The 3D Instance Segmentation Network for Synapse Reconstruction from Serial Electron Microscopy Images

Abstract Background The synapse is the key part where neurons communicate with each other. Synaptic plasticity plays a vital role in study and memory. Due to the rapid development of electron microscopy (EM) technology, imaging synapses at nanometer scale possible become possible. However, the automation and effectiveness of the synapse detection algorithm have not yet been satisfactory. The most commonly used method is a two-step solution, where first binary segmentation masks are obtained and then reconstruction results are generated by finding connected components. Results In this paper, a novel 3D instance segmentation network which can predict the synapses end to end was proposed. Then, it was also proved that the network can exploit features consistent with the biological structures of synapses by visualizing the network layer. Furthermore, our method was evaluated on two public datasets, and experimental results demonstrated the effectiveness of our proposed method. Conclusion The proposed method provided a fast and accurate solution to detecting synapses from serial section EM images. Besides, a block-wise inference strategy which adapts well to large scale EM images was introduced, and it can also help neuroscientists achieve labor-free analysis and quantification of synapses.