ESTRO 38 Abstract book

S91 ESTRO 38

OC-0180 Towards a comprehensive automatic planning with deep neural networks: dose prediction for lung IMRT A.M. Barragán Montero 1,2 , D. Nguyen 2 , W. Lu 2 , M. Lin 3 , X. Geets 1 , E. Sterpin 1 , S. Jiang 2 1 Université Catholique de Louvain- Institute of Experimental & Clinical Research, Molecular Imaging- Radiotherapy and Oncology MIRO, Brussels, Belgium ; 2 UT Southwestern Medical Center, Medical Artificial Intelligence and Automation MAIA, Dallas, USA ; 3 UT Southwestern Medical Center, Radiation Oncology, Dallas, USA Purpose or Objective Most automatic planning strategies, such as knowledge- based planning, are based on the prediction of one- dimensional dose volume objectives that are insensitive to the spatial features of the dose distribution, hampering to achieve more individualized treatments. In addition, they rely on consistent beam and dosimetric characteristics among the training database, and often fail in cases where the beam configuration or the anatomy is significantly different. Our model uses neural networks to directly predict 3D dose distributions, combining both dosimetric and anatomical information in order to increase the robustness against heterogeneous patient populations. Material and Methods The model combines two recent deep learning architectures, UNet and DenseNet, to learn from previous clinical plans. The UNet is a type of convolutional neural network able to include local and global features from the input images. It was modified with the densely connected convolutional architecture used in DenseNet, to achieve a more efficient feature propagation. We used several input channels to include anatomical information from delineated contours (PTV and OARs, 9 channels) and per- beam dosimetric data (1 channel). A set of 129 lung cancer patients treated with IMRT, with heterogeneous beam configuration (4 to 9 beams) and orientation, was used for training/validation (100 patients) and testing (29 patients). Mean squared error was used as objective loss function. The stability of the model was evaluated by using a 5-fold cross-validation approach, where the model was randomly initialized, trained with 80 patients, and validated with the remaining 20 patients, using a different training/validation combination for each fold. The accuracy of the model was evaluated by comparing the mean dose (Dmean) and other relevant metrics for clinical practice in the predicted and real doses. Results Figure 1 presents the average absolute error and its standard deviation (SD) on Dmean for the target and OARs for cross-validation (Figure 1.a, average prediction on the validation set for all 5 folds), and testing (Figure 1.b, average prediction on the test set for all 5 folds). The error on Dmean was below 2.5% of the dose prescription for all considered organs, in both cross-validation and testing. Figure 1.c and 1.d show the mostly overlapping DVHs for one of the test patients and the dose at the center of the target, respectively. Table 1 reports some relevant DVH metrics, most of them below 2%, except D2 for esophagus (> 4%) and spinal cord (> 5%). The training time was about 10h and the time employed to predict the 3D dose for a new patient was around 12 s.

Conclusion The proposed architecture was able to learn from a very heterogeneous database in dosimetric terms, and generated accurate 3D dose distributions that can be later used as voxel-wise objective to create patient-specific treatment plans. This represents an important step towards an easier and more robust implementation of automatic planning techniques. OC-0181 Prostate auto-planning in clinical practice: evaluation of plan acceptance and manual adaptations R. Van Der Bel 1 , D.A. Eekhout 1 , G.H. Wortel 1 , G. Van Der Veen 1 , R.H. Harmsen 1 , F.J. Pos 1 , T.M. Janssen 1 , E.M.F. Damen 1 1 The Netherlands Cancer Institute, Radiotion Oncology, Amsterdam, The Netherlands Purpose or Objective Implementation of auto-planning techniques may improve plan quality and consistency in plans and provide a possible time gain. In our clinic we have implemented auto-planning for prostate in 2016. In this retrospective study we aim to evaluate the effectiveness of prostate auto-planning. Therefore, we determined the acceptance rate of auto-plans and investigated the manual alterations made, their effect on the dose distribution and the clinical relevance of these interventions. Material and Methods Prostate auto-planning is performed using Pinnacle Auto- Planner (Pinnacle 9.10, Philips, Fitchburg, USA) in combination with an in-house post-script for fine tuning. A PTV coverage V 95% of 99% was strived for. For 177 prostate cancer patients, irradiated at two dose levels with 35x2/2.2Gy between January 2017 and July 2018, DVH data of both the auto-plan (including post-script) as well as the clinically used plan were available for analysis.