Research On Generating Pseudo-CT Data Based On MR Data

In recent years,PET/MRI is the most advanced multimodal molecular imaging technology among various molecular imaging technologies,with high accuracy and no radiation.As CT data provides tissue density information for attenuation correction of PET data,signals in MR image are not directly related to the ionizing radiation attenuation factor required for PET data attenuation correction,so,predicting the CT image with MR image has important clinical significance for the PET/MRI system.Therefore,the method of generating pseudo-CT data based on MR data is studied.The main researches were carried out as follows:(1)Research on pseudo-CT image prediction model based on tissue segmentation.According to the method of tissue segmentation,we need to accurately and effectively extract the CT values of different tissues and their respective labels,so we proposed a gray histogram and K-means clustering method for the extraction of brain tissue.However,it was found that the effect of classification of brain tissue on real MR image without specific sequences can not satisfy the application of matched CT values,so a model study of subsequent map registration and image translation of MR image prediction CT value was proposed.(2)Research on pseudo-CT image prediction model based on image registration and image translation.This part is divided into two aspects: image registration and image translation.The first part studied the multimodal registration between MR image and CT image.It is assumed that there is a one-to-one correspondence within the gray values between the MR image and the CT image.The maximum mutual information method with affine registration of six degrees of freedom was used here,which commonly used in medical image registration.The Powell algorithm can not scale the image,in order to solve this problem,we proposed two adaptive registration methods.The real dataset was used to preprocess the registration.The results showed that the two algorithms can improve the registration accuracy of the multimodal image of MR and CT.In the second part,we studied the image translation of pseudo CT images after the registered MR-CT dataset.Calculating a CT image from MR images is essentially a representation of one image to another.So,we proposed a Pix2 Pix model for generative adversarial nets algorithm applied to compute pseudo-CT image from MR image.Experiments were performed on adaptive affine aligned MR-CT image pairs,the results showed that it is feasible to generate pseudo CT image using MR image,and the structural similarity of pseudo CT image and real CT image reaches about 0.88.