Research On Hybrid Priors Based Medical Image Reconstruction And Deformable Registration

Medical imaging technology creates opportunities for non-invasive diagnosis and is a research hotspot in the medical field.The medical image processing technology involved mainly includes image forming process and image computing process.During the formation process,compressed sensing Magnetic Resonance Imaging(CS-MRI)allows a sampling rate much lower than the Nyquist standard,significantly accelerating the data acquisition process.The reconstruction algorithms aim at recovering an artifacts-free image from the sparse observation in k-space.Deformable registration is a very challenging task in the image calculation process.During the computation process,deformable registration is a very challenging task which aligns multiple images into the same space,and has great significance for researches like longitudinal analysis of the evolution of pathological structures.Both of these two different image processing tasks can be abstracted into a typical energy optimization problem.To solve this kind of problem,traditional methods are devoted to designing and optimizing energy functions containing specific physical priors,but the iterative numerical optimization brings an expensive calculation load and a long process.Deep learning techniques usually encode information in a large number of data samples to directly learn the mapping from the input to the target,or embed the deep architecture into the optimization to speed up calculations.However,severely depending on the selection of training samples and lacking domain knowledge constraints,these learnable mechanisms fail to guarantee the authenticity of the processing results,thus the performance in applications of complex scenes may be significantly reduced.This paper integrates the physical priors with learnable deep optimization mechanisms,respectively proposing reconstruction and registration algorithms based on hybrid priors.1)For reconstruction,this paper integrates the learning prediction with the energy optimization,and further embeds an error controlling mechanism to ensure the convergence,yielding a theoretically guaranteed CS-MRI reconstruction framework.Considering the complex real-world scenarios,this paper develops three extensions to reconstruct Rician noise polluted data,complex form data and multi-coil data.Experiments on the benchmark data and the raw data show that the proposed algorithm can efficiently and reliably converge to the optimal solution of the original problem.2)For deformable registration,this paper proposes a principled regularized bidirectional deep deformable registration,which cascades principled regularization into the deep estimation processes to inherit the advantages of both domain knowledge and data information.Experiments in two different registration scenarios show that the proposed paradigm provides a more efficient,accurate and stable deformation field estimation than the existing technologies.