Abstract: Liver vascular segmentation is significant in assisting physicians in clinical diagnosis, liver resection, and transplantation. Aiming at the problems of diversified liver vessel features and the influence of contrast agent and lesion area on the segmentation results, we propose an adaptive liver vessel segmentation network with progressive feature fusion, APFF-Net. Firstly, we focus on the global and local information of liver blood vessels by gradually fusing different levels of multi-scale feature information to better capture the morphology and distribution of blood vessels and realize fine-grained segmentation. Secondly, we design an adaptive semantic guidance module to refine the low-level features with high-level features to better distinguish liver blood vessels from other tissues so as to make a more accurate prediction for up-sampling. Finally, hollow space pyramid pooling is used in the bottom layer so that the network performs sensitive segmentation of fine blood vessels in the liver without weakening the ability to recognize coarse blood vessels, which helps to suppress the overfitting phenomenon. The sensitivity and accuracy on the 3Dircadb dataset reached 76.36% and 98.93%, respectively, and on the MSD dataset reached 68.32% and 99.85%, respectively, which were better than those of the other compared methods.