Medical Image Segmentation Based On Level Set Methods And Fuzzy Models

Currently, medical imaging becomes crucial to clinical treatment. However, due to low contrast, individual differences and multi-modalities in medical images, segmentation of medical images is still an open problem. According to the statistics, cardiovascular disease causes high fatality rate and the segmentation accuracy of ventricles on MRI which is an important non-invasive evaluation method influences highly on diagnoses. Therefore, this paper proposes three models based on level set methods for segmenting cardiac ventricles from MRI. In addition, cancer is also one of the highest causes of death. The evaluation of lymph nodes on CT and PET/CT images is important to diagnose, staging, surgical planning and prognosis during cancer treatment. We have made definition on lymph node zones for pelvis and abdomen, and then, we use AAR fuzzy model to recognize these lymph node zones. Last, we have detected and segmented lymph nodes in those lymph node zones based on shape information and level set method.First, we use convexity preserving level set method to segment endocardium of left ventricles on MRI. This model can keep endocardium contours as convexity, so it can prevent evolution of level set formulation from local convergence due to papillary muscles. When the contour is convexity, the coefficient of convexity preserving term in this model becomes 0, the contour is controlled by data term of the image itself. When the contour is concave, the coefficient of data term becomes 0, the contour is controlled by convexity preserving term and reach the true endocardium at last. Experimental results show that our convexity preserving level set model can extract endocardium contour from MRI left ventricles accurately.Second, we extend DRLSE model to a two-level-set DRLSE model to segment endocardium and epicardium of left and right ventricles from MRI simultaneously. We use DRLSE model as preliminary segmentation, and the contour comes out from preliminary segmentation will be used as initial contours of two-level-set DRLSE model. The 0-level and k-level contour of level set method represent endocardium and epicardium of left and right ventricles, respectively. The energy of two-level-set can keep 0-level and k-level contour smooth and the distance between them are varying stably. Through testing on MICCAI 2009, MICCAI 2012 and MICCAI 2013 cardiac MRI data sets, we can conclude our two-level-set DRLSE model is efficient for segmentation of left and right ventricles of MRI.Third, we use DR2 LS model to segment endocardium and epicardium of left and right ventricles from MRI. After applying DRLSE model as preliminary segmentation, we can get endocardium contour as initialization of DR2 LS model. Endocardium and epicardium are represented as 0-level and k-level contours of level set formulation, respectively. The data term can solve the intensity inhomogeneity of MRI. The distance regularization term can keep the distance between 0-level and k-level contour of level set formulation varying steadily. Experimental results show that DR2 LS model which initialization contour can be achieved from DRLSE model, can solve intensity inhomogeneity of MRI and keep anatomical structure of the two contours and finally get desired segmentation results.Then, we have made definition on lymph node zones for abdomen and pelvis, totally 10 abdominal lymph node zones and 10 pelvic lymph node zones. We use AAR fuzzy model to construct lymph node zones model and test on different hierarchies to find out the optimal hierarchy, and use optimal hierarchy to recognize lymph node zones on CT images. In order to build optimal hierarchy, we combine different organs and lymph node zones. From recognition result, the localization accuracy appears within 2-3 pixels, which is acceptable.Last, we use optimal hierarchy to recognize lymph node zones on PET/CT images, and then detect and segment lymph nodes in lymph node zones. We use ball filter, threshold CT image, threshold PET image to do preliminary detection and use ball shape as lymph node contour. Then, we use dual-tree complex wavelet moment invariants to further detect lymph nodes and DRLSE to delineate lymph nodes edges. Comparing our SUV of segmented lymph nodes with SUV from ROVER software, we can conclude our method is efficient in order to segment lymph nodes.Through experiments of left and right ventricle segmentation on MRI and lymph nodes segmentation on PET/CT, we can summarize our proposed method based on level set methods and fuzzy models to segment medical images are promising.