Purpose:　Presence　of　metal　artifacts　is　a　major　reason　of　degradation　of　computed　tomography　image　quality　and　there　is　still　no　standard　solution　to　this　issue.　A　class　of　recently　investigated　metal　artifact　reduction　(MAR)　methods　based　on　forward　projection　of　a　prior　image　that　is　artifact-free　to　replace　the　metal　affected　projection　data　have　shown　promising　results.　However,　usually　it　is　hard　to　get　a　good　prior　image　which　is　close　to　the　true　image　without　artifacts.　This　work　aims　at　creating　a　good　prior　image　so　that　the　forward　projection　can　replace　the　metal　affected　projection　data　well.　Methods:　The　proposed　method　consists　of　four　steps　based　on　the　forward　projection　MAR　framework.　First,　metal　implants　in　the　reconstructed　image　are　segmented　and　the　corresponding　metal　traces　in　the　projection　domain　are　identified.　Then　the　prior　image　is　obtained　by　two　steps.　A　processed　precorrected　image　is　generated　as　an　initial　prior　image　first　and　then　in　the　next　step　it　is　used　as　the　initial　image　of　the　iterative　reconstruction　from　the　unaffected　projection　data　to　generate　a　better　prior　image.　In　order　to　deal　with　severe　artifacts,　the　iteration　incorporates　the　total　variation　minimization　constraint　as　well　as　a　novel　constraint　which　forces　the　soft　tissue　region　near　metal　to　be　as　flat　as　possible.　Finally,　the　projection　is　completed　using　forward　projection　of　the　prior　image　and　the　corrected　image　is　reconstructed　by　FBP.　A　linear　interpolation　MAR　method　and　two　recently　reported　forward　projection　based　methods　are　performed　simultaneously　for　comparison.　Results:　The　proposed　method　shows　outstanding　performance　on　both　phantoms＇　and　patients＇　datasets.　This　approach　can　reduce　artifacts　dramatically　and　restore　tissue　structures　near　metal　to　a　large　extent.　Unlike　competing　MAR　methods,　it　can　effectively　prevent　introduction　of　new　artifacts　and　false　structures.　Moreover,　the　proposed　method　has　the　lowest　RMSE　in　regions　of　both　soft　tissue　and　bone　tissue　among　the　corrected　images　and　is　ranked　as　the　best　method　for　evaluation,　by　radiologists.　Conclusions:　Both　subjective　and　quantitative　evaluations　of　the　results　demonstrate　the　superior　performance　of　the　proposed　algorithm,　compared　to　that　of　the　competing　methods.　This　method　offers　a　remarkable　improvement　of　the　image　quality.　(C)　2013　American　Association　of　Physicists　in　Medicine.　[http://dx.doi.org/10.1118/1.4794474]