UNet　model　performs　well　in　medical　image　segmentation.　In　this　paper,　UNet　model　is　improved　by　the　same　padding　after　each　convolution,　so　that　the　image　scale　remains　unchanged　through　convolution,　and　the　edges　of　the　image　are　no　longer　cut　off.　The　improved　UNet　model　is　trained　for　semantic　segmentation　of　the　liver　in　the　portal　vein　in　CT　images,　using　binary　cross　entropy　as　the　loss　function,　and　dice　value　as　the　performance　evaluation　index.　The　average　dice　value　of　the　test　set　reaches　0.85.　Our　work　can　be　used　to　help　for　daily　work　of　liver　image　segmentation.　©　2020　IEEE.