Deep Learning Based Cardiac Magnetic Resonance Image Deblur

The use of imaging devices to record image information is of great importance in a variety of fields,but in some cases motion blur can occur in images,preventing the human eye as well as algorithms from obtaining accurate information from the image.In recent years,deep learning models have been used extensively in the field of image processing,including image deblurring as a problem.In the medical field,image deblurring algorithms have a highly practical place.Velocityencoded cardiac magnetic resonance imaging(VENC MRI)has powerful capabilities in cardiovascular blood flow analysis.However,during imaging,sometimes the uncontrollable motion of the patient dynamically adjusts the position of the heart,which leads to motion blurring,which in turn may lead to inaccurate analysis of subsequent conditions.This thesis proposes a new deblurring model,first validating its performance on a public natural dataset,and then applying this model to VENC MRI,with the following main work.For natural image deblurring,this thesis incorporates an attention mechanism based on the structural features of classical networks for multiscale deblurring to ensure that less real information is lost in the feature information transfer process.In addition,we construct a new dataset that allows the model to explicitly process the "direction of motion" information in the image,thus reducing the solution space of the deblurring model and the difficulty of fitting the model.Tested on the publicly available dataset GOPRO,the proposed deblurring model outperforms SRN-Deblur and Deblur GAN in terms of numerical performance,with PSNR and SSIM reaching 31.158 and 0.927 respectively.For cardiac velocity-encoded MRI images,this thesis implements a defuzzification process to visualise and compare the blood flow patterns in the atria.To achieve the goal,the principle of cardiac VENC MRI blurring is first investigated,whereby the cardiac VENC MRI deblurring dataset is simulated and constructed to approximate the real situation;the atrial region is then segmented and the blood flow patterns in it are shown using swirl arrow plots to facilitate comparison of the deblurring effect.In order to verify the ability to process real images,the ability of the deblurring model to process real blurred images was tested using time-frame images of the same sequence number from two heartbeat cycles of the same subject as the mapping pair of blurred and clear,and using the similarity of cardiac blood flow in time-frame images of the same sequence number in different cycles.Finally,an easy-to-use software was designed and implemented to facilitate the deblurring of VENC MRI and the analysis of atrial blood flow using the algorithm proposed in the paper.