Automatic Liver Segmentation from Abdominal CT Volumes Using Graph Cuts and Border Marching
Miao Liao,Yu-qian Zhao,Xi-yao Liu,Ye-zhan Zeng,Bei-ji Zou,Xiao-fang Wang,Frank Y. Shih
DOI: https://doi.org/10.1016/j.cmpb.2017.02.015
IF: 6.1
2017-01-01
Computer Methods and Programs in Biomedicine
Abstract:BACKGROUND AND OBJECTIVE:Identifying liver regions from abdominal computed tomography (CT) volumes is an important task for computer-aided liver disease diagnosis and surgical planning. This paper presents a fully automatic method for liver segmentation from CT volumes based on graph cuts and border marching.METHODS:An initial slice is segmented by density peak clustering. Based on pixel- and patch-wise features, an intensity model and a PCA-based regional appearance model are developed to enhance the contrast between liver and background. Then, these models as well as the location constraint estimated iteratively are integrated into graph cuts in order to segment the liver in each slice automatically. Finally, a vessel compensation method based on the border marching is used to increase the segmentation accuracy.RESULTS:Experiments are conducted on a clinical data set we created and also on the MICCAI2007 Grand Challenge liver data. The results show that the proposed intensity, appearance models, and the location constraint are significantly effective for liver recognition, and the undersegmented vessels can be compensated by the border marching based method. The segmentation performances in terms of VOE, RVD, ASD, RMSD, and MSD as well as the average running time achieved by our method on the SLIVER07 public database are 5.8 ± 3.2%, -0.1 ± 4.1%, 1.0 ± 0.5mm, 2.0 ± 1.2mm, 21.2 ± 9.3mm, and 4.7 minutes, respectively, which are superior to those of existing methods.CONCLUSIONS:The proposed method does not require time-consuming training process and statistical model construction, and is capable of dealing with complicated shapes and intensity variations successfully.