Cherenkov-excited　luminescence　scanned　tomography　(CELST)　is　an　emerging　tomographic　optical　imaging　modality.　However,　recovering　spatial　distribution　of　luminescent　source　from　boundary　measurements　is　a　typically　ill-posed　problem.　To　improve　the　performance　of　CELST　reconstruction,　an　end-to-end　reconstruction　algorithm　is　developed　by　combining　dilated　convolution　and　attention　mechanism　based　on　Unet　(DA-Unet).　Its　performance　is　validated　with　numerical　simulations.　The　results　reveal　that　DA-Unet　has　superior　reconstruction　performance　with　high　spatial　resolution.　It　achieves　image　quality　with　PSNR　of　more　than　35　dB　and　SSIM　of　larger　than　0.95.　Furthermore,　the　DA-Unet　can　reconstruct　luminescent　source　even　with　less　boundary　measurements.