The　3D　data　acquisition,　reconstruction　and　accurate　segmentation　of　teeth　are　of　great　significance　for　diagnosis　and　treatment　planning　in　dentistry.　A　3D　tooth　model　was　reconstructed　based　on　the　cone-beam-CT　(CBCT),　and　an　improved　GCNN　network　was　proposed　to　improve　the　semantic　segmentation　of　each　tooth　in　the　3D　tooth　model.　Accurate　semantic　segmentation　of　individual　tooth　can　be　achieved　by　deep　learning　of　the　fine　local　details　as　well　as　rough　global　structure　of　each　tooth.　The　framework　of　the　improved　network　consists　of　two　parts:　(1)　an　example　segmentation　network,　which　is　used　to　obtain　the　general　shape　and　relative　position　information　of　each　tooth;　(2)　a　fine-grained　segmentation　network,　which　is　used　to　learn　the　fine　details　of　individual　tooth.　A　penalty　mechanism　for　mis-assigned　labels　was　further　used　to　improve　the　accuracy　of　tooth　segmentation.　The　results　showed　that　the　end-to-end　deep　learning　framework　used　in　this　study　can　achieve　accurate　segmentation　in　the　3D　tooth.　The　mean　intersection　over　union　(MIoU)　score　of　the　proposed　improved　GCNN　network　achieves　0.91,　which　was　much　better　than　that　of　PointNet　++　(MIoU:　0.78)　and　GACNet　(MIoU:　0.88).　©　2020,　Beijing　China　Science　Journal　Publishing　Co.　Ltd.　All　right　reserved.