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

Area of Science:

• Neuroscience
• Biotechnology
• Computer Science

Background:

• Volume electron microscopy (vEM) enables nanoscale neural circuit reconstruction.
• vEM data often exhibits severe anisotropy, impacting axial resolution and reconstruction accuracy.
• Existing models struggle with instance segmentation of synapses from anisotropic vEM datasets.

Purpose of the Study:

• To develop a dedicated framework, SynReEM, for accurate synapse reconstruction from anisotropic vEM data.
• To overcome limitations of conventional models in handling voxel instance attributes from anisotropic datasets.
• To improve the accuracy and reliability of 3D synapse reconstruction in neural circuits.

Main Methods:

• SynReEM employs structural encoding of synapse annotations for optimized components.
• Biological priors are incorporated for continuity and inclusion constraints, with online pseudo-labels for convergence.
• A dual-headed branch decodes semantic and instance information simultaneously, fused with watershed algorithm for reconstruction.

Main Results:

• SynReEM demonstrates superior performance in synapse reconstruction across three vEM datasets (Synapse178, AC3/AC4, SynWTAD).
• The framework effectively addresses challenges posed by anisotropic imaging in vEM data.
• Accurate instance reconstruction of synapses is achieved, improving biological architecture mapping.

Conclusions:

• SynReEM provides a robust solution for synapse reconstruction from anisotropic vEM data.
• The method enhances the feasibility of large-scale vEM-based neural circuit analysis.
• This framework advances the accuracy of 3D synapse segmentation and reconstruction in neuroscience research.