SynReEM: Synapse Reconstruction via Instance Structure Encoding in Anisotropic Electron Microscopic Volumes

Abstract

Volume electron microscopy (vEM) has revolutionized the nanoscale reconstruction of synapses in neural circuits. However, large-scale vEM techniques relying on serial sectioning suffer from severe anisotropy, where axial resolution is far worse than lateral resolution. This anisotropic imaging induces discontinuities in biological architectures across 3D space, compromising reconstruction accuracy and instance segmentation of synapses. Although synapse reconstruction can be realized via aggregation of segmented voxels or detected superpixels, conventional semantic and instance-level models fail to learn voxel instance attributes robustly from strong anisotropic datasets. Here, we present SynReEM, a dedicated framework for synapse reconstruction. Specifically, we first conduct structural encoding on synapse annotations to optimize structural components, making instance segmentation feasible within a semantic context. Then, we incorporate biological priors to impose continuity and inclusion constraints on model outputs, leveraging online pseudo-labels to enhance model convergence. Furthermore, we design a dual-headed branch for simultaneous semantic and instance decoding from shared feature maps, fuse the multi-task outputs, and adopt the watershed algorithm to achieve accurate instance reconstruction. Comprehensive evaluations on three vEM datasets containing synapses (Synapse178, AC3/AC4, and SynWTAD) consistently confirm the superior performance of our proposed SynReEM method.