Liver Vessel Segmentation Based On 3D Deeply Supervised Network

Liver vessel segmentation is a key step for liver-disease diagnosis, therapy, and liver surgery. This paper proposes an automatic liver vessel segmentation method. First, a 3D deeply supervised network is efficiently applied to generate vessel prediction probability and initial segmentation. Then, to further learn vessel intensity feature, an intensity model of liver vessel is built according to the initial segmentation. Finally, the graph cuts combined with the prediction probability and intensity is used for the ultimate liver vessel segmentation. Our method avoids the use of traditional vessel filter to generate hand-craft feature, which usually relies on complex mathematical theory to form the geometrical constrains, and does not need manually label the initial object regions for each CT volume. Our experiments show that the proposed method obtains an accuracy 97.0%, sensitivity 78.5%, specificity 97.6% and RMSD 3.69 mm, and receives satisfying segmentation results of thin vessels.