ESTRO 2021 Abstract Book

S650

ESTRO 2021

including user-dependent variability in delineation and planning, different machines and beam configurations; and a homogenized database (HomDB), intentionally built to reduce this variability by re-contouring and re- planning all patients with a fixed class-solution protocol. Two experiments were conducted to analyse the effect of data quality, in particular the user-dependent variability (Experiment 1), and data quantity (Experiment 2). Note that in DL, the quality is a measure of the consistency between the training and the test set, and thus HomDB has higher quality than VarDB. Experiment 1 used 26 patients planned with the same machine and beam setup (E26-VarDB versus E26-HomDB). Experiment 2 increased the training set by groups of 10 patients (E26, E36, E46, and E56) for both databases. Model evaluation metrics were the mean absolute error (MAE) and box plots for selected dose-volume metrics and the global MAE for all body voxels Results

For Experiment 1, E26-HomDB reduced the MAE for the considered dose-volume metrics compared to E26- VarDB (e.g. reduction of 0.2Gy for D95-PTV or 1.2Gy for Dmean-heart). The box plots in Figure 1 also show a reduction of the error for the HomDB database, since they are narrower than those for E26-VarDB. For Experiment 2, increasing the database size slightly improved performance for HomDB models (e.g. decrease in global MAE of 0.13Gy for E56-HomDB versus E26-HomDB), but increased the error for the VarDB models (e.g. increase in global MAE of 0.20Gy for E56-VarDB versus E26-VarDB). This is confirmed with the box plots in Figure 1, which show a slight improvement for the lungs and heart Dmean when increasing the size of the training set for HomDB (Figure 1, right), whereas none of the plotted metrics improved when increasing the training set for VarDB (Figure 1, left). In particular, for the lungs and liver Dmean, the prediction error gradually increased from the E26-VarDB to the E56-VarDB model. Conclusion This study raises awareness of the importance of the training database for the applications of DL for dose prediction. A small database may suffice to obtain good DL prediction performance, provided that homogenous training data is used. Data variability reduces the performance of DL models, which is further pronounced when increasing the training set.

PD-0819 Dose mimicking by deep learning based fluence prediction: one model for different class solutions L. Vandewinckele 1,3 , S. Willems 2,4 , M. Lambrecht 1,3 , F. Maes 2,4 , W. Crijns 1,3