Segmentation Of Pulmonary And Intestinal Lesions Based On Multiscale Theory

Lung cancer and colorectal cancer are the two major malignant tumors with high mortality rates,and they are often characterized by nodules and polyps.The segmentation of pulmonary nodules and intestinal polyps,as the basis of key steps in character judgment,volume measurement,doubling rate tracking,and so on,can provide reliable data support in the stage of disease analysis and treatment.Besides,accurate results of lesion segmentation can also improve the accuracy of image processing methods such as lesion feature extraction,benign and malignant judgment,three-dimensional reconstruction and visualization.Therefore,whether from the perspective of improving the level of medical information or from the perspective of promoting the development of medical image processing-related research,it is of great significance to conduct in-depth research on the segmentation of pulmonary nodules and intestinal polyps.This dissertation aims at the problems that restrict segmentation performance such as transition regions,abundant textures,uneven pixel brightness distributions and the complex surfaces in the lesion images of pulmonary nodules and intestinal polyps.The relationships between lesion features and scales were analyzed according to the lesion characteristics of two types of medical images.To achieve stable and accurate lesion segmentation,the research has been conducted on the edge-preserving smoothing,pixel similarity,edge features and regional pixel brightness/color distribution of the lesion images at different scales.The main contents and achievements are as follows:(1)In medical image analysis,the extraction and diagnosis of lesions are based on appropriate scales.This dissertation analyzed the relationship between lesion brightness,contour,pixel distribution and analysis scales,and proposed a multiscale smoothing model for medical images from the perspective of pixel changes.By analyzing the relationship between the diffusion performance of smoothing kernel functions and edge-preserving smoothing in the image local structure information,two ideal conditions that need to be met for edge-preserving smoothing are summarized.On this basis,the problems of edge-preserving smoothing of several classical smoothing kernel functions were analyzed and discussed,and an improved linear smoothing kernel function was designed in combination with the spatial characteristics of vision.To make up for the problem of the step effect in the classical kernel functions,this dissertation designed two nonlinear smooth kernel functions based on the relaxation of the edge-preserving smooth condition and analyzed the diffusion performance of the two functions.Experimental results shown that the improved linear smoothing kernel function in this dissertation could improve regional consistency and had a lower computational cost.While the two nonlinear smoothing kernel functions could preserve the edge information of lesion effectively and smooth the uneven textures and brightness distributions in lesion images.(2)Aiming at the problem that the accuracy of pulmonary nodules segmentation is easily affected by pixel similarity,this dissertation proposed a DR pulmonary nodules segmentation model based on multiscale pixel similarity.In this model,the cohesiveness of the pixel inside the pulmonary nodules and the difference of tissue was analyzed,and the pattern representation based on the content of pulmonary nodules was designed by using structural characteristics of pulmonary nodules.The transfer of boundary conditions was given in this pattern,in overcoming the nodule segmentation failures caused by the transition region to some extent.To improve the similarity of pixels and overcome the influence of texture and regional inconsistency on segmentation,the smooth energy term of segmentation was constructed with the combination of the linear smoothing kernel function proposed in this dissertation,and the scale component for multi-scale information fusion was obtained.In the solving process of the segmentation model,this dissertation combined the idea of finite elements to divide the bounded domain and discretize the continuous problem through theoretical derivation to obtain the numerical solution of the segmentation.The model had good performance of accuracy and stability in lesion segmentation and has strong robustness,especially when processing the lesion samples that contain transition regions between pulmonary nodules and the surrounding tissues.(3)Aiming at the problem that the segmentation results deviate from the nodule edge caused by the abundant textures and uneven pixel brightness in the pulmonary nodules lesion and surrounding tissues,a DR pulmonary nodules segmentation model was proposed based on multiscale edge.In this model,the multiscale characteristics of the edge were analyzed and combined with the curve evolution theory,the evolution curve was driven to converge to the lesion contour neighborhood according to the finiteness of the geometric measurement to the pulmonary nodule lesions.To overcome the convergence shift caused by the abundant texture and uneven pixel brightness distributions in the lesion and surrounding tissues during the curve evolution,the edge-preserving smoothing energy term was constructed to improve the accuracy of curve positioning by using the edge-preserving smoothing kernel function improved in this dissertation.To improve the efficiency of curve evolution and reduce the time cost of segmentation,an improved lightweight image pyramid is designed to provide a conformal initialization curve for pulmonary nodules segmentation.Besides,to solve the problem of the reduction of segmentation accuracy or even the disappearance of segmentation curve caused by over-smoothing,this dissertation analyzed the segmentation index of the adjacent scale segmentation results and designed the segmentation scale selection function to realize the lesion segmentation on the appropriate scale.The model could overcome the influence of abundant textures and uneven pixel brightness distributions and had a good performance in segmentation accuracy.(4)Aiming at the problems of low segmentation accuracy and poor stability caused by the complex lesion surfaces and complex external structures in intestinal polyp images,an intestinal polyps segmentation model based on multiscale mixed features was proposed.In this model,the edge features of different scales and appearance models were combined to construct the energy functional of segmentation.To overcome the effects of intestinal mucosal texture,food residues,and tissue lesions on the accuracy of the surface model,the improved edge-preserving smoothing kernel function was used to suppress the inconsistency of regional pixel brightness/color,and combined with the histogram shape analysis,the number of sub-regions of the image content was obtained and the parameter estimation of the surface model was optimized.Besides,to obtain the segmentation results on a suitable scale,the similarity of two adjacent intestinal polyps segmentation was defined and the segmentation termination conditions were designed according to the similarity changes.Based on inheriting the low-level features of intestinal polyp images,this model expanded the scale information of features.When processing intestinal polyp lesions with complex surface conditions,including textures and weak edges,it could still obtain relatively accurate segmentation results.