ESTRO 38 Abstract book

S987 ESTRO 38

determine the optimal trade off between PTV coverage and OAR doses. RP and MCO were applied to 10 Pancreas planning cases. Previous work, Houston P., Laverick N. Planning comparison of knowledge based planning and multi- criteria optimisation for VMAT pancreas planning. ESTRO 2017 compared the plans with regards to plan quality. The treatment plans were here investigated for robustness to patient set-up by shifting the patient and re-measuring dose volume statistics. Plans were shifted ±0.3cm and ±0.5cm in the x, y and z directions and worst case dose values recorded for each PTV and OAR. Results

Biomedical Physics, Warsaw, Poland ; 4 Poznan University of Technology, Faculty of Technical Physics, Poznan, Poland Purpose or Objective Optimal values of treatment plans parameters for CyberKnife (CK) prostate’s treatment, could varied from patients-to-patients. An artificial neural network (ANN) could be powerful tool for a radiotherapy technique for mapping individual patient anatomy and other treatment plan’s parameters. ANN could be support tools for prediction optimal parameters for CK’s treatment. This study was performed to investigate the feasibility of ANN in prediction CK treatment planning parameters for prostate cancer. Material and Methods A set of retrospective clinical data from 200 patients with prostate cancer were used to build and train the ANNs to predict radiotherapy treatment plan parameters. 22 new patients were used to test the models. Inputs were chosen from general parameters such as: prescription dose, volumes of PTV and six OARs and geometry parameters defining distance between mass centers of PTV and OARs, respectively. A fully connected ANN was used with two hidden layers of 150 neurons, 18 input and 14 output neurons. As an activation function the Rectified Linear Units (ReLU) were used, with 0.4 dropout rate in order to avoid over fitting. The net was trained in 2000 epochs, with mean squared error as a loss function. Results Errors of ANN's prediction of the plans parameters for patients from test set were evaluated in terms of mean absolute value of differences (mean(abs(diff))) between predicted and original values and root-mean-squares (rms (diff)) of these values. Table 1 presents mentioned errors for chosen parameters (prescribed isodose (%), number (#) of nodes, beams and collimators, estimated treatment time (minutes), PTV CI and PTV Coverage (%)) in comparison with original mean values (mean test) of parameters in the test set. Conclusion This investigation was preliminary and lead us to confirm the feasibility of artificial neural network in prediction CyberKnife treatment planning parameters for prostate cancer. The errors in comparison with mean values of parameters in test set were not high, but not yet good enough to be useful for clinical purposes: the errors were comparable with standard deviations of parameters of test set. Moreover several steps have to be undertaken in future. In order to predict features of treatment planning parameters, the crucial information is in geometry dependencies of PTV and OARs, which could be found specially tailored parameters. We suppose it could be obtained by convolutional neural network with bigger amount of training examples. EP-1821 Fast Robust Optimization using a Patient- Specific Scenario Selection Methodology G. Buti 1 , K. Souris 1 , J.A. Lee 1 , E. Sterpin 2 1 Université Catholique de Louvain, Molecular Imaging-

Figure1: Comparison of PTV and GTV coverage when shifts of 3mm and 5mm from planned position are applied. MCO v15.5 and RP v15.5 plans show comparable robustness to 0.3cm shifts. MCO v15.5 gives higher PTV D98% coverage after a 0.5cm shift than RP v15.5. GTV coverage is not significantly affected by either a 0.3cm or 0.5cm shift, as expected due to GTV to PTV margins of 1cm.

Figure 2: D50% and D2cc for ipsi-lateral kidney when shifts of 3mm and 5mm from planned positon are applied. The improvements in plan quality with MCO conferred a significant advantage in the robustness of OAR dose to shifts in patient position. Average ipsilateral kidney D50% was lower with MCO v15.5 shifted 0.5cm when compared to unshifted RP plans, 837cGy and 1200cGy respectively. Duodenum, which is often the dose limiting organ in pancreas planning, did not display a significant change in robustness of D33% or D2cc. Conclusion MCO treatment planning does not degrade the robustness of PTV coverage to shifts in patient position and may in some cases improve that robustness. OAR doses are either comparable or improved when planned with MCO and possible shifts in patient are accounted for. This general robustness to changes in patient shifts indicates that current margins, imaging and set-up protocols are appropriate for MCO planned pancreatic radiotherapy treatments. EP-1820 Preliminary results of using artificial neural networks for prediction CK planning parameters A. Skrobala 1,2 , J. Ginter 3 , B. Pawalowski 1,4 , M. Skowron 3 , M. Adamczyk 1 , A. Jodda 1 , J. Litoborska 1 1 Greater Poland Cancer Centre, Department of Medical Physics, Poznan, Poland ; 2 Poznan University of Medical Sciences, Department of Electroradiology, Poznan, Poland ; 3 University of Warsaw, Department of

Radiotherapy & Oncology, Brussels, Belgium ; 2 Katholieke Universiteit Leuven, Experimental Radiotherapy & Oncology, Leuven, Belgium