Anomaly　detection　in　medical　images　is　a　fundamental　problem　in　computer-aided　medical　image　analysis.　How　to　transform　the　medical　prior　knowledge　carried　by　abnormal　examples　into　supervision　or　constraint　for　network　training　is　a　crucial　issue.　We　proposed　a　rehabilitation　bridge　model　to　learn　the　mapping　from　the　abnormal　to　normal　manifestations　in　feature　space.　The　rehabilitation　bridge　will　increase　the　discriminative　difference　efficiently.　We　applied　the　proposed　rehabilitation　bridge　model　to　detect　the　absence　of　bone　wall　in　the　jugular　bulb　and　sigmoid　sinus　in　the　temporal　bone　CT　images.　Experimental　results　show　that　it　can　improve　the　performance　of　anomaly　detection　with　the　supervision　of　a　small　number　of　abnormal　examples.