NuGraph: Graph-Based Reasoning over 3D Primitives for Nucleus Segmentation Correction

Correcting segmentation errors in large-scale 3D nuclei reconstructions requires reasoning about which fragments belong to the same nucleus across densely packed regions. Existing correction methods rely on local pairwise fragment matching, which cannot resolve the global topology of nuclear clusters and fails to recover missing morphology. We propose N u G raph , a graph-based reasoning framework that operates over atomic 3D primitives obtained by decomposing erroneous masks. N u G raph encodes primitive geometry via a 3D point-cloud backbone and performs global relational reasoning through graph attention, capturing inter-primitive dependencies across entire clusters rather than isolated pairs. A primitive–proposal contrastive loss aligns local primitive features with nucleuslevel semantics, improving grouping accuracy in dense regions. The resulting proposals are then refined by a shaperefinement network that predicts signed distance fields to restore smooth morphology. To train without manual error annotations, we develop a self-supervised data engine that synthesizes realistic segmentation errors from clean nuclei labels. To benchmark correction at brain scale, we curate NucEMFix, the first brain-wide EM benchmark of nuclei error cases across FAFB and MICrONS (8,000+ annotated error nuclei). N u G raph attains 87.99% F1 on NucEMFix-F (FAFB) and 86.20% on NucEMFix-M (MICrONS), outperforming both re-segmentation baselines (e.g., +8.6% over nnU-Net) and pairwise correction methods, while reducing curation effort by over 100× relative to manual proofreading. Code and data are available at https://mingzhiwang618.github.io/NucEMFix .