A Transformer-Based Framework for Large-Scale EM Segmentation Stitching

Accurate neuron stitching across large-scale electron microscopy volumes is crucial for reconstructing complete neural circuits. We propose TransStitch, a distributed Transformer-based framework that addresses these challenges by integrating multimodal feature fusion with topology-aware self- and cross-attention mechanisms to model global structural dependencies across adjacent electron microscopy blocks. To refine uncertain predictions, a dynamic 1-nearest-neighbor strategy progressively converts the probabilistic connectivity matrix into discrete associations without relying on a fixed threshold. Additionally, a mapping-based lazy relabeling strategy reduces merging complexity from voxel to fragment level, significantly improving computational efficiency and scalability. Extensive experiments on public electron microscopy datasets (SNEMI3D, CREMI-C, FIB25) with ground truth demonstrate superior stitching accuracy of proposed method compared to baseline, while qualitative evaluation on a large-scale, self-collected zebrafish whole-brain dataset confirms coherent 3D reconstruction across tens of thousands of sections. These results highlight TransStitch as an accurate and scalable solution for large-scale connectomics reconstruction.