Manifold Regularization And Spasrse Coding Based Medical Images High-Resolution Reconstruction

Medical image high-resolution reconstruction technology can convert a low-resolution image to a high-resolution image through some technologies, and can provide more detailed information for doctors to diagnose, which can effectively improve diagnostic accuracy and the correct rate. Since it does not involve hardware and low cost, and it has become a hot spot in recent years. This thesis mainly analyzes and researches several classic high-resolution methods, by introducing manifold regularization and integration algorithm, manifold-regularization sparse coding medical image high-resolution reconstruction and integrated high-resolution reconstruction algorithm based on broad-learning are proposed, and these algorithms are applied to the medical image to verify their validity. Specific work are as follows:1. The classical sparse coding-based high-resolution image reconstruction algorithms cannot preserve the image local smoothing structures, so a novel manifold-regularization sparse coding medical image high-resolution reconstruction algorithm is proposed. Improved algorithm introduces the manifold regularization to constrain the image patches lay on the intrinsic smoothing manifold, and incorporate the image nonlocal self-similarity into sparse representation model to improve the accuracy of the sparse coefficients, which can make the coefficients obtained closer to its true value. And in the thesis we also adopt a new approach to find similar patches, making a similar patch can be guaranteed the accuracy and smoothness of the edge, and finally apply it to the CT and MRI medical image experiments. Through analyzing the reconstructed high-resolution images and the objective evaluation value obtained from experiment for 100 medical images, it is concluded that high-resolution reconstruction algorithm proposed can effectively improve the quality of the image, and has a better visual effect.2. Classic high-resolution reconstruction algorithms based on learning have their own advantages, such as high-resolution through neighbor embedding can reconstruct a smooth partial structure; Image high-resolution via sparse representation can reconstruct a good local texture structure; Manifold-regularization high-resolution image reconstruction can be implemented locally smooth and showed good visual effects and so on. In order to make full use of the advantages of reconstruction image methods, integrated high-resolution reconstruction algorithm based on broad-learning is proposed. By neural network(broad-learning) which can add adaptive hidden nodes to improve the fitting accuracy, this algorithm learns the regional proportion in the reconstructed image, and low quality image is integrated to high quality images by the image patch and its corresponding weights. Finally we apply it to the medical image experiments, 50 brain CT image reconstruction results reveal that image constructed through broad-learning integration algorithm has clear texture and a better visual effect.3. Design and implement high-resolution medical image reconstruction system. The system integrates algorithms proposed in this thesis and other relates algorithms, it can reconstruct image by different algorithms, and display the high/low resolution image through the system interface, compared to using singal algorithm to reconstuct high-resolution image, this system is more convenient and flexible.