Multi-organ Segmentation Of Medical Images Based On Prior Knowledge

With the popularization of medical image analysis technology in medical research and clinical diagnosis and treatment,many clinical applications need to quickly segment multiple organ regions from medical images to assist the simulation of multiple organs and morphological and functional measurements.Compared with single organ segmentation,the simultaneous segmentation of multiple organs not only improves the speed of the algorithm,but also enhances the segmentation accuracy and robustness of the algorithm by virtue of the interdependence of morphological and positional relationships among multiple organs.The purpose of this thesis is to develop an algorithm for segmenting multiple organs simultaneously from a medical image,and to improve the accuracy and robustness of the segmentation algorithm by using prior knowledge of shape and position among multiple organs.In this thesis,mouse CT images are selected as target images because of the low dose of X-rays and poor soft tissue contrast.The multi-organ segmentation is an urgent problem to be solved.The main work of this thesis is shown below:Firstly,in order to provide prior knowledge for subsequent image segmentation,the thesis studies how to build a multi-organ shape model.This thesis expands on the existing multiresolution statistical shape model,and solves the problem that existing methods cannot model three-dimensional shape of multiple organs due to memory limitations.Also,this thesis proposes multi-resolution shape for multiple organs based on prior knowledge.This method can model the morphological changes of multi-organ systems from three resolution levels,including multiple organs,single organs,and local organ ranges.Experimental results show that the multi-resolution multi-organ shape modeling method presented in this thesis has better generalization and specificity than traditional methods.Secondly,combined with the multi-resolution statistical shape model constructed in this study,low-dose multiple organs in mouse CT images are segmented.Due to the serious lack of soft tissues contrast and the blurred edge in low-dose CT images,this study combines user interaction guidance and shape prior knowledge to achieve multi-organ segmentation.The Variational Hermite Radial Basis Function is used to implement organ morphology estimation based on a small amount of user delineation and guide the multi-resolution shape model for further refined segmentation.Experimental results show that the proposed method achieves better segmentation accuracy than the traditional statistical shape model,and has good consistency for different user operations,achieving rapid segmentation of multiple organs in low-dose CT images.Finally,a multi-organ shape modeling method based on artificial neural network is further studied based on the previous parts of this study.The Stacked Autoencoder network was used for nonlinear shape modeling.Compared with the traditional linear Principal Component Analysis method,the nonlinear shape modeling based on Stacked Autoencoder neural network can automatically learn the multi-resolution deformation features and achieve the similar effect of the modeling method in Chapter 3,but the algorithm is of less complexity.This method can also learn the deformation features that traditional linear modeling methods cannot learn.