SynReEM: Synapse Reconstruction via Instance Structure Encoding in Anisotropic Electron Microscopic Volumes
We present SynReEM, a framework for reconstructing individual synapses in anisotropic volume EM using structural encoding, biological constraints, and joint semantic and instance prediction.
Jinyue Guo*, Yanchao Zhang*, Hao Zhai*, Yi Jiang, Qi Zhang, Yunfeng Hua, Jing Liu, Hua Han
Institute of Automation, Chinese Academy of Sciences
Shanghai Institute of Precision Medicine, Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine
Department of Ultrastructural Pathology, Capital Medical University
Abstract
Serial-section volume electron microscopy (vEM) enables 3D reconstruction of synapses, but its coarse axial sampling makes structures difficult to connect reliably between sections. We present SynReEM, a framework that encodes synaptic structure so that a semantic segmentation model can distinguish individual synapses. During training, biological priors impose continuity and inclusion constraints, while online pseudo-labels guide learning. A shared feature representation feeds separate semantic and instance prediction branches. Their outputs are combined and processed with watershed segmentation to recover individual synapses. Evaluations on Synapse178, AC3/AC4, and SynWTAD show improved reconstruction performance across three anisotropic vEM datasets. Code, dataset links, and reconstruction videos are available at https://github.com/fenglingbai/SynReEM.
Citation
Guo J, Zhang Y, Zhai H, Jiang Y, Zhang Q, Hua Y, Liu J, Han H (2026). SynReEM: Synapse Reconstruction via Instance Structure Encoding in Anisotropic Electron Microscopic Volumes. IEEE Transactions on Medical Imaging, 45(9): 4662-4677. https://doi.org/10.1109/TMI.2026.3706567.