In order to make up for the shortcomings of the U-Net model and the level set method in segmenting liver CT images, this paper proposes a segmentation algorithm which combines U-Net model with the level set method. The algorithm uses the U-Net model as a prior network to segment the liver in the training phase, and then sends its output segmentation result as a prior feature map to the level set method for further segmentation; by calculating the difference between the two segmentation results, back propagation error, update network parameters, and finally get a complete segmentation algorithm model. Through experimental comparisons on the public data set 3Dircadb, this paper proposes that the sensitivity of the segmentation algorithm and the average Dice coefficient can reach 94.85% and 95.18%, respectively, and the average area overlap error and relative area error are only 9.97% and 7.69%. Compared with other common segmentation algorithms, this algorithm obtains better segmentation results.