Research On Medical Image Segmentation Algorithm And Application

In recent decades,with the advancement of medical image scanning techniques and computer technology,various modality of medical images have been widely used and played more and more important roles in clinical.Medical image segmentation is the basis for disease detection and computer assisted diagnosis and treatment,such as tumor localization,pathological analysis,tissue volume measurement,computer aided surgery,computer guided surgical radiotherapy,development of treatment planning,anatomical structure research,etc.However,we can’t find a method that is able to achieve good segmentation performance for all modalities of images with complex target structures.Manual delineation is tedious and inefficient that adds heavy work to clinicians.Therefore,it is important to study effective computer-aided segmentation methods for different images in clinical application,which has been attracted highly attention of researchers at home and abroad.In this thesis,we mainly studied the gland segmentation from pathological images and vessel from retina images.We focused on the method of fractional scattering network,minimal path model with dunamic Riemannian metric and graph based perceptual grouping.Our main work and contributions are briefly outlined as follows:(1)An image segmentation algorithm was designed with the fractional scattering network(ScatNet).Based on the method to construct the complex wavelet scattering network,we generalized the traditional S to fractional order domain.We analyzed the translational invariance and small deformation stability of the fractional scattering network mathematically,meanwhile,we derived the the non-expansion of the energy propagation process in fractional scattering network.Since the traditional scattering network is a special case of the fractional scattering network,and different fractional transform domain can be obtained by adjusting the fractional orders,the proposed fractional scattering network has more potentional application.After that,a gland segmentation method from pathological image was designed with the proposed fractional scattering network.Finally,the effectiveness of the gland segmentation method by the proposed fractional scattering network in pathological image was validated in the experiments,and we also compared the classification accuracy of the fractional scattering network with different fractional orders.(2)An image segmentation algorithm was proposed based on the dynamic minimal path model.Based on the existing anisotropic minimal path model(or geodesic model),we proposed an anisotropic tubular minimal path model with fast marching front freezing scheme to avoid the shortcut problem and short branches combination problem.The crossing-adaptive anisotropic radius-lifted tensor field was proposed in the radius lifted space.The anisotropy of the metric was removed during the crossing points and its high anisotropy was kept on the points of non-crossing section.Besides,the non-local path feature,which only can be detected during the geodesic distance computation,was introduced into the metric to steer the evolution direction.Then the image segmentation algorithm based on dynamic minimal model was designed.Two initial points are given manually as the starting point and the end point.The minimal path between the two points and the radius information of the structure were obtained by the algorithm.Finally,the algorithm was valided quantitatively and qualitatively on retinal vessels,rivers and roads.We also compared our proposed method with other geodesic algorithms.(3)An interactive image segmentation algorithm utilized graph-based perceptual grouping was proposed.The graph-based perceptual grouping methd was intergrated by minimal path model and graph optimization.First of all,the appearance and orientation feature was detected by a matched filter to construct the vesselness map that also gave the optical direction of the vessel.Then the segments representing the vessel centerlines were obtained by non-maximum suppression on the vesselness map.After that,the graph was constructed,in which the the curve segments were regarded as the nodes and the minimal paths were used to connected the neighboring nodes.Finally,graph optimization was implemented by Dijkstra’s algorithm,and the optimal path between the given points by the user was searched in the graph.Since the curve segment was used instead of the image pixel as the node,the segmentation efficiency was improved.Besides,the smoothed vessels were extracted by connecting the gaps between the curved segments with a minimum path.We also compared our proposed method with other graph optimization based segmentation method.