Due　to　diversity　among　tumor　lesions　and　less　difference　between　surroundings,　to　extract　the　discriminative　features　of　a　medical　image　is　still　a　challenging　job.　In　order　to　improve　the　ability　in　the　representation　of　these　complex　objects,　the　type　of　approach　has　been　proposed　with　the　encoderdecoder　architecture　models　for　biomedical　segmentation.　However,　most　of　them　fuse　coarse-grained　and　fine-grained　features　directly　which　will　cause　a　semantic　gap.　In　order　to　bridge　the　semantic　gap　and　fuse　features　better,　we　propose　a　style　consistency　loss　to　constrain　semantic　similarity　when　combing　the　encoder　and　decoder　features.　The　comparison　experiments　are　done　between　our　proposed　UNet　with　style　consistency　loss　constraint　in　with　the　state-of-art　segmentation　deep　networks　including　FCN,　original　U-Net　and　U-Net　with　residual　block.　Experimental　results　on　LiTS-2017　show　that　our　method　achieves　a　liver　dice　gain　of　1.7%　and　a　tumor　dice　gain　of　3.11%　points　over　U-Net.　©　Springer　Nature　Switzerland　AG　2019.