Prediction Of Liver Respiratory Motion In 4D-CT Images Using Bayesian Theory

Liver cancer is a kind of common malignant tumor.It ranks the fourth and the third in the morbidity and mortality,respectively,which causes great security risks to our people.Accurate treatment of liver cancer can greatly improve the survival rate and life quality of patients.Because respiratory movement is the main factor that leads to the displacement of liver,the prediction of respiratory movement displacement of liver is very important for accurate treatment of liver cancer.Computed tomography(CT)is a commonly used imaging diagnostic technique in clinic.4D-CT can acquire images of multiple respiratory phases in one respiratory cycle.It is an effective method to predict respiratory movement by constructing a surrogate signal based on the information of abdominal surface and a motion correlation model between surface and liver.At present,most commonly used methods are to implant fiducials in the human body to track the movement of the soft tissue in the body,and then to model theire relationship,but the implanted fiducials will cause certain trauma.On the other hand,the model constructed in this way can only reflect the movement information of local position,and can not feed back the overall internal movement information,so it is not suitable for tracking the change of liver motion information.Besides,real-time requirement of respiratory motion prediction is very critical,but most of current methods focus on the accuracy improvement,large number of complex calculation in the modeling procedure leads to extensive time cost.Based on Bayesian theory,prediction methods of liver respiratory movement was analyzed in this thesis.The main contribution and innovation are as follows:(1)Construction of a prediction model of liver respiratory movement based on Bayesian principleA prediction model for compensation of liver respiration was constructed based on posterior probability of Bayesian formula.The model is built by constructing the correlation between the displacement vector field of the abdominal surface and the displacement vector field of the whole liver,which avoids the harm caused by traditional methods those need to track the movement change through implantation of fiducials,and uses the displacement vector of the whole liver,and can give feedback on the movement information in detail,which makes the model prediction more accurate and have potential in clinical application.(2)Internal and external breath amplitude correlationA model training method based on the correlation of internal and external respiratory amplitude was proposed for the prediction model of liver respiratory movement.It is found that there is a positive correlation between the change of respiratory amplitude on the surface of abdomen and that of liver Combined with the idea of internal and external breathing range correlation,the internal and external breathing range can be calculated automatically during model training,and then the most relevant data set of breathing range is selected.The advantage of this method is to reduce the interference of irrelevant information on model training and shorten the time required for model construction.(3)Respiratory motion prediction model based on registration of sparse points and correlation of internal and external respiratory amplitudeA novel respiration prediction model is proposed based on the original Bayes respiration prediction model,combining sparse surface points registration and correlation of internal and external respiration amplitude.A sparse point set registration algorithm based on CT image surface is proposed based on the breath motion model constructed by Bayesian principle.During Bayesian model construction,a B-spline registration was used to align liver tissue in vivo.Although this algorithm can feed back the movement information of the whole liver,there are few displacement in many parts of the liver except for the surface of the liver,so the error of the parts with large displacement will be averaged.Based on the sparse point set registration algorithm of CT image surface,the part of liver surface that can reflect the information of liver displacement at best was selected,which can give more feedback on the change of liver respiratory movement.Because of the surface point set is sparse,the registration efficiency is greatly improved,and the time cost is reduced during model construction and model prediction.A posterior probability was used to construct model parameters in the statistical model based on principal component analysis(PCA).On this basis,model parameters are further optimized according to the similarity between the input and predicted thoraco-abdominal surfaces.In this thesis,mean absolute error(MAE)is used to evaluate the prediction performance of the model.In the single period 4D-CT data training,the prediction of displacement vector field of ten respiratory phases is verified by leave-one-out method.In the experiment of dual-period CT data,the first period data set is used for training,and the second period data is used for prediction and evaluation.The experimental results showed that for the single period data,the respiratory movement model based on Bayesian principle achieves similar results with the traditional PCA statistical model,but the optimized model has a significant improvement in prediction accuracy.For the dual-period data,the Bayesian model with or without optimization both achieve more accurate prediction accuracy than the traditional PCA statistical model.