Convolutions　neural　networks　have　obtained　promising　results　in　various　medical　image　segmentation　tasks.　However,　these　methods　ignore　the　problem　of　domain　shift,　which　will　lead　to　a　model　trained　in　a　source　domain　performing　poorly　when　applied　to　different　target　domains.　In　this　work,　we　propose　a　two-stage　segmentation　network,　and　utilize　histogram　matching　to　eliminate　domain　shift.　Specifically,　the　first　stage　obtains　the　region　of　interest　by　performing　coarsely　segmentation　on　down-sample　images.　Then　the　second　stage　segments　the　left　atrium　(LA)　based　on　the　region　of　interest.　The　method　is　evaluated　on　LAScarQS　2022　data-set,　acquiring　average　Dice　of　0.87790　for　LA　segmentation.　Besides,　the　two-stage　network　is　about　four　times　faster　against　a　single-stage　network　in　the　test　phase.　©　2023,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Switzerland　AG.