Automatic Liver Segmentation Based on Shape Constrained Differeomorphic Demons Atlas Registration

Abstract

Liver segmentation from abdominal Computer Tomography (CT) images plays an important role in liver disease diagnosis as well as liver surgical planning. In this paper, a hybrid approach is proposed for fully automatic liver position search and liver segmentation in CT images. First liver intensity range is detected based on prior knowledge of liver volume. Then region of interest (ROI) is extracted using atlas-based affine and non-rigid registration. At the last step, to achieve more accurate segmentation, major liver tumors are detected using gray level and distance prior knowledge, and then a modified diffeomorphic demons registration with shape constrain is applied. Thirty CT image datasets are tested, and the effectiveness is evaluated using volume mean overlay coefficient (Dice) and a comprehensive metric. Results show that our method can be a potential tool in clinical application.