A Feasibility of Respiration Prediction Based on Deep Bi-LSTM for Real-Time Tumor Tracking

Abstract

In radiotherapy, the position of thoracic-abdominal tumor is changing due to respiratory motion. Real-time tracking of thoracic-abdominal tumors is of great significance in improving the treatment effect of radiotherapy. The accurate prediction of thoracic-abdominal tumor motion is required to compensate for system latency in image-guided adaptive radiotherapy systems. The purpose of this paper is to identify an optimal prediction model to improve the treatment effect of radiotherapy. A seven-layer bidirectional long short term memory (Deep Bi-LSTM) and one output layer deep neural network is proposed to predict respiration motion for a latency about 400 ms. 103 malignant lung tumor patients’ respiratory motion data is used to train model. Mean absolute error (MAE), root mean square error (RMSE), and normalized mean square error are introduced to evaluate the performance of predictive results. Deep Bi-LSTM has great performance in the cases with relative long latency, average MAE of 0.074 mm, RMSE of 0.097 mm, and normalized root mean square error (nRMSE) of 0.081 with latency about 400 ms are obtained from predictive results of Deep Bi-LSTM. It demonstrates that the prediction accuracy of our proposed Deep Bi-LSTM is about five times better than traditional autoregressive integrated moving average model and about three times better than adaptive boosting and multi-layer perceptron neural network when the latency of 400 ms. The method can be applied to improve tracking accuracy and increase efficiency in thoracic-abdominal radiotherapy, which is practical and attractive for clinical application in the near future.