Registration Of Multi-phase Liver CT Data Using Edge Textures

The purpose of liver registration is making an alignment of anatomic structures in multi-phase images.In this way,different pathological information is integrated,which helps doctors to make an accurate diagnosis or perform a surgery.Landmark extraction is a crucial step in liver registration.Since anatomical features of a liver is hard to be detected automatically by computer,most of the feature points are manually labeled by radiologists,which takes time and often exists certain human error.Additionally,feature points are mainly located in spine and the intersection of blood vessels while only a few of them are located in the liver region,which is unsuitable for liver registration.Therefore,this paper presents a fully automated landmark detection method to register livers on multi-phase CT images.To achieve that,liver segmentation is used to extract center points of a liver;edge texture features and Support Vector Machine(SVM)are applied to detect discriminated surface landmarks on the liver.The main contents of the proposed method are as follows:(1)Automatic detection of internal landmarks.Edge detection and 3D-labeling algorithms are employed to extract a liver in two phase images.By comparing with the average liver and adjacent slice,false positive points(FPs)are eliminated.Finally,based on the largest slice of the liver,every slice is matched and center points of corresponding slices are used as internal landmarks.(2)Automatic detection of surface landmarks.Firstly,3D liver model is built to analyze edge features and 5 representative surface points are chosen.Secondly,3-dimensional gray level co-occurrence matrix(3D-GLCM)is built and 17 texture features are extracted to represent the edge information of different surface points.To analyze the differences of textures,features are extracted both before and after the liver segmentation respectively.Results show that liver segmentation is helpful to expand the differences of features between every surface point,which makes the division boundaries wider.Thirdly,leave-one-out method and SVM classifier are applied to optimize texture features.Combination of ASM,COR and IDM makes contribute the most to the accuracy of classification.When applying this combination to surface point classification,it obtains 98.5%accuracy.The result demonstrates that edge textures and SVM classifier are essential to the automated landmark detection.Finally,elimination of FPs is implemented by center coordinates of the liver.Further,matching points in two phase images is considered.(3)Registration of a liver in two phases.Based on both internal and surface detected points,thin plate splines(TPS)deformation is employed to register a liver.After registration,surface-based mean error decreases by 0.93mm,which indicates that the automated TPS algorithm is able to handle the liver registration.