Purpose: To develop an efficient residual neural network (ResNet) model to predict the cutout factor of electron therapy instantly. To overcome the issues of long model training time and sparse input data of conventional deep learning method in this clinical application.
Methods: The accuracy of electron cutout factor by Monte Carlo (MC) simulation (RayStation) was first evaluated by comparing with the measurement on 35 clinical cases. MC method was then used as the standard later to generate the data for model training/evaluation. Shapes of 281 previously used clinical electron cutouts, treated with one of four different energies (6/9/12/15 MeV) and three different cones (10/14/20 cm), were selected for the study. These shapes were varied in planning software by off-center distance and size to generate a 600-sample digital library per cone applicator (thus 1800 shapes for 3 cone sizes and 7200 digital cutouts for all energies). A ResNet prediction model with 7 blocks and 11 layers was developed for each energy/cone combination, while 400 shapes were randomly chosen as training data, 50 as validation and 150 as testing. In order to reduce training time, a 1D distance histogram was defined to characterize the shape and replace the conventional 2D image input of ResNet model. The model predictions of 1800 testing cases were compared with the MC calculated results.
Results: The MC calculated cutout factor agreed with the measurement (0.7Â±0.5%). Distance histogram accelerated the training process about 40 times faster. The model prediction accuracy was 0.20Â±0.16% for all tested cases and maximum discrepancy was 0.69%, 0.96%, 0.90% and 0.89% for energy 6, 9, 12 and 15 Mev, respectively.
Conclusion: We simplified ResNet model to accelerate the training process and overcome the limited availability of training data through data augmentation. The model can instantly and accurately predict clinical electron cutout factor.