ESTRO 2021 Abstract Book

S513

ESTRO 2021

Table 2: Bi-Lateral Hippocampus dose metrics (SD) Gy

Conclusion VMAT-HS, VMAT-MCO and HyperArc result in superior HS compared to VMAT. In patients where only the contra lateral HPC can be spared, VMAT-MCO and HyperArc are superior to VMAT-HS.

OC-0645 Head and neck radiotherapy on the MR-Linac: a multicentre planning challenge on MRIdian- platform M. Chamberlain 1 , J. Krayenbühl 2 , J.E. van Timmeren 1 , L. Wilke 3 , N. Andratschke 1 , S. Tanadini-Lang 1 , H. Garcia Schüler 2 , M. Guckenberger 1 , P. Balermpas 1 1 University Hospital Zurich, Radiation Oncology, Zürich, Switzerland; 2 University Hospital Zürich, Radiation Oncology, Zürich, Switzerland; 3 University Hospital Zurich, Radition Oncology, Zürich, Switzerland Objective: Purpose of this study is to evaluate planning on an MRI-Linac system (MRIdian, Viewray®) for head and neck cancer through comparison of planning approaches of several centers. Materials and Methods: 14 planners using the MRIdian planning system participated in a treatment planning challenge, centrally organized by Viewray®. A pre-contoured real-life case of a patient with a cT4b cN1 p16- negative squamous cell carcinoma of the posterior oropharyngeal wall was contoured and distributed to all participating centers through an online platform. The prescription was on three dose levels with a simultaneous integrated boost (SIB) technique in 35 factions: 70 Gy should be applied to the high-risk volume (PTV1), 59.4 Gy to the intermediate risk volume (PTV2) and 54 Gy to the “elective”, low risk volume (PTV3). Homogeneity, conformity, sparing of organs at risk, and other parameters were evaluated according to ICRU- recommendations anonymously and were then compared between centers. Differences amongst centers were assessed with means of Wilcoxon test. Each plan had to fulfil hard constraints based on DVH parameters and total delivery time. A “plan quality metric “(PQM) was evaluated. The PQM was defined as the sum of 16 submetrics, characterising different DVH goals. Results: The median (range) number of beams used was 17.5 (11 - 60) and the median number of segments 139 (114-208), resulting in a total number of 909.5 MUs (539 - 1474), beam-on time of 1.5 minutes (0.8 - 2.5) and median total delivery time of 18.8 minutes (14.9 - 22.4). The median (range) IMRT-planning experience for head and neck of all users amounted to 7 years (2 - 20) and the Viewray planning experience 1 year (0 - 2). For most dose parameters, the median score of all centers was higher than the threshold that results in maximum score. Six centers achieved the maximum number of points for the OAR dose parameters and none had an unacceptable performance on any of the metrics. Each planner was able to achieve all the requirements except for one with longer delivery time. The number of segments correlated to improved quality metrics and inversely correlated to brainstem-D 0.1cc , and to PTV1-D 0.1cc . Total planning experience inversely correlated to spinal canal dose. Conclusions: MR-Linac-based planning for head and neck cancer is already feasible with good quality. Generally, increased number of segments and increasing planning experience are able to provide better results regarding planning quality without significantly prolonging overall treatment time. OC-0646 Automatic tool for head and neck patient referral based on dose prediction with deep learning M. Huet Dastarac 1 , A.M. Barragán-Montero 1 , S. Michiels 1 , S. Teruel Rivas 1 , E. Sterpin 1 , J. Lee 1 1 UCLouvain, Center of Molecular Imaging, Radiotherapy and Oncology (MIRO), Brussels, Belgium Purpose or Objective Using deep learning [DL] to predict three‐dimensional dose distributions is becoming popular for automatic planning of radiotherapy treatments. In this work, we go one step further and use DL dose prediction combined with NTCP model-based selection to detect patients that would benefit from proton therapy. Unlike the conventional approach, which requires the manual generation of two treatment plans (photons and protons), the presented DL method can perform the patient selection in a matter of seconds. Materials and Methods Two databases of patients with oropharyngeal cancer were used to build the DL models. The first one consisted of 60 patients with proton [PT] plans with Pencil Beam Scanning. The second one had 80 patients with conventional radiotherapy [XT] plan with Tomotherapy. The treatment plans were generated manually in our center by professional planners. Each patient had a CT image and a set of contours for the regions of interest. The overlap between these two databases consisted of 47 patients, from which we selected a test set of 10 patients. The DL model was a UNet-like-architecture with hierarchically dense connections. Two models were