Non-rigid registration of multi-phase liver CT data using fully automated landmark detection and TPS deformation

In case of the complicated anatomical structure of the liver, landmark points on a three dimensional (3D) liver surface is hardly distinguished as corresponding pairs visually and automated landmark placing will be extremely time saving for liver registration. This paper presents a fully automated landmark detection method to register livers on multi-phase computed tomography (CT) images. Edge texture features and Support Vector Machine (SVM) are applied to detect the discriminated landmarks of the liver, including both surface and internal points. Using the information of liver shape, 3D gray level co-occurrence matrix is calculated into texture features, from which the most informatics there features are selected by our optimization algorithm for choosing a sub-set of features from a high dimensional feature set. Then automated landmarks detection begins at scanning surface points on the pre-contrast and portal venous phase images, where positive outputs of the SVM classifier are regarded as initial candidates and final candidates are obtained by eliminating false positives (FPs). Finally, relied on the detected landmarks, thin plate splines (TPS) algorithm is used to register livers. Five surface landmarks, together with internal landmarks of the liver center from every 25 mm slice interval, can be detected automatically with sensitivity of 88.33% and accuracy of 98.5%. Surface-based mean error (SME) is decreased from 3.80 to 2.87 mm on average, while SME value has increased 32.4 and 8.0% on average respectively when comparing with the rigid and B-spline methods. The results demonstrate that edge textures and SVM classifier are effective in the automated landmark detection. Together with TPS algorithm, fully automated liver registration is able to be achieved on multi-phase CT images.