Abstract
Accurate 3D neuron segmentation from large-scale electron microscopy (EM) volumes is a cornerstone of connectomics, yet it is hampered by a trade-off between accuracy and computational efficiency. While 3D Convolutional Neural Networks (CNNs) offer high accuracy by predicting 3D affinity maps, their high computational cost and limited input size impede their application to large-scale data. To overcome this, we introduce EM-SliceCrafter, a novel framework that employs a lightweight 2D CNN for both efficient and precise 3D neuron segmentation. Our method generates a 3D affinity map by computing distances between embedding maps of adjacent 2D slices. To enrich the 2D network's contextual understanding, we propose a dual-distillation strategy. First, Cross-Dimension Affinity Distillation (CAD) transfers inter-slice structural knowledge from a 3D teacher network to enhance 3D continuity. Complementarily, Internal-Dimension Affinity Distillation (IAD) leverages a powerful teacher built on the pre-trained Segment Anything Model 2 (SAM2) encoder, using a learnable adapter, to distill fine-grained intra-slice boundary details. Furthermore, a Feature Grafting Interaction (FGI) module deepens knowledge transfer by integrating embeddings across the student and both teacher networks. Extensive experiments on diverse EM datasets, featuring various imaging modalities and resolutions, show that EM-SliceCrafter improves segmentation accuracy over state-of-the-art methods while achieving a 33× speedup in GPU forward inference. The code is available at https://github.com/liuxy1103/EM-SliceCrafter.