Recently,　the　Reverse　Distillation　(RD)　model　has　shown　importance　in　medical　image　analysis,　especially　in　anomaly　detection.　However,　when　detecting　finer　lesions　in　lung　CT　scans,　there　is　a　core　challenge　of　maintaining　consistent　representation　in　the　teacher-student　network　to　reduce　the　false　positive　rate.　To　address　this　issue,　we　propose　an　anomaly　detection　method　with　learning　Consistent　Representations　for　Reverse　Distillation　(CR4RD)　of　intra-image　and　intra-batch.　On　the　one　hand,　we　introduce　a　hard　global　cosine　similarity　loss　in　the　reverse　skip　connection　distillation　framework,　which　enhances　the　intra-image　consistency　of　representations,　enabling　the　model　to　more　comprehensively　evaluate　the　overall　normality　of　CT　scan　images.　On　the　other　hand,　we　propose　a　unified　regularization　loss　to　achieve　intra-batch　consistency,　guiding　a　uniform　distance　distribution　within　the　same　batch　and　consistency　of　distance　distribution　among　samples　within　the　batch.　Finally,　the　teacher　and　trained　student　network　will　maintain　consistent　representations　of　normal　regions　and　show　significant　differences　in　anomaly　regions　for　the　anomaly　detection　task.　Experimental　results　on　real　lung　CT　scan　datasets　demonstrate　that　the　proposed　method　is　not　only　efficient　but　also　superior　in　anomaly　detection　compared　with　several　state-of-the-art　methods.