ESTRO 38 Abstract book

S1170 ESTRO 38

dose inaccuracy for plan potentially affected by overmodulation issue. Material and Methods 70 stereotactic RapidArc plans were planned in Eclipse v. 13.7 (Varian Medical System) and delivered on TrueBeam stx. Delivery Quality Assurances (DQA) of all plans were performed: delivered doses were measured with PTW SRS1000 chamber array combined with PTW Octavius 4D and compared with calculated dose using PTW Verisoft v.6.1. Each plan was characterized by 14 features extracted from RP-Dicom files describing different aspects of their complexity that could lead to dosimetric disagreements [Carlson et al - Phys Med Biol. 2016]. A subset of 45 plans (training set) was used to train a gaussian process regression model using the features as predictor inputs and the measured dose differences as response output. Training model was achieved with 5 fold cross validation and PCA for predictors reduction. Further open-loop validation was performed with the remaining 25 plans not belonged the training set (validation set). Features extraction, training process and validation analysis were performed using in-house Matlab ( the Mathworks) scripts. To evaluate an application of the model, a dose-painting plan was optimized in several steps with different plan complexity solutions, achieved by using different Max MU objective constraints, then DQA results were compared with prediction model. Results The cross validation analysis on the training set showed a Root Mean Square Error (RMSE) equal to 0.01 and R 2 equal to 0.92. The open loop-validation resulted in a RMSE equal to 0.003 and R 2 equal to 0.96, that indicate a good model generality. With PCA analysis the model was trained only by 3 principal components with an explained variance set to 95%. The Figure 1 shows, for the evaluation set, a correlation plot of the predicted versus actual dose differences normalized to the maximum measured dose. The dose-painting plan was optimized 3 times before to achieve a predicted dose difference within 3% of the maximum dose. In the Table1 prediction and actual values for each dose-painting plan are reported.

the Gun/Target (GT) direction for the MV isocentre for Elekta beam modulator and Elekta Agility linear accelerators respectively.. The lateral and vertical errors were relatively smaller both under 0.8 mm diameter. The mechanical deviation of the couch was within 1.5 mm diameter on both the linacs. There were some outliers which were due to suboptimal beam steering and couch isocentric rotation setup which were picked up during this procedure and were corrected by our in-house engineers. The CBCT imaging isocentre was within 0.7 mm on the beam modulator linac whereas both the CBCT and exactrac isocentres were within 0.5 mm on the agility linac in comparison to the MV isocentres on each machine.

Conclusion Congruence of mechanical, radiation and imaging isocentres of two types of Elekta linear accelerator heads were quantified to ensure safe treatment of stereotactic patients and to assess if the planning target margins used are sufficient for target coverage. EP-2114 Predicting inaccuracy of overmodulated RapidArc plans using Machine Learning model A. Botti 1 , E. Cagni 1 , M. Orlandi 1 , R. Sghedoni 1 , D. Lambertini 1 , A. Barani 1 , V. Bertolini 1 , M. Iori 1 1 Azienda USL - IRCCS di Reggio Emilia, Medical Physics Unit, Reggio Emilia, Italy Purpose or Objective Significant differences between measured and calculated doses of RapidArc plans with very small PTV volume or overmodulated MLC pattern were observed [Fog et al - Phys. Med. Biol. 2011]. The aim of the study was to train a Machine Learning (ML) model to predict the calculated

Conclusion