Research On Cross-Modality And Super-Resolution Reconstruction Of Medical Images Based On Weakly-Supervised Learning

In recent years,image super-resolution and cross-modal technology have been continuously developed and applied in various fields.As a radiation-free,fast,and multimodal imaging method,Magnetic resonance imaging is useful in clinical medical diagnosis.However,the acquisition of medical image data is very limited due to personal privacy issues.On the one hand,due to the scanning cost of magnetic resonance imaging and patient comfort,it is difficult to obtain a mass of magnetic resonance imaging of human body in modern medicine.With only a few data sets,training a good superresolution model is one of the hottest research issues in the field of medical image processing.On the other hand,magnetic resonance imaging technology can theoretically provide doctors with a variety of observation data.However,in practice,only one mode of imaging can be obtained.How to efficiently and accurately reconstruct magnetic resonance imaging in another modality based on a limited data set is the goal pursued by all cross-modal synthesis techniques,and it is also one of the most challenging difficulties.The efficient and accurate reconstruction of magnetic resonance imaging from one modality to another is the purpose of all cross-modal synthesis techniques and one of the most challenging issues.In this thesis,weakly supervised machine learning will be combined with convolutional sparse coding to analyze the super-resolution reconstruction of single magnetic resonance imaging,and systematically explain and implement the single-image super-resolution reconstruction based on convolutional sparse coding.This thesis discusses cross-modal synthesis algorithms based on convolutional sparse coding,and proposed a unified weakly supervised learning model.The main contents of this thesis are as follows:(1)This thesis discusses a super-resolution reconstruction algorithm based on convolutional sparse coding.The feature map of a single image with a specific filter can be extracted by using Fourier transform.(2)For magnetic resonance imaging from different modalities,this thesis innovatively make pair between the data sets by using Gaussian distance.By performing convolutional sparse coding on paired different modal data,a pair of filters are learned for cross-modal synthesis,which adapted to the same set of feature maps.(3)Aiming at the memory pressure caused by the large number of feature maps,the BCCB matrix is used to transform the convolution operation into matrix multiplication,which results the filter can be updated according to a single image.There is a certain mapping relationship between different feature maps.The transformed feature map can be sparse by adding some constraints,and the SA-ADMM algorithm is used to optimize the mapping function.(4)Based on the super-resolution technology and convolutional sparse coding,this thesis incorporate cross-modal synthesis technology,and use the data differences from different modalities to constrain the mapping function,which can be ensured for always performing on the paired data.A large number of experiments show that the method proposed in this thesis has achieved good results in both super-resolution reconstruction and cross-modal synthesis.