Denoising of Medical Images with Motion Artifacts

Abstract

newline Motion artifacts occur in magnetic resonance imaging (MRI) due to the motion or movement newlineof the patient during the MRI scan procedure. Such motion artifact-induced MRI scans are newlinevery difficult to read and sometimes lead to a faulty diagnosis. Current technology to solve newlinethis issue is based on rescanning the MRI. The problem with the rescanning method is that it newlineis time-consuming and wastes a lot of resources and overall diagnosis is delayed in the newlineprocess. Also, there is no guarantee that there will be no motion artifact after rescanning. newlineRecent developed deep learning technology allows users to recreate the image from their newlineblurred versions. Hence deep learning can be the solution to resolve motion-related artifacts newlinefrom the images. In this thesis work, we have presented modified Restricted Boltzmann newlineMachine (mRBM) and conditional Generative Adversarial Network (cGAN). We have used newlinea database collected from the list of hospitals mentioned in the appendix, consisting of 218 newlineimage pairs that include both motion artifacted scans and desired rescans obtained by newlinerepeating the MRI on the patients newlineA Restricted Boltzmann Machine (RBM) can train itself using probability distribution over a newlineset of inputs. So RBM had the potential to be used in artifact-free image generation. In the newlinepresent thesis work, it was proposed to modify the existing RBM for denoising motion newlineartifact-induced MRI scans. In the proposed mRBM the number of weights and biases to be newlinetuned are confined to visible and hidden layers and hence significantly speeding up the newlinetraining process. For a 256 x 256 pixel image, mRBM output could be achieved in about 2 newlineseconds once trained. mRBM could reach a Root Mean Squared Error (RMSE) of 0.003. newlineThe thesis also covers the detailed description of another method developed to overcome the newlineproblems with mRBM using cGAN. This method uses deep learning for removing the motion newlineartifacts from MRI scans called modified Pix2Pix. The proposed method could achieve newlineRMSE up to 0.0078 and PSNR of 26.19 dB with an