Abstract: The performance of neural network-based methods for medical image segmentation is still unsat-isfactory,mainly due to the limited generalization of neural networks,the unbalanced quality of medical images,and irregularities and infiltration of tumors.To sufficiently utilize image-specific information,we propose prior-embedded networks(PEN).By introducing image-specific spatial priors into neural networks,PEN focuses on lesion regions to learn discriminative features and ignore unrelated information,and there-fore extracts crucial features for improving segmentation performance.Two medical image segmentation frameworks,2D U-Net and 3D nnU-Net,are used as backbone networks,whose performance is evaluated in the liver tumor segmentation task on the LiTS data set.With five-fold cross-validation,the results show that the PEN significantly improves segmentation performance by 22.4%on the training set compared with U-Net,and by 1.2%and 4.4%on the test set compared with ensemble nnU-Net and single nnU-Net,respec-tively.