Synapses　are　fundamental　components　of　how　neurons　communicate　with　each　other　and　have　attracted　widespread　attention　from　neuroscientists.　Due　to　the　rapid　development　of　electron　microscopy　(EM)　technology,　imaging　synapses　at　nanometer　scale　has　become　possible.　However,　the　automation　and　efficacy　of　the　synapse　detection　algorithm　have　not　yet　met　expectations.　The　most　popular　approach　involves　a　two-step　process　in　which　binary　segmentation　masks　are　first　obtained　and　then　connected　components　are　used　to　produce　reconstruction　results.　In　this　paper,　an　intelligent　system　to　detect　and　segment　synapses　from　serial　section　EM　images　is　proposed.　Specifically,　a　novel　3D　instance　segmentation　network　that　can　predict　the　synapses　end-to-end　is　presented.　The　network　can　exploit　and　summarize　features　consistent　with　the　biological　structures　of　synapses,　which　is　similar　to　the　process　of　manual　annotation.　A　block-wise　inference　strategy　that　adapts　well　to　large-scale　EM　images　is　then　introduced.　Finally,　two　public　datasets　are　used　to　evaluate　our　method.　Experimental　results　demonstrate　the　superiority　of　the　proposed　approach,　thus　enabling　computer-assisted　analysis　of　synapses　for　neuroscientists.