Research On Convolutional Neural Networks For MR Image Fusion And Magnetic Susceptibility Reconstruction Based On Magnitude And Phase Data

Magnetic Resonance Imaging(MRI)is an imaging technology that can accurately display the internal tissue structure of an object.However,most MRI methods only use part of the information in the magnitude and phase data,resulting in unnecessary information waste.In addition,the traditional post-processing technology of MRI requires manual design of feature extraction rules,which not only has a complex process but also cannot extract features effectively.In recent years,Convolutional Neural Network(CNN)has achieved great success in the field of image processing with its unique advantages of automatic learning image features.Inspired by this,in order to obatain more comprehensive and accurate descriptions of tissue structure,this thesis carries out related research on convolutional neural networks for MR image fusion and magnetic susceptibility reconstruction based on magnitude and phase data.The main work is divided into the following two parts:This thesis performs image fusion on the magnitude and field maps from the magnitude and phase data to make full use of the effective information in the data and reduce the redundancy of the images.In order to improve the robustness of the fusion system,the artifact suppression network is used to judge and suppress the artifacts in the fusion images.Then the weight prediction network is designed to distinguish the gradient of the corresponding regions in the two images.The region with rich texture will be assigned larger fusion weight and recorded in the weight map.Finally,a fusion network is designed to perform feature-level fusion of the input images under the guidance of the weight map.Compared with other methods,the fusion result of this method contains more effective information,not only has good visual effect,but also achieves a leading position in multiple objective evaluation indexes.Quantitative susceptibility mapping is a new post-processing technique to reconstruct susceptibility maps from magnetic resonance phase data.In order to solve the ill-posed inverse problem in the reconstruction process,the spatial adaptive network is used to directly learn the mapping between the field map and the magnetic susceptibility map.The spatial adaptive network improves the feature extraction capability through the combination of the Inception block and the multi-scale coding layer,uses channel weight to assist the network in feature selection,and fuses the magnitude information in the spatial adaptive module to improve reconstruction accuracy.Compared with other methods,spatial adaptive network makes full use of information in magnitude and phase data,which not only reconstructs more accurate susceptibility maps on the data of healthy human brain,hematoma human brain and simulated human brain,but also has good training stability.The research in this thesis shows that CNN has its unique advantages in image fusion and reconstruction algorithm.Through reasonable network design,we can effectively use the information in magnitude and phase data to provide a more comprehensive and accurate description of the organization structure for doctors.