ESTRO 2021 Abstract Book

S500

ESTRO 2021

explained by the sparing of the TVBs.

Conclusion The results of this study suggest the feasibility of the inclusion of LOARs in treatment planning for stage III NSCLC. The LS-plans show a statistically significant decrease in both the TVBs and GVs parameters while maintaining adequate target coverage and cOARs constraints. OC-0635 Breast IMRT or 3D-CRT planning? A decision-making framework using convolutional neural networks P. Gallego Franco 1 , J. Pérez-Alija 1 , J. Chimeno 1 , M. Lizondo 2 , N. Jornet 1 , C. Ansón 1 , A. Latorre 1 , M. Barceló 1 , N. García 1 , H. Vivancos 1 , M. Adrià 1 , A. Ruiz 1 , P. Carrasco 1 , E. Ambroa 3 1 Hospital de la Santa Creu i Sant Pau, Medical Physics, Barcelona, Spain; 2 Consorci Sanitari de Terrassa, Medical Physics Unit, Terrassa, Spain; 3 Consorci Sanitari de Terressa, Medical Physics Unit, Terressa, Spain Purpose or Objective Breast planning is a case where it is troublesome to know whether a patient will benefit from an IMRT planning approach or not. While IMRT may seem better for complex patient anatomies, it could also unnecessarily increase the complexity of the plan without adding any further patient benefits compared with a 3DCRT technique. The purpose of this study was to develop a deep convolutional neural network model (DCNN) for predicting 3D-CRT lung and heart DVH. This predictive model can then be used as a decision support tool for IMRT versus We selected 195 patients with left breast cancer treated with 3D-CRT. We included patients with axillary and supraclavicular lymph nodes but excluded those with an internal mammary nodal chain. We partitioned our set in training, validation, and test (176, 10, and 10 patients, respectively). For the model creation, we trained the DCNN and renormalized all plans to 2 Gy/fraction to consider the different prescribed doses. For our DCNN model, we implemented a transfer learning approach using a pre- trained VGG-16 and replacing its three last layers with a fully connected neural network. Input data was the planning CT contour information. Output was a 2D lung and heart DVH for every slice. All slices were subsequently added up to account for the final 3D OAR DVH. The uncertainty of the prediction was assessed by running a modified k-folds validation. The DCNN was trained and validated (in previous work) with patients from one institution. We tested the robustness of the model with a cohort of 19 patients from another institution treated with IMRT. In all cases, the criteria for the PTV coverage were V95 >= 98%, V107% < 2%, Dmean = 100% (99%-101.5%), Dmax < 110%. We considered a plan was suited for 3DCRT if the DCNN prediction plus the one-sigma uncertainty estimated 3DCRT breast Planning. Materials and Methods

fulfill the following constraints for both the ipsilateral lung and the heart: V20 < 30% for the Ipsilateral Lung, V25 < 10% and Dmean < 6 Gy for the Heart. Results

Our CNN model detected 5 out of 19 patients suitable for a 3DCRT plan. These five patients were replanned following our standard institution planning criteria. In all 5 cases, we achieved the dose prediction in 3DRT fulfilling all OAR and PTV criteria (Figure 1 shows the OARs for three of these patients, one patient per row).