Deep Learning In Real-time Tracking Method For Respiratory Motion Of Thoracic And Abdominal Tumors

Cancer is still one of the diseases with the highest mortality rate in the world,which seriously threatens the healthy life of all mankind.About 60% of cancer cases require radiotherapy,and most of them are thoracic and abdominal cancer patients.The treatment of thoracic and abdominal tumors is more difficult than other tumors,the location of the tumors in the body changes with the respiratory movement,which poses a great obstacle to the treatment of cancer.In the course of radiotherapy,the radiation beam may be irradiated to healthy tissues due to the influence of respiratory movement,which will cause a series of complications.Precise radiotherapy technology is imminent.Because of the movement of the target area of thoracic and abdominal tumors,real-time and accurate tracking of the moving area of tumors is of great significance to improve the accuracy of radiotherapy and radiotherapy efficacy.This paper presents a method of respiratory motion compensation based on Recurrent Neural Network(RNN).In this paper,the method of respiratory motion compensation is divided into two parts.Firstly,the prediction algorithm is trained according to the respiratory data of in vitro markers,and then the correlation model is based on the two parts of respiratory data in vivo and in vitro.Through these two algorithms,the location of tumors in vivo can be predicted about 400 ms.The prediction time is to overcome the system latency in radiotherapy and achieve the goal of precise radiotherapy.For the respiration prediction algorithm,103 volunteer respiratory databases are used in this paper.Firstly,the in vitro respiratory data were preprocessed,including filter the noise,normalization and smoothing of respiratory data.Then the whole data set is divided into training set(80%)and test set(20%),the processed data is used for our Deep Bi-LSTM network training.LSTM(Long Short-Term Memory)is a special RNN,which is more suitable for processing sequence data.The network composed of LSTM is easier to train than ordinary RNN network,and can largely eliminate the problem of gradient explosion and gradient disappearance in the training process of RNN.Our Deep Bi-LSTM network consists of 1 input layer and 7 LSTM layers.Each layer of LSTM connects to a Dropout layer,the Dropout Ratio is set to 0.2.The last layer is fully connected layer with tanh function as activation function.Due to the limitation of the amount of correlation data,the correlation model uses a relatively simple multi-layer perceptron(MLP).MLP only contains three layers: input layer,output layer and hidden layer.Deep Bi-LSTM respiration prediction algorithm has better effect than traditional prediction method,RMSE(Root-Mean-Square Error)is used to evaluate the performance of the model.In this experiment,our Deep Bi-LSTM prediction algorithm improves significantly with RMSE=0.097 mm,which is about five times higher than the conventional method.Combining the prediction algorithm with the correlation model,RMSE=0.572 mm meets the precision requirement of precise radiotherapy.Respiratory motion compensation method based on LSTM can accurately predict the patient’s respiratory movement.This method can be applied to real-time tracking of thoracic and abdominal tumors,improve the accuracy and efficiency of the radiotherapy process of tumors,which has great clinical significance.