Research map / Probabilistic and scientific modeling / Dynamical systems
Joint Learning Neuronal Skeleton and Brain Circuit Topology with Permutation Invariant Encoders for Neuron Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NeuNet, a permutation-invariant neuron classifier fusing a point-cloud skeleton encoder (FPS/ball-query grouping with 1x1 Conv1D) and a GCNII-style connectome GNN, and releases reprocessed HemiBrain and H01 neuron classification datasets.
Problem
Automated neuron type classification from whole-brain volume EM data, using both neuronal skeleton morphology and synaptic connectome topology, which prior work had not jointly modeled.
Approach
Combines a bottom-up skeleton point encoder (FPS sampling, ball-query grouping, kernel-size-1 Conv1D with max pooling, shown permutation invariant) with a deep GCNII-based connectome encoder, concatenating both embeddings for an MLP classifier.
Key results
- NeuNet reaches 0.9169 overall accuracy on HemiBrain versus 0.8627 for PointNet++ (second best).
- NeuNet reaches 0.9363 overall accuracy on H01, far above the best baseline (AGNN 0.6334).
- Removing the connectome encoder drops HemiBrain accuracy to 0.5865 in the NeuNet(+Connnetome)/NeuNet-S variant table, while removing skeleton info (NeuNet-C) drops H01 accuracy to 0.5123, showing the two modalities are complementary.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace feature concatenation with max pooling (ablation) | NeuNet · HemiBrain | accuracy: 0.9169 -> 0.9149 | No clear effect |
| replace feature concatenation with average pooling (ablation) | NeuNet · HemiBrain | accuracy: 0.9169 -> 0.9135 | No clear effect |
| replace feature concatenation with max pooling (ablation) | NeuNet · H01 | accuracy: 0.9363 -> 0.9412 | No clear effect |
| replace feature concatenation with average pooling (ablation) | NeuNet · H01 | accuracy: 0.9363 -> 0.9419 | No clear effect |
| add connectome information to the skeleton-only model (ablation) | NeuNet · HemiBrain | accuracy: 0.8612 -> 0.9169 | Helped |
| add skeleton information to the connectome-only model (ablation) | NeuNet · HemiBrain | accuracy: 0.5865 -> 0.9169 | Helped |
| add connectome information to the skeleton-only model (ablation) | NeuNet · H01 | accuracy: 0.6356 -> 0.9363 | Helped |
| add skeleton information to the connectome-only model (ablation) | NeuNet · H01 | accuracy: 0.5123 -> 0.9363 | Helped |
Surprising findings
- Ablation tables in the paper are internally inconsistent: the NeuNet-S and NeuNet-C rows are labeled differently across tables, and the text's reported drops do not clearly match the tabulated values.
- Max-pooling and average-pooling fusion variants perform about the same as concatenation, and on H01 they slightly exceed it (0.9412 and 0.9419 vs 0.9363).
Limitations
- Baseline results are reported with no variance or repeated runs, and the appendix hyperparameter tables are inconsistent with the main text.
- Connectome encoder uses 64 layers and the skeleton encoder 3 layers in the text, the reverse of the appendix description, so the architecture is somewhat ambiguous.
Experimental setup
Datasets: HemiBrain H01
Benchmarks: HemiBrain H01
Metrics: accuracy within-class accuracy
Baselines: MLP PointNet++ GCNII DGCNN AGNN Point-Transformer
Builds on: GCNII PointNet++ DGCNN Point-Transformer
Closest papers by meaning
- GraPHFormer: A Multimodal Graph Persistent Homology Transformer for the Analysis of Neuroscience Morphologies (CVPR 2026 · similarity 0.86)
- $\textit{NeuroPath}$: A Neural Pathway Transformer for Joining the Dots of Human Connectomes (NeurIPS 2024 · similarity 0.85)
- Learning Multimodal Volumetric Features for Large-Scale Neuron Tracing (AAAI 2024 · similarity 0.85)
- Efficient Connectivity-Preserving Instance Segmentation with Supervoxel-Based Loss Function (AAAI 2025 · similarity 0.84)
- Neuron Platonic Intrinsic Representation From Dynamics Using Contrastive Learning (ICLR 2025 · similarity 0.84)
- Biologically Plausible Brain Graph Transformer (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.