Geometric Flows Based Medical Image Segmentation Methods And Applications

Medical Image segmentation is referred to extracting regions of interest from medical images, which plays a key role as the basic premise medical application of in medical image analysis and understanding, such as clinical diagnosis, pathological analysis of volume measurement, surface feature extraction and 3D reconstruction of human organs. In recent years, with the development of biological medicine imaging technology, more and more medical images are widely used, at the same time, traditional segmentation methods have been gradually unaffordable to the actual need of the increasing complexity of medical images. Therefore, research on medical image segmentation has been of great importance. With curve evolution theory, level set methods and information entropy principle as the theory foundation, this paper makes in-depth study on geometric flows based image segmentation methods, 3D medical image segmentation and assessment of image segmentation, which are the current research hot spots.Chan-Vese models segment very slow because each iteration process of evolution must be calculated in the whole region of image. In view of this defect, starting from theoretical analysis of the dynamic changes in regions, this paper analyzes the basis of its regional average gray value of the analytical formula for the progressive improvement of iterative formula, so that fast algorithms such as narrow band can be used. The new model greatly improve the efficiency of the segmentation, which makes the model much more practical.To further improve geometric active countours of the activities of leak of layer intermittent weak or marginal edge on the verge, a geodesic geometric flow with prior is presented by modificating edge detection functions through the introduction of information of neighbors as a priori to change the stop condition of curve evolution. This model brings local characteristics of the segmented regions into the adjacent and guides the convergence of contours to actual borders of objects, which improves accuracy and stability of geometric active contours.The conventional watershed algorithm vulnerables to the impact of noise and quantization error from heavy over-segmentation. GVF-Watershed algorithm is presented, which uses the scalar gradient map of gradient vector flow (GVF) as the input image. Since the gradient vector flow mutation of gradient information spreads slowly through the image, this algorithm can accurately detect the true edge of image, and by enhancing the immunity from noise and quantization error, reduce over-segmentation.To improve quality and efficiency of 3D medical image segmentation, this paper combines GVF-Watershed algorithm with models based on geometry flows to form the GW-LSM framework, the basic idea of which is to segment roughly with GVF-Watershed (GW) and then precisely with level set methods (LSM) based on narrow band. The former is faster to capture all edges, but easy to over segmented, while the latter has features such as topology changing adaptability and precision cutting, but is over calculated. Combining the two ones, they interact in a synergistic way. Within this framework, a 3D segmentation approach is presented for images those have very weak similarity among layers, which markedly enhances the efficiency of segmentation, and a 2.5D segmentation approach is presented by introduction of the similarity between the gradient as a priori for images those have strong similarity among the layers, which improves accuracy and speed of segmentation.To study of new and effective assessment methods of medical image segmentation and evaluate new approaches presentd in this paper, according to the nature of image segmentation and the special features of medical images, from the homogeneity of intra-regional perspective, this paper puts forward the concept of entropy segmentation, as well as an evaluation approach with the proposed segmentation information entropy as the measure, and from a perspective between the regional heterogeneity, this paper proposes an evaluation methodology based on the correlation regional entropy by introduction of the correlation information entropy as a nonlinear measure. The former is a relatively simple to calculated, while the latter gives an assessment in the unit closed interval. These assessments give both effective and accurate objective evaluation of the medical image segmentation, and will play a positive role in segmentation optimization and new methods studying.