Seismic　imaging is　a geophysical technique　to　investigate　the　subsurface　based　on　seismic　wave.　Seismic　image　is　a　crucial　channel　to　acquire　salt　deposits　positions　which　are　important　evidences　to　locate　oil　and　gas　resources　area　underground.　Although　the　development　of　geophysical technology　makes　it　easier　to　collect　more　seismic　images,　difficult　wave　shapes　and　structures　in　images　still　hinder　the　use　and　analysis　of　these　valuable　data.　In　this　paper,　we　adopt　a　novel　deep　learning　method　to　solve　this　issue.　An　individual　decoder,　named　boundary　decoder,　is　introduced　in　U-Net-based　framework　to　improve　segmentation　performance　of　seismic　images.　Boundary　decoder　is　supervised　by　labeled　salt　body　boundaries　and　further　assists　original　segmentation　decoder　to　generate　more　precise　results　with　its　boundary　priori　knowledge　guidance.　As　a　result,　we　called　this　novel　network　as　boundary　U-Net　(BU-Net).　In　experimental　evaluation,　the　proposed　approach　can　increase　IoU　by　2.6%　and　optimize　the　recognition　of　semantic　boundary　and　details　of　salt　bodies　greatly.　©　2021,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.