Research On Pseudo CT Image Synthesis Using Patch-based Learning

In recent years,the rapid development of medical imaging technology provides patients with comprehensive and accurate diagnosis information,and was widely used in various clinical diagnosis and auxiliary treatment.Medical image synthesis refers to the prediction and synthesis of images obtained by one medical imaging device and the image obtained by another medical imaging device,also known as medical image mode mapping.Due to the different characteristics and application scope of medical images of different modes,image synthesis could realize the complementation of data and information between medical images,enrich and perfect the medical image information,reduce the excessive collection of medical images,and better serve the clinical diagnosis and disease treatment.Because of the risk of ionizing radiation and the relatively poor imaging of soft tissue by CT scanning,the clinical application of CT imaging technique was limited.Compared with CT image,MRI can’t only obtain higher quality soft tissue contrast effect,but also has no ionizing radiation in the process of scanning.Therefore,the pseudo CT image synthesis based on MRI image has become a new method of medical image analysis,and it was also the research hotspot in medical image processing domain.The main work of this dissertation is summarized as follows:(1)A method of pseudo CT image synthesis based on non-local self-similarity and joint dictionary learning was proposed.According to the problem of reducing the accuracy of image synthesis caused by weighted average of non-similar patchs in the synthesis of pseudo CT image,the image synthesis method of pseudo CT image was proposed based on non-local self-similarity and joint dictionary learning-based(DL-B).In this method,the mapping relationship between MRI image and CT image was learned by using the non-local self-similarity constraint of image and the joint sparse dictionary,so as to realize the prediction and synthesis of pseudo CT image from MRI image.The proposed algorithm was compared with atlas-based method and tissue segmentation-based method,and the experimental simulation was carried out from four aspects: skull base,skull cavity,intra-group and inter-group.The simulation results show that the proposed DL-B algorithm can obtain better pseudo CT image synthesis,and its performance is better than the existing atlas algorithm and tissue segmentation algorithm.(2)A method of pseudo CT image synthesis based on joint dictionary learning and random forest(DL-RF)was proposed.In this algorithm,the improved local binary pattern(LBP)and Gabor filter were used to extract image features,and principal component analysis(PCA)was used to reduce the dimension of image feature vector.Then,the joint sparse dictionary model was constructed,and the mapping relation matrix was trained by using the tree structure of random forest to obtain the mapping relation function of MRI and CT,and the output of reconstruction was used to realize pseudo CT image synthesis.The simulation results show that the proposed DL-RF method is more effective than the atlas-based method,the tissue segmentation-based method and the dictionary learning method.Simultaneously,the simulation and comparison between the intra-group experiment and the inter-group were carried out.(3)A method of pseudo CT image synthesis based on group feature extraction and alternative regression forest(ARF)was proposed.Firstly,the central voxel feature,patch sub-region feature and Gabor feature were used to sample the original image to obtain the patch group feature,and PCA was used to reduce the dimension of the group feature vector data.Secondly,the alternant regression forest model with self-iterative enhancement was used for training.Finally,the trained alternant regression forest model was used to predict the voxel information of CT image corresponding to the regression of MRI image patch,and the pseudo CT image synthesis was completed by reconstruction output.The simulation results show that the ARF method is more effective than the method based on atlas-based、the tissue segmentation-based、the dictionary learning-based and the DL-RF method,and the simulation and comparison between the intra-group experiment and the inter-group experiment were carried out.