Cherenkov-excited　luminescence　scanned　tomography　(CELST)　can　recover　a　high-resolution　3D　distribution　of　luminescent　sources　within　tissue.　However,　reconstructing　the　distribution　of　the　quantum　field　from　boundary　measurements　is　a　typical　ill-posed　problem.　In　this　work,　we　propose　a　novel　two-step　reconstruction　network　(TSR-Net)　based　on　a　fusion　mechanism,　that　integrates　two　encoder-decoder　networks　(ED-Net)　using　a　concatenation　block.　Firstly,　an　ED-Net　is　trained　to　learn　the　CT　structural　features　of　tissues　with　the　measured　data.　Then,　the　trained　ED-Net　is　fixed　and　cascaded　by　another　ED-Net　for　a　second-step　training　to　predict　the　3D　distributions.　Numerical　simulations　reveal　that　the　proposed　approach　can　not　only　accurately　reconstruct　the　intensity　values　of　the　luminescent　sources,　but　also　achieve　a　reconstruction　resolution　of　1mm　with　low　target-background　contrast.　Furthermore,　the　well-trained　network　is　still　effective　in　the　reconstruction　of　tissues　with　different　shapes,　which　indicates　an　excellent　generalization　ability　of　the　algorithm.　©　2023,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.