ESTRO 2021 Abstract Book

S657

ESTRO 2021

purpose of this work is to create dose treatment plans for locally advanced breast cancer patients using a Convolutional Neural Network (CNN).

Materials and Methods Data of 60 patients treated for left-sided breast cancer including lymph nodes level 1 and 2 (30) or level 1,2,3 and 4 (30) was used. The dose description was 15x2.67Gy. The dataset was split in training, validation and test data (36/12/12) and a 4-fold cross validation was used. The plans were generated in Raystation v9b and were optimized in such a way that heart and lung doses were minimized while all other clinical goals were still fulfilled. The used CNN architecture was based on a hierarchically densely connected U-net (HD U-net), previously used for head and neck cancer patients [1]. The inputs for the CNN model were 2D masks containing the contoured volumes of the relevant ROIs: PTV, heart, lungs, humerus and external. These volumes were represented by distance maps, in which each pixel has a numerical value based on the Euclidean distance from that pixel towards the boundary of the ROI. These distance maps keep proximity information about the ROIs in the craniocaudal direction, without using a full 3D input. The PTV mask was overlayed with a dose mask containing the prescription dose. The CNN model was trained for 800 epochs. The batch size was 24, by selecting 8 slices of 3 patients per batch through Gaussian sampling. The model was evaluated on the clinical goals used in our institution. A Wilcoxon signed rank test was performed to investigate significant differences between treatment plans. Results The training/validation losses of the 4-fold cross validation all showed a convergence in the loss and there was no sign of overfitting. Dose predictions were generated by the CNN model and mimicked in Raystation (Figure 1), which is used to obtain a clinically deliverable plan. In cases where the D 98% clinical goal for the breast PTV (PTVb) was not fulfilled after mimicking, the dose was scaled to 38.05Gy for D 98% of the PTVb. An evaluation on the clinical goals can be seen in Table 1. The predicted plans show significant differences (p<0.05) in the doses compared to the optimized plans, while still fulfilling all clinical goals. The mimicked plans show significant differences in the doses for the heart, lung and D 2% of all PTVs. In total, ten of the twelve mimicked plans were clinically acceptable.