Research On 3D-MRI Super-resolution Reconstruction Algorithms Using Deep Convolutional Neural Networks

Magnetic resonance imaging(MRI)is a safe and non-radioactive medical imaging technology.It can provide high-contrast human tissue images and highlight the detailed information of different tissues with flexible multi-parameter imaging.However,MRI has longer scanning time,and patients cannot keep still for a long time,thus its imaging process suffers from motion artifacts.Increasing scanning thickness can accelerate MRI,but the larger the scanning thickness,the lower the spatial resolution of MRI image.Besides,it will produce partial volume effects,which is bad for disease analysis and diagnosis.Therefore,how to obtain high-resolution MRI images is an urgent problem that needs to be solved in the field of MRI research.Recently,super-resolution(SR)reconstruction algorithms based on deep convolution neural networks(DCNNs)provides a new idea for improving the spatial resolution of image.DCNN has a strong nonlinear learning ability,which makes a breakthrough in SR reconstruction research.However,there are many challenges when applying DCNN to SR reconstruct MRI images.This thesis carries out innovative research on some important issues in DCNN-based MRI image SR reconstruction research field.Novel SR reconstruction algorithms for anisotropic and isotropic MRI images are proposed in this thesis,and they greatly improve MRI image SR reconstruction accuracy.The main innovations and work of this thesis are as follows:(1)This thesis proposes a residual-learning-based algorithm(RLSR)for single anisotropic 3D-MRI image SR reconstruction.RLSR promotes features and gradient information flow the network via local and global residual learning,and it can use two-dimensional network to learn three-dimensional features based on the cross-layer self-similarity of MRI image.RLSR can SR reconstruct the medium and high-frequency information of MRI images in any slice-select direction,which effectively improves the SR reconstruction accuracy of MRI images with fewer hardware resources.In addition,based on the characteristic and complementarity of T1-weighted(T1w)and T2-weighted(T2w)MRI images,RLSR carries out DCNN-based multi-modality MRI image SR reconstruction research.Specifically,RLSR introduces the information of high-quality T1 w image into its training so that it can simultaneously SR reconstruct T1 w and T2 w images.Compared with single-modality MRI image SR reconstruction algorithms,RLSR greatly improves the SR accuracy of T2 w image.(2)This thesis proposes a dilated convolutional encoder-decoder(DCED)algorithm for isotropic MRI images SR reconstruction.In DCED,dilated convolution encoders are used to infer and learn from the context information in larger image areas,and deconvolution decoders are applied to alleviate gridding artifacts and recover high-frequency details of 3D-MRI images.DCED can expand the receptive field of SR network without introducing additional parameters.Compared with the method using different dilated factors to alleviate gridding artifacts,DCED uses deconvolution to avoid bounding artifacts and improve SR reconstruction accuracy,and its quantitative and qualitative evaluation results are better than compared SR algorithms.In order to further improve the SR reconstruction accuracy of MRI images,this thesis proposes a feature-level geometric self-similarity 3D wavelet fusion algorithm,which can retain high-frequency details of MRI images when fusing different images and obtain superior fusion results than the pixel-level mean fusion method.(3)This thesis proposes a parallel convolution and deconvolution SR(CDSR)algorithm to reconstruct 3D-MRI images.CDSR uses parallel convolution and deconvolution to directly extract context information and multi-level features from low-resolution 3D-MRI image.The features of each level are SR reconstructed using corresponding deconvolution layer and then fused at the end of the network.CDSR can automatically learn up sampling parameters,thus it does not rely on interpolation upsampling preprocessing method.In order to alleviate the training difficulty of deep networks,this thesis proposes cross-scale residual learning by introducing low-resolution images into the output features of CDSR,which enables the network to focus on learning the differences between high-and low-resolution MRI images and speeds up the convergence of training.Experimental results on normal and tumor MRI images show that CDSR can restore the details of high-resolution MRI images more clearly,and it achieves state-of-the-art quantitative and qualitative SR reconstruction accuracy.