.png){kind=icon}
.png){kind=icon}
Return
A Transformer-Based Framework for Large-Scale EM Segmentation Stitching
.png){kind=icon}
PRE
AI
DOI：10.1142/S0219691326500116.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
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.
Keywords:
Large-scale reconstruction
neuron stitching
multimodal feature fusion
lazy relabeling
Journal
I
IF:
0.8
Papers:
39
Citations:
717