Incorporating historical sub-optimal deep neural networks for dose prediction in radiotherapy.

• The non-modulated dose distribution, which is the initial quantity of the inverse optimization phase in radiotherapy, is used as part of the input to aid the training of the segmentation-based network. • A novel ensemble method that takes advantages of the sub-optimal models during the historical training phase is used to improve the predictions on test data. • Empirical experiments demonstrate that the non-modulated dose distribution provides effective auxiliary information which facilitates the convergence of the network during the training phase. Moreover, the dose distribution predictions can be further improved by employing the HSE method. As the main treatment for cancer patients, radiotherapy has achieved enormous advancement over recent decades. However, these achievements have come at the cost of increased treatment plan complexity, necessitating high levels of expertise experience and effort. The accurate prediction of dose distribution would alleviate the above issues. Deep convolutional neural networks are known to be effective models for such prediction tasks. Most studies on dose prediction have attempted to modify the network architecture to accommodate the requirement of different diseases. In this paper, we focus on the input and output of dose prediction model, rather than the network architecture. Regarding the input, the non-modulated dose distribution, which is the initial quantity in the inverse optimization of the treatment plan, is used to provide auxiliary information for the prediction task. Regarding the output, a historical sub-optimal ensemble (HSE) method is proposed, which leverages the sub-optimal models during the training phase to improve the prediction results. The proposed HSE is a general method that does not require any modification of the learning algorithm and does not incur additional computational cost during the training phase. Multiple experiments, including the dose prediction, segmentation, and classification tasks, demonstrate the effectiveness of the strategies applied to the input and output parts. [ABSTRACT FROM AUTHOR]