The Level Set Method And Application Of Medical Image Segmentation Incorporating Anatomical Knowledge

With the rapid development and wide application,modern medical imaging technology has become the most important means of auxiliary diagnosis of clinical diseases.With the continuous upgrading and performance improvement of these imaging technologies,computer-aided diagnosis has become an important step in doctors’ daily work such as clinical diagnosis and treatment plan determination,among which the key problem is medical image segmentation.This dissertation discusses the basic theory of medical image segmentation,lists several classical traditional methods of medical image segmentation and deep learning methods at present,and explains in detail their advantages and disadvantages and complementarity in image segmentation.This dissertation focuses on the basic principles and theories of the level set method,and integrates the anatomical knowledge into the level set method,and puts forward several image segmentation methods based on anatomical geometric features,so as to improve the accuracy,automation and robustness of segmentation algorithm.These methods effectively make up for the limitation that data-driven deep learning methods are difficult to integrate knowledge,and provide an effective model for the research of knowledge-driven medical image segmentation methods.This dissertation systematically discusses the research topics,and each method is interrelated and progressive.The original work of this dissertation mainly includes the following contents:1.A level set variational model with variable scale local approximation and integration is proposed to eliminate the inherent defects of the famous RSF model and make the algorithm more efficient and stable.This method is applied to the CMA images of cerebral vessels and CBCT images of teeth,and satisfactory segmentation results are obtained.2.Integrating the features of double-layer anatomical structure into the region-based level set method,a double-layer level set segmentation model is proposed to realize the automatic segmentation of white matter and gray matter of brain structure.The above method provides a new idea for the research of subsequent methods,and can also be used as the initial level set function of other level set methods.3.This dissertation puts forward the convexity preserving level set evolution(CPLSE)for left ventricle segmentation,makes full use of the anatomical geometric characteristics of the left ventricle,and overcomes the interference of trabecula and papillary muscles to the segmentation algorithm.The convexity of the level set contours is controlled by the curvature,so that the contour of the level set function can be finally deformed into a convex structure which meets the requirements of clinical anatomy.This method is general and can be applied to the segmentation of various convex objects.It is also instructive to the segmentation for non-convex objects.4.This dissertation puts forward convexity preserving bi-layer level set(CP-BILLS)model.During the CP-BILLS evolution,the 0-level set and k-level set simultaneously evolve and move toward the true endocardium and epicardium under the guidance of image information and the impact of the convexity preserving mechanism as well.Moreover,this model develops an algorithm for automatic selection of an optimal k-level,which eliminates the manual selection of the k-level.As a result,the obtained endocardial and epicardial contours are convex and consistent with the anatomy of cardiac ventricles.5.This dissertation proposes convex shape decomposition(CSD)structure of cardiac ventricles based on anatomical knowledge to realize the segmentation of RV and myocardium of both left and right ventricles.In order to improve the degree of automation of our algorithm,this dissertation combines the deep learning method with the level set method,makes full use of the advantages of the deep learning method to automatically extract image features,instead of manually setting the initial level set function in the level set method,and realizes the full automation of the segmentation model.This dissertation makes full use of the advantage that the level set method can be easily extended to high-dimensional space,and expands the level set method with anatomical knowledge into three-dimensional space to realize the segmentation of three-dimensional image data.As an aid and supplement of cardiac magnetic resonance image segmentation,this dissertation provides a solution to the problem of slice misalignment in cardiac image data that needs to be solved urgently in clinic.