This　paper　proposes　a　CNN　(Convolutional　neural　network)　based　blood　vessel　segmentation　algorithm.　Each　pixel　with　its　neighbors　of　the　fundus　image　is　checked　by　the　CNN.　The　preliminary　segmentation　results　of　fundus　images　were　refined　by　a　two　stages　binarization　and　a　morphological　operation　successively.　The　algorithm　was　tested　on　DRIVE　dataset.　While　the　specificity　is　0.9603,　sensitivity　is　0.7731,　which　is　very　close　to　that　of　manual　annotation.　The　sensitivity　is　2%　better　than　the　ones　found　in　current　studies.　The　CNN　based　algorithm　improves　the　segmentation　of　blood　vessels　performance　significantly.