In this paper, we propose an automated liver segmentation method to overcome the challenging issues of high degree of variations in liver shape / size and similar density distribution shared by the liver and its surrounding structures. To improve the performance of conventional statistical shape model for liver segmentation, in our method, the signed distance function is utilized so that the landmarks correspondence is not required when performing the principle component analysis. We improve the Chan-Vese model to bind the shape energy and local intensity feature to evolve the surface both globally and locally toward the closest shape driven by the PCA. In our experiments, 20 clinical CT studies were used for training and 25 clinical CT studies were used for validation. Our experimental results demonstrate that our method can achieve accurate and robust liver segmentation from both of low-contrast and high-contrast CT images.