Source-linked AI summary
SkNeXt enables topology-guided neuronal reconstruction from petabyte-scale microscopy data
Jiayi Ding, Hu Zhao
TL;DR
Large microscopy volumes make complete neuronal reconstruction expensive in computation, data movement, and proofreading. SkNeXt uses proofread skeletons as topology-preserving priors and spatial indices for detailed reconstruction and selective data access, achieving large-scale mouse-brain reconstruction on a single GPU while reducing manual effort.
Problem
Terabyte- to petabyte-scale microscopy datasets make complete neuronal reconstruction expensive in computation, data movement, and manual proofreading.
Method
SkNeXt reconstructs and proofreads compact SWC skeletons, then uses them as topological priors and spatial indices for high-resolution instance and semantic reconstruction.
Results
Approximately 0.5 person-hours per neuron were required for skeleton proofreading versus approximately 4 person-hours for direct manual reconstruction, while a single GPU processed a 0.3-PB hippocampal dataset in approximately three days.
Takeaways & Limitations
Skeleton-guided reconstruction preserves neuronal identity and topology while restricting high-resolution processing to regions traversed by neuronal processes.
Takeaways & Limitations
Manual proofreading remains necessary in regions containing highly interwoven neurites, weak signals, or extremely thin neuronal processes.
Abstract
from arXiv · showhide
Recent advances in high-resolution fluorescence and electron microscopy have enabled nanoscale imaging across increasingly large brain volumes, but the resulting terabyte- to petabyte-scale datasets make complete neuronal reconstruction prohibitively expensive in computation, data movement, and manual proofreading. Here, we present SkNeXt, a topology-first framework for scalable neuronal reconstruction from large volumetric microscopy datasets. Instead of densely processing entire image volumes, SkNeXt first converts neuronal morphology into compact SWC skeletons that preserve long-range connectivity. Proofreading is therefore focused on sparse neuronal trees, allowing branch, continuity, and connectivity errors to be corrected before high-resolution reconstruction. The corrected skeletons then serve as persistent structural priors for recovering detailed morphology while preserving neuronal identity and topology. Crucially, SkNeXt also uses neuronal skeletons as spatial indices for selective data access, retrieving high-resolution image regions only along reconstructed trajectories and bypassing most background and signal-free volumes. This substantially reduces I/O and computational overhead, allowing reconstruction cost to scale with neuronal morphology rather than total dataset size. Using SkNeXt, we reconstructed neurons from a petabyte-scale super-resolution fluorescence dataset of the mouse brain on a single GPU within one week, without requiring exhaustive dense inference across the complete imaging volume.
1 Introduction
SkNeXt addresses the cost and reliability challenges of reconstructing long-range neuronal morphology in terabyte- to petabyte-scale brain datasets. It establishes topology with compact skeletons, proofs those structures, and uses them to guide detailed reconstruction and selective image access.
- Motivation: Petabyte-scale brain imaging shifts the bottleneck from acquisition to scalable neuronal reconstruction, especially for long-range neurons whose identity and connectivity span many image blocks.Local errors can accumulate into fragmentation, erroneous mergers, incorrect branch assignments, or lost distal projections.
- Motivation: Skeletons preserve continuity and branching topology, whereas dense segmentation captures fine boundaries but is computationally expensive and locally ambiguous.This complementarity motivates reformulating reconstruction as a topology-first problem.
- Framework: SkNeXt first reconstructs neuronal skeletons from reduced-resolution images, then uses proofread skeletons as structural priors for high-resolution morphology while preserving neuronal identity across separated regions.Proofreading focuses on correcting topological errors in sparse neuronal trees before detailed reconstruction.
- Framework: Corrected skeletons also act as spatial indices, retrieving high-resolution image regions along neuronal trajectories while bypassing most background and signal-free volumes.This makes reconstruction cost scale more closely with neuronal morphology than with total imaging volume.
- Framework: SkNeXt reconstructed neurons from petabyte-scale mouse-brain fluorescence data using a single GPU, without exhaustive dense processing of the complete volume.The framework is intended for datasets where exhaustive dense processing is increasingly impractical.
2.1 Datasets preparation
The study used densely labeled mouse-brain fluorescence data sampled across multiple regions and morphological contexts. The selections varied neuronal density, signal quality, background, and neurite overlap to represent whole-brain reconstruction conditions.
- 2.1 Datasets preparation: The datasets consisted of super-resolution fluorescence microscopy volumes from densely labeled mouse brain.The data were selected for large-scale neuronal reconstruction experiments.
- 2.1 Datasets preparation: Representative regions included somata, dendritic arbors, local axonal branches, and long-range projecting axons across varied imaging conditions.The sampled volumes differed in neuronal density, signal intensity, background level, and neurite overlap.
2.2 Automatic SWC skeleton reconstruction
Automatic SWC reconstruction uses a downsampled multi-task network to predict foreground, skeleton, orientation, and branch-point information. These predictions are merged into a graph whose confidence- and orientation-aware minimum-cost tree is refined and exported in SWC format.
- 2.2 Automatic SWC skeleton reconstruction: The network processes volumes downsampled by factors of 2, 4, and 4 along Z, Y, and X, producing foreground, skeleton, orientation, and branch-point predictions.The reduced-resolution input lowers computational and I/O requirements while retaining long-range trajectories.
- 2.2 Automatic SWC skeleton reconstruction: Foreground and skeleton outputs use binary cross-entropy and Dice losses, while branch-point prediction uses focal loss because branch points are sparse.Orientation prediction uses a sign-invariant cosine loss over valid non-branching skeleton locations.
- 2.2 Automatic SWC skeleton reconstruction: Overlapping block predictions are merged, high-probability skeleton candidates are extracted, and branch-point maxima are retained before graph construction.The resulting graph connects nearby candidates when intervening regions are supported by foreground and skeleton predictions.
- 2.2 Automatic SWC skeleton reconstruction: Edge costs combine physical distance, mean skeleton and foreground probabilities, and local orientation agreement, assigning lower costs to high-confidence consistent connections.This cost guides topology-aware graph reconstruction.
- 2.2 Automatic SWC skeleton reconstruction: Tracing builds a minimum-cost tree from a soma or root, bridges short supported discontinuities, favors confident bifurcations, and removes spurious or redundant branches before SWC export.Permissive initial thresholds preserve neurite recall, followed by confidence- and length-based pruning and resampling.
2.3 Skeleton proofreading
Skeleton proofreading corrects topology in automatically reconstructed SWC trees rather than dense voxel labels. This sparse representation reduces manual interaction by limiting edits to a relatively small number of nodes and edges.
- 2.3 Skeleton proofreading: Proofreading removes false-positive or over-connected branches, reconnects supported discontinuities, and corrects incorrect parent–child relationships in automatically reconstructed SWC skeletons.Nodes and branches are subsequently assigned soma, axon, or dendrite identities based on anatomical continuity and local morphology.
- 2.3 Skeleton proofreading: Proofreading sparse skeleton graphs requires editing nodes and edges instead of inspecting and modifying large three-dimensional label volumes, substantially reducing manual interaction.The workflow therefore concentrates correction on topological errors in a compact representation.
2.4 Semantic and instance segmentation
SkNeXt performs high-resolution semantic and instance reconstruction only around proofread neuronal skeletons, preserving global identities while recovering detailed morphology and subcellular labels.
- High-resolution reconstruction uses foreground, skeleton, and semantic probability maps for soma, dendrite, and presynaptic bouton structures.
- Selective inference processes expanded image regions around skeleton trajectories while retaining central blocks to reduce boundary artifacts and bypass background volumes.
- Proofread SWC trajectories are aligned to high-confidence predicted centerlines, preserving connectivity and neuronal identity during position refinement.
- Marker-controlled watershed grows instances from uniquely identified skeleton seeds through foreground regions, separating neighboring processes while maintaining long-range identities.
- Semantic predictions are mapped onto reconstructed neurons to associate somata, dendritic structures, and presynaptic boutons with individual neuronal identities.
2.5 Implementation details
The implementation uses chunked storage and selective loading for large microscopy volumes, with paired ConvNeXt-V2 encoder–decoder networks for skeleton and semantic-instance reconstruction.
- SkNeXt was evaluated on datasets of approximately 0.2 to 2 PB using a single NVIDIA RTX PRO 6000 Blackwell GPU.
- Raw datasets were converted to Imaris or OME-Zarr chunked formats so inference loaded only image chunks required by reconstruction.
- Both reconstruction stages used ConvNeXt-V2 encoder–decoder networks with 32 × 256 × 256 input patches and feature channels [32, 64, 128, 256, 320].
3 Results
SkNeXt was tested on single-view and 0.3-PB hippocampal fluorescence datasets, combining sparse skeleton proofreading with skeleton-guided reconstruction at large scale.
- The single-view dataset produced skeletons preserving major branches and long, thin axonal processes before dense instance segmentation.
- The 0.3-PB experiment required manual correction and annotation of automatically reconstructed skeletons, while proofreading remained focused on sparse topology.
- Approximately three days was sufficient to complete segmentation inference on the 0.3-PB hippocampal dataset using a single NVIDIA RTX PRO 6000 Blackwell GPU.
- Figure 2 reports segmentation results for a single view.
- Reconstructed instances preserved corrected-skeleton identities, avoiding apparent merging or fragmentation along examined paths and separating neighboring processes.
4 Conclusion
SkNeXt separates topology reconstruction from dense segmentation, using proofread skeletons to preserve neuronal identity and guide large-scale morphology reconstruction. The workflow reduces manual effort and enables petabyte-scale, single-GPU reconstruction, although difficult regions still require proofreading.
- 4 Conclusion: SkNeXt separates topology reconstruction from dense voxel-level segmentation, preserving neuronal identity across long processes while reducing fragmentation and merging errors.Skeletons are reconstructed first and then used as topological priors for instance and semantic segmentation.
- 4 Conclusion: Figures 3 and 4 illustrate the progression from hippocampal neuron skeleton reconstruction to instance segmentation.The final workflow also associates somata, axonal boutons, and dendritic spines with their neuronal identities.
- 4 Conclusion: 0.5 person-hours per neuron versus approximately 4 person-hours reduced proofreading effort for large-scale hippocampal reconstruction.Proofreading focuses on sparse SWC skeletons, which then provide identity-aware seeds for dense segmentation.
- 4 Conclusion: Approximately three days on a single GPU completed instance and semantic segmentation of a 0.3-PB hippocampal dataset.Skeleton-guided selective access restricts full-resolution inference to regions traversed by neuronal processes.
- 4 Conclusion: Manual proofreading remains necessary in regions with highly interwoven neurites, weak signals, or extremely thin neuronal processes.SkNeXt reduces, but does not eliminate, human labor for large-scale reconstruction.