3D MR Image Super Resolution Reconstruction Algorithm Based On Sparse Representation

Magnetic resonance imaging(MRI) plays a key role in the examination of various diseases, which is a non-invasive, multiple parameters imaging method with high-contrast for soft tissue. However, scan time constraints and patient comfort often result in motion artifacts. Fast MRI acquisitions are preferred to avoid motion artifacts; so many MRI scans are performed with relatively few slices and with rather large slice thickness. As a result, highly anisotropic data sets are acquired that have higher resolution within the slices than in the slice-selection direction due to partial volume effects. Such low resolution MR images posed limitation on the computer aided diagnosis, analysis, and research on disease. Thus, reducing the voxel size and reconstruction isotropic 3D MR images with high spatial resolution is much desired. Super resolution(SR) techiniques have emerged as efficent methods to improve the resolution of images. The SR algorithms indeed improve the resolution, SNR(Signal-to-Noise Ratio) and acquisition time trade-offs compared with direct high-resolution acquisition.In this work, focusing on single frame and multi-frames super resolution reconstruction, we proposed three novel sparse representation based super resolution algorithms to reduce the thickness of the low resolution 3D MR image and reconstruct high resolution 3D MR image. The main contributions and results of this thesis are as follows:(1)we proposed cross plane direction dictionary learning based single frame super resolution algorithm(cpSFSR): To reconstruct high resolution 3D MR image, cpSFSR algorithm utilizes the self-similarity cross plane directions and trains the over-complete dictionary from in-plane high resolution slices to upsample in the out-of-plane dimensions. cpSFSR algorithm does not require extra training sets and provides trade off between the reconstruction quality and robustness for the trained over-complete dictionary. Abundant experiments, conducted on simulated and clinical MR images, show that cpSFSR algorithm is more accurate than classical interpolation in terms of PSNR(Peak Signal-to-Noise Ratio) and SSIM(Structural SIMilarity). When compared to a recent upsampling method based on the nonlocal means approach, the cpSFSR algorithm did not show improved results at low upsampling factors with simulated images, but generated comparable results with much better computational efficiency in clinical cases.(2)We proposed Multi-Modality MR images based Single Frame Super Resolution algorithm(mmSFSR): mmSFSR algorithm exploits the complementary between T1-weighted MR image and T2-weighted MR image, and introduces the prior knowledge of T1-weighted MR image into the super resolution reconstruction process of the low resolution T2-weighted MR image. In order to utilize the prior knowledge of T1-weighted MR image and the self-similarity of T2-weighted MR image, we constructed an optimization model with multiple constraints, including constraint for sparse representation and constraint for nonlocal self-similarity, then using mm SFSR algorithm to solve the optimization problem. Experiments conducted on simulated and clinical MR images show mmSFSR algorithm performance better than classical interpolation, cpSFSR algorithm and nonlocal means reconstruction algorithm based on multi-modality priors in terms of PSNR, SSIM, intensity profiles and visual inspection.(3)We proposed an orthogonal MR images based multi-frames super resolution reconstruction algorithm(oMFSR). oMFSR algorithm exploits the correspondence between the high-resolution slices and the low-resolution sections of the orthogonal input scans as well as the self-similarity of each input scan to train pairs of over-complete dictionaries that are used in a sparse land local model to upsample the input scans. So oMFSR algorithm does not require extra training sets. We firstly utilize the sparse representation theory to solve the multi-frames super resolution reconstruction problem. The proposed framework includes three steps: upsampling each orthogonal scan by sparse representation algorithm, fusing resampled orthogonal scans using 3D wavelet fusion, and reconstruction constrained by error back-projection. Experimental results show that MFSR algorithm could fuse all the details information from all the input low resolution orthogonal 3D MR images and gets much better reconstruction result than SFSR algorithm, even though MFSR algorithm requires more computational time. Furthermore, oMFSR algorithm outperforms cpSFSR, the averaging of upsampled scans and the wavelet fusion method in terms of PSNR, SSIM, intensity profiles, and visual inspection.