ESTRO 38 Abstract book

S490 ESTRO 38

layers from the VGG-16 and added a new fully connected neural network to properly transfer learning to our data. Our data was split into training, validation and test sets (98, 26 and 10 patients in each set respectively).To apply this DCNN to our problem, for each patient we extracted the contour information included in the CT and feed this data to the modified VGG-16. The model was trained on single slices of the patient instead of a whole 3D image. DCNN output is a 2D rectum DVH for every slice. Rectum DVHs from our test set were compared to DVHs predicted by our model. An actual plan was considered suboptimal if the mean square error (MSE) of the prediction was greater than the MSE of the validation plus 2 standard deviations. If predicted DVH was better than the actual DVH a replan was carried out to study the cause of this discrepancy. Results DCNN with transfer learning successfully predicted rectum DVHs for our test patients. The predicted DVH was comparable to the actual DVH except for two cases (20%) where the model predicted a substantially better DVH (Fig.1). In these two cases, replanning decreases doses to the rectum without worsening the PTVs or any organ at risk.

2 Hospital de la Santa Creu i Sant Pau, Servei de Radiofísica i Radioprotecció, Barcelona, Spain

Purpose or Objective In breast radiotherapy, the use of dynamic techniques (IMRT, VMAT) usually include a skin flash region outside the body contour to ensure correct irradiation of CTV accounting for uncertainties in position, breathing or possible anatomical changes. In VMAT, the common approach consists in optimizing the VMAT plan on an extended CT including a virtual bolus out of the body contour in the area of the PTV. We present a detailed method to obtain the optimal thickness and HU value assigned to this virtual bolus, regarding plan robustness and dosimetric impact of the strategy. Material and Methods 7 bilateral breast patients treated for whole breast radiotherapy were retrospectively selected for the study.For each patient, we defined 16 modified-CT (CT’) by adding to the original CT (CT0) all combinations of width (0.5cm & 1.0cm) and HU (from -700 to 0, 100 HU steps) of the bolus. We optimized a VMAT plan (RapidArc®,Varian Medical Systems, POv13.5.35, AAAv13.5.35) on the CT0 ( NoAction plan ) and on each CT’ ( CT’ plans ) by using the same optimization objective template per patient and a PTV excluding 5mm of skin .We recalculated CT’ plans to CT0 by fixing MU ( CT0 plans ). For the dosimetric impact of the strategy, CT’ plans and CT0 plans were compared on the basis of doses to PTV (Dmean, D98 and D2) and doses to OARs, particularly to heart (Dmean, V30), lungs (V5) and liver (Dmean). The robustness is assessed by shifting the isocenter 0.5cm and 1.0cm in the breathing direction, and recalculating CT0 plans and NoAction plan by fixing MU ( CT0-shifted plans ). CT0-shifted plans and CT0 plans were compared on the basis of D98 and D2 relative differences on the portion of the PTV in the buildup region (1cm-depth inner margin beneath the skin, PTVskin), which is likely to increase the sensitivity of the analysis. Optimal parameters were those that maximized the plan robustness on CT0 and minimized dosimetric impact of this strategy. Results All PTV dose indexes increased when the re-calculation was made on CT0. Minimum relative Dmean differences were found between -400 and -600 HU depending on the bolus thickness (Fig1). Doses to OARs are not significantly affected. Regarding to robustness, for shifts of 1.0 cm there are significant differences between choosing a bolus thickness of 0.5 cm or 1.0 cm, which is not observed for shifts of 0.5 cm (Fig2). Best robustness is found for -500 UH and 1.0 cm thickness. It is worth noting that when no pseudo skin flash strategy is applied, relative dose differences up to 20% can be found.

Table I shows a dosimetric comparison between the original plan (blue line), the predicted one (orange line) and the re-optimized plan (green line) for the suboptimal patients. Our replan strategy achieved an improvement of 28.7% and 15.0% in V40 for both patients.

Conclusion We successfully used a pre-trained DCNN to predict rectum DVH’s for VMAT prostate patients, allowing for effectively detect suboptimal plans and guiding plan optimization to improve dose distribution. To our knowledge this is one of the first attempts to apply the transfer learning of DCNN to dose prediction in radiotherapy. Thus we demonstrate that transfer learning of DCNN is a feasible solution to the lack of training data in radiation oncology. Future work will include DVH prediction for the remaining organs at risk. PO-0918 Optimal parameters to perform the Pseudo Skin-Flash on VMAT on breast radiotherapy M. Lizondo 1 , A. Latorre-Musoll 2 , N. Espinosa 2 , A. Coral 2 , C. Cases 2 , N. Jornet 2 , P. Carrasco 2 , P. Delgado-Tapia 2 , A. Ruiz-Martinez 2 , I. Valverde-Pascual 2 , M. Barcelo 2 , M. Ribas 2 1 Institut de Recerca Hospital de la Santa Creu i Sant Pau, Servei de Radiofísica i Radioprotecció, Barcelona, Spain ;