A Study On Prediction Algorithm Of MRI-based For Generating Pseudo-CT

As one methods of the treatment of tumor,radiation therpy mainly uses radiation to kill tumors.With the improvement of technology,it has evolved from initial systemic radiation therapy to conformal radiation therapy,and then to image-guided radiation therapy(IGRT).IGRT based on images to overcome errors that caused by positioning and tumor movement.At present,computer tomography(CT)is used as a baseline image during radiotherapy,and doctors can outline the clinical target volume(CTV),planned target volume(PTV)and organ at risk(OAR).The electron density(ED)of the tissue can also be obtained from the CT for dose calculation.Although CT’s image quality has high spatial resolution and simple and fast operation,its low contrast density has low resolution,which limits the contour information of the tumor,and the patient will also be exposed to ionizing radiation.Compared with CT,magnetic resonance imaging(MRI)not only provides superior soft tissue contrast,but also has a variety of sequence images.Doctors can distinguish tissues or organs based on different sequence images.It is often used for radiation therapy with CT using for the outline of target volume and OAR.Considering the disadvantages of CT ionizing radiation affecting the patient’s health,simultaneous scanning of both to increase the patient’s financial burden and the introduction of errors in the registration fusion process,the use of MRI alone has attracted more and more attention from researchers.However,MRI is not related to ED and cannot be directly used for dose calculation and X-ray-based patient positioning.It is necessary to apply related algorithms to obtain electron density information or HU value according to MRI,that is to generate pseudo CT(pCT).In recent years,many experts and scholars have studied the conversion of MRI-based CT images,such as tissue segmentation-based methods,atlas-based methods,and 2D deep convolutional neural networks(2D-DCNN).The method of tissue segmentation is time-consuming and laborious,and the error at the joint between air and bone is large.Although the atlas-based method relatively improves the accuracy of prediction,there is a large part of the problem of misconversion between bone and air,and the method requires higher registration.The 2D-DCNN combines deep learning with neural networks,adopts an end-to-end learning method,and predicts the results based on the training model.Although this method improves the errors at the bones and the air,it only applies the intra-layer features of the image,resulting in more serious misconversion problems of the prediction results and there are more serious artifact,so it needs to be further improved.This research proposes a cascaded network based on convolutional neural network and long short-term memory network(CNN-LSTM)and a 3D-based deep convolutional neural network(3D-DCNN)to connect the relationship between image layers and layers.CNN-LSTM cascades CNN and recurrent neural network(RNN).The CNN part uses the down-sampling layer of the U-net network to extract features from the image through multiple 3×3 convolution kernels,and the pooling layer after convolution layer plays the role of secondary extraction of features to obtain features with spatial invariance.The convolution and pooling are repeated several times,and the output is used as the input of the RNN.Due to the conventional RNN cannot handle the long-term information dependency problem,we use LSTM to extract information between adjacent image layers.Mean absolute error(MAE)was used as the loss function,and Adam stochastic optimization function was used to back-propagate the obtained error.The parameters were further optimized until the minimum prediction error was obtained.The 3D network model can be regarded as a complex 3D data end-to-end mapping function,which is mainly divided into upsampling and downsampling.Upsampling mainly uses the network structure of multiple convolutional layers and pooling layers to learn the relationship between MRI and CT;Downsampling mainly uses deconvolution and depooling operations for inverse operations.Thousands or even millions of parameters in the convolution process are mainly used to extract features,such as edges,lines and corners.As the number of convolutional layers increases,the extracted features are more and more advanced,but this does not mean that more convolutional layers will help the training of the network,and it is also affected by the size of the convolution kernel and the number of feature map.During the training process,these parameters are continuously optimized with the prediction error.The article use the MAE to measure the error between the output prediction value and the true value of the network,and use the Adam optimization function to optimize the parameters until the minimum error is obtained.Once the model is trained,the results are predicted based on the learned parameters.This study collected CT and MRI data of 13 patients with head brain tumors treated by Cyber Knife.Five prediction algorithms based on tissue segmentation-based methods,atlas-based,2D-DCNN,CNN-LSTM,3D-DCNN were compared respectively through retrospective offline analysis.The experiments show that the MAE,root mean square error(RMSE),registration error,structural similarity(SSIM)and DSC of bone of the CNN-LSTM and 3D-DCNN proposed in this paper are better than the other three methods.Among them,the accuracy and robustness of the 3D-DCNN algorithm are better than other methods,which can obtain pCT more accurately.