Sparse Sampled MRI Reconstruction Based On Dual Domain Processing

Magnetic resonance imaging is non-radiation and non-invasive.But the imaging process takes a lot of waiting time and can be uncomfortable for patiens.At present,reducing the amount of data collected is usually used to reduce the imaging time.However,collecting only part of the information will lead to aliasing effects.The MRI reconstruction algorithm is to reconstruct a high-quality image close to the reference image from the aliased data.From the perspective of the type of data processed,reconstruction algorithms can be divided into reconstruction algorithms on frequency domain and reconstruction algorithms on image domain.At present,most reconstruction algorithms only focus on frequency domain data or image domain data.In order to make comprehensive use of the two forms of data,this thesis combines frequency domain reconstruction algorithm with image domain reconstruction algorithm to accurately reconstruct aliase-free images directly from down-sampled frequency domain data.Based on this idea,this thesis conducts the following research.Firstly,this thesis verifies the effectiveness of joint reconstruction.Reconstruction experiments with different sampling rates have been conducted on two datasets.One is the Calgary-Campinas public dataset,and the other is provided by Xin Gao Yi company.The experimental results show that the Grappa algorithm reconstructed in the frequency domain can better recover the details of the image and preserve a high image contrast.However,the Grappa algorithm has a very strict limitation on the sampling rate.When the sampling rate is low,the phenomenon of noise amplification will occur.The Total Variation algorithm and TBMDU algorithm reconstructed on the image domain can better eliminate image artifacts and noise,but more image details will be lost.The joint reconstruction algorithm in frequency domain and image domain can comprehensively utilize the advantages of the two algorithms,so as to obtain the optimal reconstruction results at the same sampling rate.However,in the case of low sampling rate,due to the limitation of the traditional reconstruction algorithm model,it is impossible to effectively suppress artifacts and preserve image details.Therefore,this thesis combines the idea of joint reconstruction with the convolutional neural network,and proposes a dual-domain joint reconstruction network named Hybrid Net.Hybrid Net can not only use the correlation of frequency domain data to fill in the missing data,but also effectively eliminate image artifacts and noise.At the same time,this thesis also proposes a data consistency module,which can correct the data of the whole frequency domain space rather than just the sampling position.Through experiments on the same two dataset,the effectiveness of Hybrid Net has been proved.