below 2.5Gy. For photon H&N plans, a dataset size of at least 121 plans resulted in all mean 3 Siris Medical, CTO, Mountain View, USA 4 Siris Medical, CEO, Mountain View, USA

Results Automatically generated plans were selected by the RTT without any adaptation in 54% of non-boost treatments and in 41% of boost treatments (Fig.2). For both classes, reasons for not selecting an automatically generated plan were very similar: in 45% of the cases, optimization goals were modified in order to change trade-offs between PTV coverage and OAR doses (at the discretion of the RTT). Our study found that in most of these cases the plan only marginally differed from the automatic plan. In another 40% of the cases a new plan was manually created, e.g. to replace the automatic tangential beam set-up with a more favourable set-up. The final 15% comprised cases in which automatic delineation was erroneous or other technical issues. In the majority of left-sided cases, the 5mm heart clearance plan was preferred. The variety in chosen beam energies is related to patient geometry.