AUTOMATED NEURITE SEGMENTATION IN THE MARMOSET PREFRONTAL CORTEX USING LARGE-VOLUME SERIAL TEM

Large-volume electron microscopy (vEM) enables dense reconstruction of synaptic connectivity at the scale of cortical microcircuits, yet extracting neurite trajectories from terabyte- to petabyte-scale datasets remains a major bottleneck. We acquired a large serial-section vEM dataset from marmoset prefrontal cortex using a high-throughput imaging system (Blade, Voxa, USA) mounted on a transmission electron microscope (TEM). The dataset covers a 1.1 × 1.6 mm² region across ~1,000 ultrathin sections (50 nm thickness), totaling 514 TB and acquired in ~1 month.
To prepare the volume for automated analysis, we stitched tiled TEM images and performed inter-section alignment using FEABAS [1], generating a consistently registered stack suitable for downstream dense segmentation. For neurite extraction, we applied PyTorch-Connectomics [2] with an affinity-learning formulation, where the training target indicates whether neighboring pixels belong to the same neurite. The trained convolutional neural network was applied to the aligned stack to produce dense neurite segmentation maps, which were subsequently refined through interactive proofreading to correct residual merge/split errors and yield traceable neurite paths.
We demonstrate an end-to-end pipeline from high-throughput acquisition to stitching/alignment, deep-learning-based segmentation, and proofreading-assisted tracing, and we report representative neurite tracing results in marmoset prefrontal cortex. The modular, tile-based workflow is intended to be practical for small laboratories, minimizing compute requirements while scaling to hundreds of terabytes. Our achievement using the volume EM technique greatly contribute to proceed the connectome analysis in cortical microcircuit investigation.
[1] https://github.com/YuelongWu/feabas
[2] Lin et al. PyTorch Connectomics. arXiv:2112.05754.