A New Deep-Learning-based Model for Predicting 3D Radiotherapy Dose Distribution In Various Scenarios

Abstract

Deep neural networks have been proved to be able to predict accurate dose prediction to improve radiotherapy planning efficiency. However, existing deep-learning-based methods could not predict dose distribution accurately for complicated cases, e.g. tumors at various locations and multi- prescriptions. Based on a new network Channel Attention Densely-connected U-Net (CAD-UNet) proposed by the authors, volume-normalized weight was firstly multiplied to the Mean Squared Error, defined as VN-MSE, as the loss function in the dose prediction area. A cohort of VMAT plans for lung cancer patients was selected for this study. The results show that the new model CAD-UNet with VN-MSE can successfully predict dose distribution of lung cancer cases with single and multiple prescriptions, outperforming CAD-UNet with MSE loss and HD-UNet. The new model demonstrates its potential to be applied for dose prediction in more complicated scenarios.