Research And Application On Medical Image Reconstruction

Medical imaging usually refers to the technology and process of generating images of the subjects’ internal anatomy for diagnostic and therapeutic purposes during a digital health examination.Common medical imaging includes anatomical imaging and functional imaging.The images generated by the technology can provide the tissue,structure,and function information of the subject.Medical image diagnosis is one of the most important diagnostic bases in clinical data.Generally,doctors need to determine whether there is a lesion area in the subject with the help of medical images to evaluate and grade the disease according to the degree of the lesion severity.In the process of image acquisition,the medical images obtained due to the gravity of the samples,scanning cost,improper operation,or other factors,may cause physical tearing of organs or tissues,low resolution or disordered sequences,resulting in image degradation.Medical image reconstruction aims at different degradation reasons,taking advantage of algorithmic thinking and computing power,and through a series of image processing technologies,the degraded image can be reconstructed into a complete image under ideal conditions.The reconstructed medical image can be used for research modeling and quantitative analysis in the field of classification,detection,and segmentation,providing a clinical reference for data annotation and statistical analysis,and achieving the purpose of improving the accuracy and efficiency of diagnosis.Therefore,medical image reconstruction plays an extremely important role in medical image processing and analysis.Image registration,interpolation,and matching are important methods for medical image reconstruction related applications,which are conducive to integrating the image information of the subjects in various aspects,improving the image resolution and restoring complete images,as well as providing technical support for disease prediction and analysis,auxiliary diagnosis and so on.In the process of solving practical problems in clinical practice,due to the particularity of the acquisition and applications of medical images,including the complex requirements caused by the operation process,the acquisition time,and the workload of doctors,the solution process is usually combined with multiple tasks such as registration,interpolation,and matching.Image interpolation is an indispensable key step in registration.Registration is conducive to the improvement of matching accuracy.The matching result can also be used in the process of interpolation and registration.To solve the practical problems in the clinical basic research of medical imaging,aiming at the problems faced in the task of image reconstruction,based on the key technologies in the field of image processing and analysis,the main research contents of this paper include the following parts.(1)Piecewise registration of rodent light sheet fluorescence images for anatomical differences reconstruction.In the process of collecting the rodent light sheet fluorescence images,the gravity of the mouse brain or the careless operations can easily lead to the physical tearing of organs or tissues.According to the above problem,this thesis proposes an image reconstruction method based on piecewise registration of MRI and light sheet fluorescence images by making use of the advantage that MRI can be used in intracranial or in vivo scanning.Considering the consistency of the multimodal images in the anatomical structure of the mouse brain,as well as the atlas and labeled images in Waxholm space,the random walk and morphology methods are used for unsupervised segmentation of MRI and light sheet fluorescence images,respectively.The registration is applied to the segmented parts of MRI and light sheet fluorescence images to obtain the global deformation fields.The image fusion stage combines the mask based on the MRI and the Laplace equation to fuse each part of the deformation fields and obtain the restored light sheet fluorescence images through image warping and resampling.This method solves the problems of unsupervised segmentation and image fusion of multimodal images in piecewise registration and has been preliminarily applied in intra-and inter-subject registration of mouse brain,providing an idea for the subsequent processing and analysis of rodent light sheet fluorescence images.(2)Multimodal brain MRI reconstruction based on spatial information and interpolation.In the acquisition process of brain T2 Flair-weighted MRI,due to the limitations of imaging conditions,including scanning time,different imaging sequences,or acquisition costs,usually only three-dimensional MRI data with high-resolution can only be acquired at a certain anatomical plane.According to the above problem,this thesis proposes a multimodal image reconstruction method based on spatial information and interpolation,taking advantage of the short scanning time of T1-weighted MRI.This method combines the multi-modal information between the T2 Flair-weighted MRI with high thickness and the T1-weighted MRI with low thickness,introduces the spatial constraint relationship,and combines the conditional generative adversarial network to construct a high-resolution image reconstruction algorithm.In terms of visualization and quantitative evaluation,this method has been proved to be able to effectively improve the resolution of an anatomical plane,providing high-resolution data for subsequent work such as segmentation of lesion regions or disease analysis.(3)Histopathological image reconstruction based on image matching.In the multimodal imaging research related to human histopathological images,since histopathological imaging only slices local tissues,it cannot meet the needs of combining high-resolution images of other modalities to analyze the correlation between images on a global scale.According to the above problem,this thesis proposes a histopathological image reconstruction method based on image matching,which integrates the potential idea of non-overlapping jigsaw puzzles.This method is based on the assumption that the histopathological slice is rectangular,and constructs a fragment reconstruction task with unknown orientation and permutation.Combining the advantages of the Mahalanobis gradient distance and the projected power method,as well as using the quadratic programming model,this method transforms the combinatorial optimization problem between fragments into a constrained non-convex optimization problem and establishes an image matching model.The Mahalanobis gradient distance is used to calculate the adjacent relationship of the fragments,which not only considers the intensity information but also introduces the gradient information of the boundary pixels,which improves the accuracy of the measurement.The projected power method can represent the adjacent relationship of the fragments in the highdimensional space,combined with the random initial sorting status,using the iterative power method and space projection,to further ensure that the optimization algorithm converges to the optimal solution in the high-dimensional space,and reduce the time of image reconstruction and restoration.In terms of visualization and quantitative evaluation,the proposed method verifies the effectiveness of the image matching algorithm with non-overlapping unknown orientation and permutation fragments and provides a potential idea for subsequent reconstruction of irregular fragments and multimodal image analysis based on histopathological images.In summary,this thesis starts from the actual needs of medical image diagnosis and basic research,designs models and algorithms according to the characteristics of the datasets and the problems to be solved,to achieve the purpose of medical image reconstruction tasks.The actual needs include the reconstruction of physical tearing from the light sheet fluorescent images of rat brain,the high-resolution reconstruction of T2 Flair-weighted brain MRI,and the matching of human histopathological images.The characteristics of the dataset problems include medical image dimensions and different modal information complementation and integration,the different modal image acquisition errors,various individual physiological structural differences,3D image spatial information problems,image intensity values or color distribution and scope,the different organisms due to structural differences significantly the effects on the algorithm,deep learning data augmentation,annotation,evaluation standard,and the gold standard.The models and algorithms include the random walk,morphological transformation,atlas introduction,affine and elastic registration,nearest neighbor search,Laplace equation,conditional generative adversarial network,and so on.The experimental results in visual and quantitative evaluation show that the model and algorithm proposed in this paper can effectively accomplish the task of medical image reconstruction,and contribute to solving the practical problems faced by medical imaging in basic clinical research and analysis.