A　non-rigid　image　registration　algorithm　is　proposed　based　on　an　improved　variational　optical　flow　model　which　is　incorporated　by　a　novel　anisotropic　diffusion.　We　propose　to　employ　the　Hessian　matrix　based　operator　to　avoid　motion　boundary　blurring　existing　in　traditional　optical　flow　approaches,　and　hence　improve　its　accuracy.　Additionally,　the　issue　of　registration　of　medical　images　with　relatively　larger　deformation　is　taken　into　account　by　delaying　of　linearization　to　the　minimizing　stage.　The　correctness　of　the　proposed　methodology　for　non-rigid　medical　image　registration　is　verified　by　experimental　results.　It　is　demonstrated　from　the　comparison　in　the　experiment　that　the　accuracy　of　medical　image　registration　estimated　by　Hessian　matrix　based　diffusion　method　outperforms　that　obtained　by　traditional　optical　flow　model.