ESTRO 2020 Abstract Book

S113 ESTRO 2020

OC-0223 Prediction of proton dose distributions with deep learning for automatic treatment planning A.M. Barragán Montero 1 , M. Huet 1 , S. Teruel Rivas 1 , K. Souris 1 , D. Nguyen 2 , S. Jiang 2 , J. Lee 1 , E. Sterpin 1 1 Université Catholique de Louvain- Institute of Experimental & Clinical Research, Molecular Imaging- Radiotherapy and Oncology MIRO, Brussels, Belgium ; 2 University of Southwestern Texas, Medical Artificial Intelligence and Automation Laboratory MAIA, Dallas, USA Purpose or Objective Deep neural networks (DNN) are becoming a popular tool for automatic treatment planning in radiation therapy. Recently, several groups have reported the excellent performance of DNN to predict three-dimensional dose distributions for VMAT and IMRT treatments. The output dose can later be used to automatically generate a treatment plan, removing all human intervention and associated variability, which ensures high plan quality. Proton therapy could greatly benefit from the power of DNN, especially nowadays, when the centers with accumulated clinical experience in proton planning are very few and the number of new centers is growing. The present work is the first to investigate the use of DNN for proton dose prediction for head and neck (H&N) cancer. Material and Methods The model combines two deep learning architectures, UNet and DenseNet. The UNet, a type of convolutional neural network able to include local and global features from the input images, was modified to achieve a more efficient feature propagation, by mimicking the DenseNet architecture. We compared two models that used different input data in separated channels: 1) Contours (C), which uses binary masks from the targets (CTVlow, 50Gy/54.25Gy, and CTVhigh, 70Gy) and organs, and 2) Contours plus CT (C+CT), which uses both the binary masks and the CT image. A set of 62 patients treated with pencil beam scanning, with the same beam configuration (4 beams), was used for training (50 patients) and testing (12 patients). All plans were generated in RayStation v8a (RaySearch Laboratories, Sweden), using robust optimization with 4 mm for setup errors and 3% for range errors. The stability of the model was evaluated by using a 5-fold cross-validation, where the model was trained 5 times with 40 patients and the objective function was evaluated in the remaining 10 patients, which were alternated along the 5 folds. The accuracy of the model was evaluated by comparing the mean dose and other metrics for clinical practice in the predicted and real doses, as well as the dice coefficient for the isodose lines. Results The results for the average absolute error and its standard deviation (SD) on the mean dose for the CTVs and organs are presented in Figure 1.a. for the test set (average prediction for all 5 folds), for the C and C+CT models. Averaged over all organs, the prediction error was equal to 6.67±4.83% (mean±SD) for the C model and 3.13±1.90% for the C+CT model. The dice coefficient (Figure 1.b) confirmed also an improved spatial accuracy when including the CT as input, although the high dose region (70% to 90% isodoses) remains difficult to predict. The training time was about 8h and the inference time was around 20 s.

ImageNet image dataset (1.2 million natural images of 1000 object categories). We dropped the fully-connected layers from the VGG-16 and added a new fully-connected neural network. The inputs for the CNN were a 2D image of the volumes contoured in the CT. We only retained the geometrical information of every CT. The outputs were the corresponding dose maps. Rectum and bladder DVHs were computed for each patient summing up all the dose- volume information in every slice. To ensure the quality of the data, we selected all potential outliers and proceeded to re-optimize them. The already trained CNN was tested using the second test of patients (25). All patients were re-optimized by the same operator unfamiliar with the results of the prediction. A confusion matrix was used to report the number of false positives, false negatives, true positives, and true negatives. Results Figure I shows the clinically approved, replanned, and predicted bladder and rectum DVHs for three representative cases. The results demonstrated that for most cases the actual DVH, either clinically approved or replanned, was within the DVH predicted by our model.

Our algorithm achieved 100% and 81.25% of true positive and true negative prediction, respectively. We have an overall accuracy of 87.5%, a misclassification rate of 27% and a precision of 100%. Conclusion We have successfully developed a model for reliable prediction of rectum and bladder DVH in prostate patients treated with a VMAT technique by applying a CNN pre- trained previously on a set of natural images. Our model demonstrates excellent performance, and its results validate the ability not only to detect sub-optimal plans retrospectively but also to predict achievable DVHs as a reliable guide and an optimal target for treatment planning. To our knowledge, this is the first attempt to apply transfer learning to the prediction of DVHs that accounts for the ground truth problem while training the model.