Reconstruction Of Cardiac MRI Based On Unsupervised Deep Learning In Gradient Domain

As a non-invasive medical imaging technology,cardiac magnetic resonance imaging(CMRI)can not only provide all-round information such as cardiac structure and function,but also realize pathological imaging of its histological characteristics,which plays an important role in the diagnosis,prognosis and risk analysis of cardiac diseases.However,CMRI requires acquisition of images with high temporal and spatial resolution,different contrast,and whole-heart coverage,resulting in excessively long acquisition times.In addition,due to the dual effects of breathing motion and heart beating,artifacts and blurring may be introduced during the imaging process.Therefore,speeding up imaging and reducing the influence of motion are the keys to improve CMRI.In recent years,deep learning methods have explored prior information from a large amount of image data through structures such as convolutional neural networks,which can solve a variety of image processing problems.Since the supervised model is only effective for specific tasks and lacks generality and robustness,this thesis proposes a model based on unsupervised deep learning in the gradient domain(UDLGD)for CMRI.Through the conditional noise score matching network,a deep generative prior for cardiac dynamic data is obtained,which is then combined with traditional total variational sparse constraints for good performance.Specifically,gradient transformation and channel fusion strategy are applied to the fully sampled cardiac dynamic sequence to reduce the data dimension and utilise its sparse properties.Then,the gradient of dynamic data distribution is fitted by the generative network for obtaining the deep generation prior.Finally,under the constraints of deep generative priors,total variational sparseness and undersampled data,the annealed Langevin dynamic sampling and a three-step alternating minimization method are used to make the random Gaussian noise gradually approach the original cardiac dynamic images.The proposed UDLGD model uses an unsupervised generative network to estimate the gradient information of the data distribution,makes full use of the spatiotemporal correlation of cardiac dynamic sequences,and combines deep generative priors with traditional sparse algorithms to construct a novel sparsity constraint.Compared with classical CMRI methods,the UDLGD model excels in improving visual effects,reducing detail errors and speeding up imaging,which significantly improves the quality of reconstruction.