Study On Prediction Algorithm Of Tumor Position Based On Dynamic Tumor Tracking Radiology In Radiotherapy

With the development of radiotherapy technology,more and more researchers have been focused on adaptive radiation therapy.Dynamic tumor tracking radiotherapy can illuminate a prescribed dose more accurately and efficiently than other methods of solving tumor movement problems caused by breathing and heartbeat.However,there is a delay in the Multi-Leaf Collimator(MLC)control system in dynamic tumor tracking radiotherapy,which causes radiation that should be irradiated on the tumor to illuminate healthy tissue,thereby causing damage to the healthy tissue,and because of the delay,the dynamic tumor tracking radiology technology cannot achieve the desired therapeutic effect.Therefore,we have proposed a method to avoid the effect of this delay on the dynamic tumor tracking radiotherapy – use the mathematical model to construct a prediction model to predict the future tumor position.This lag is known to be almost 500 ms and the objective of the current study is accordingly to generate the one-second future position estimate of the tumor with the tolerance of 1mm prediction error in the three-dimensional space.However,the models constructed by different researchers that can predict the tumor position have not achieved the required accuracy.Therefore,it is significant to propose a new predictive model that can achieve the prediction accuracy.The thesis focus on the construction of new mathematical models and prediction algorithms to predict the future tumor position.The goal of this experiment was to propose a new prediction model and to predict the different patients which the average root mean square error of all predicted outcomes to be less than 1 mm.The paper mainly proposes three new mathematical models,including prediction models based on recurrent neural network,nonlinear autoregressive networks with exogenous inputs,and long-short-term memory networks.According to the structural characteristics of the three models and the feature of the input signals.We adjust the structure of these three models.At the same time,two model optimization algorithms which including backpropagation through time algorithm and real-time recurrent learning algorithm are proposed.The prediction model was used to be trained by the tumor trajectory coordinates of seven cancer patients obtained from the oncology department of Yamaguchi University Hospital and predict the future movement of them.The training process involves the processing of the input signal,the optimization of the model using the weight adjustment algorithm and the feedback in real time to the optimization process.The paper uses each patient's own tumor trajectory to train the same patient's prediction model.Each patient uses three prediction models including the X-axis,Y-axis,and Z-axis prediction models,and use them to predict the same patient's own tumor trajectory.These three trained models are used to predict tumor position 500 ms ahead,the root means square error and gated duty cycle were used to measure the accuracy of the predicted result.Among all results of prediction models,the results predicted by the prediction model based on the nonlinear autoregressive network with exogenous input have the best performance,and the average root means a square error of the seven patients is less than 1 mm.which is achieve the design requirements.Comparing the prediction results of three prediction models proposed in this experiment,the correlation model and the prediction results of the prediction model proposed by the previous researchers,the prediction model proposed in this experiment is more accurate than the related model,and the result of prediction model based on the nonlinear autoregressive network with exogenous input has improved compared with the best results obtained by previous researchers,which verifies the rationality of the model structure and shows that the model is more suitable for dynamic tumor tracking radiotherapy than other models.