Respiratory Prediction With Wavelet Decomposition And Adaptive Neuro Fuzzy Inference System

Radiotherapy plays a increasingly important role as one means of the treatment of tumor.With the development of computer technology and radiation physics,radiotherapy becomes more precise from traditional conventional radiotherapy and conformal radiotherapy to intensity modulated radiotherapy and image guided radiation therapy.Pursuing high precision,high dose,high efficacy and low loss are the goals of accurate radiotherapy all the time.The target tumor could be removed from the beam range by respiratory motion while the normal tissue moves into the beam range in radiotherapy for thoracic and abdominal tumors.The motion has an adverse effect on the accuracy in radiotherapy target positioning causing the influence of radiotherapy curative effect and the increase of the damage of the normal tissue.In order to reduce the impact of breathing motion and ensure the high dose irradiation of the clinical target volume(CTV),the traditional method is to expand the target boundary of CTV to internal target volume(ITV)by considering the breathing movement displacement of the target,and then to formulate the radiotherapy plan.The motion of tumor could also be reduced or compensated through respiratory control technique,respiratory gating technique,image-guided real-time tracking technique or 4-Dimensional radiotherapy technology with the development of science and technology.It takes a certain time from information acquisition to beam adjustment to make the response of accelerator or the movement of the robot arms lag behind the motion of the tumor in these technologies,resulting in the system time delay.Predicting respiratory motion could compensate system time latency and thus improve targeting accuracy in image-guided tracking or gating radiotherapy of thoracic and abdominal tumors.At present,various kinds of breathing prediction algorithms have been put forward by scholars at home and abroad such as neural network(NN),support vector machine(SVM),adaptive neural fuzzy inference system(ANFIS)and the prediction algorithm using by Synchrony Respiratory Tracking System in CyberKnife Robotic Radiosurgery System.The NN might fall into local minimization or create over fitting issues easily.The SVM is high complexity and the prediction results depended on the selection of parameters.The ANFIS uses the positions of the respiratory motion as input parameters of the model to construct a fuzzy set for prediction.The prediction ability is strong,but the prediction error become too large for the irregular signal,especially if the amplitude of respiratory signal changes suddenly.The prediction accuracy of Synchrony has been verified by a large number of clinical data.However,the prediction of Synchrony is relatively large for irregular or abnormal breathing and still does not meet clinical needs.Breathing motion varies from person to person,and the baseline,frequency,and period of the respiratory movement changes over time for the same patient.It affects the prediction precise.This paper aimed to develop a more precise prediction algorithm to predict irregular respiratory motion pointing at the complexity and uncertainty of respiratory motion.A prediction algorithm based on wavelet decomposition and adaptive neuro fuzzy inference system(WANFIS)was proposed.The respiratory signal was first decomposed into baseline,low frequency and high frequency components,which were then predicted respectively using linear fitting,ANFIS and moving average.The baseline was smooth and predicted by linear prediction.The low frequency component was the up-and-down motion without the baseline and used ANFIS to predict.The proposed ANFIS model structure utilized a novel model structure which not only includes position but also velocity as the input parameters.A fuzzy set with N×N matrix was constructed combining position and velocity.During performing prediction by ANFIS,if the amplitude of the input parameter,either position or velocity,in the newly coming input data was out of the range of the training set built by historical respiratory data,it was accordingly adjusted to fall in the range.Then the parameter with the adjusted amplitude was applied as the input in the ANFIS model to carry out prediction.The high frequency component was mainly noise and selected the moving average method.The three components of predicted output were finally combined to be the respiration prediction result.The WANFIS prediction algorithm was evaluated using respiratory motion data of 42 cases of thoracic and abdominal tumors treated by CyberKnife.These patients were particularly selected to have irregular breathing patterns.The clinical data was retrospectively analyzed.The results of WANFIS were compared respectively with the results calculated by NN,SVR,ANFIS and Synchrony in CyberKnife.The prediction results of five algorithms showed that in term of normalized root mean squared error(nRMSE),mean error(Mean),standard deviation(STDEV),maximum error(Max)and the number of errors greater than 1mm,WANFIS is better than NN,SVR,ANFIS and Synchrony.The proposed WANFIS algorithm in overall had better performance than other four methods in accuracy and robustness.It was able to predict irregular respiratory motion more precisely and compensate system time latency more efficiently.