Several Novel Medical Image Segmentation Methods Based On Variational PDEs

Medical image segmentation is a fundamental problem in medical image processing and analysis.Through image segmentation,anatomical structures and lesion areas can be locat-ed and one can also process feature extraction,quantitative analysis and visualization in these regions,so as to provide a reliable basis for clinical diagnosis and medical research.Image seg-mentation is the premise of image fusion and registration and reconstruction,also it is founda-tion of computer aid diagnosis,operation simulation and surgical navigation.However,medical images are usually corrupted with intensity inhomogeneities,weak boundaries and unwanted edges,which make the segmentation an inherently difficult task.Medical image segmentation is still an urgent and important problem to be solved.The main contributions of this dissertation are as follows:We propose a segmentation algorithm for medical images especially ultrasound images with intensity inhomogeneities and weak object boundary.Compared with traditional active contour models,the proposed algorithm has the following advantages.First,by restricting the energy functional on an evolving banded region,we can depress the influence of image intensity inhomogeneities on the algorithm.So the proposed algorithm can get promising segmentation results.Second,by maximizing the total intensity variation in the local patch along the normal direction of the zero level set,the edge descriptor can capture the visually invisible boundaries.The efficiency of the proposed model is demonstrated by experiments on real 3D prostate tran-srectal ultrasound images and 2D prostate ultrasound images.Experiments on 2D ultrasound images show that,compared with the manual segmentation results,our method can obtain a sensitivity of 94.87%± 1.85%,a DSC of 95.82%± 2.23%.Validation results on real 3D TRUS prostate images show that,compared with the manual segmentation results,our model can ob-tain a DSC of 94.03%±1.50%and a sensitivity of 93.16%± 2.30%.We focus on the segmentation of medical image with boundary dropout.We introduce a new formulation of shape constraint based on point-distance contour to describe shape priors.Then we incorporate it into the level set energy functional to get a point-distance based active contour model.The proposed point-distance shape constraint has the following advantages over the existing methods.First,it does not need to optimize the shape parameters or estimate shapes from training set.Second,it is more flexible in dealing with different shapes.It can handle various kinds of shapes like:circle,quasi-circle,ellipse,super-ellipse,squares and cardioid shapes etc.We test our algorithm on real ultrasound thyroid nodule images and real ultrasound kidney images.The validation results show our model can obtain a similarity of 91.14%±2.95%and a DSC of 92.78%± 1.90%on ultrasound thyroid nodule images and a similarity of 93.65%± 2.2%and a DSC of 94.96%± 1.22%on ultrasound kidney images.For the segmentation of images with intensity inhomogeneities,we propose a region inten-sity homogeneity energy functional based in the image of region intensity homogeneity factor.By incorporating the region intensity homogeneity energy functional into the active contour model,the proposed method can handle two typical kinds of images:one with homogeneous intensities in the objects and inhomogeneous intensities in the backgrounds,the other one with homogeneous intensities in backgrounds and inhomogeneous intensities in objects.The exper-iments on medical images and nature images show the advantage of the proposed method.