Neural Atlas Graphs (NAGs)

• Neural Atlas Graphs (NAGs) are structured graph representations that encode spatial, anatomical, or functional data to unify multi-modal information.
• They facilitate analysis by integrating connectivity, spatial relationships, and statistical associations across fields like neuroscience, vision, and LiDAR mapping.
• NAGs empower scalable, end-to-end learning and editing for diverse applications, ranging from brain connectomics to dynamic scene decomposition.

Neural Atlas Graphs (NAGs) are a class of structured representations that encode the spatial, anatomical, or functional organization of neural, biological, or perceptual domains. NAGs generalize the concept of an atlas—traditionally a labeled template in neuroanatomy or computer vision—by using a graph structure in which nodes represent discrete units (regions, neurons, objects, or volumes) and edges encode connectivity, spatial relationships, or functional dependencies. Across diverse subfields, NAGs unify multi-modal or population data, enable analysis of structure-function relationships, support large-scale mapping and editing, and are increasingly realized with neural network-based methods optimized end-to-end.

1. Conceptual Foundation and Definitions

A Neural Atlas Graph (NAG) is broadly defined as a population- or measurement-representative graph, where each node corresponds to a spatial or semantic unit (e.g., anatomical ROI, neuron, object, or LiDAR submap), and edges encode statistical, anatomical, or functional relationships. The graph structure supports aggregation, comparison, and editing across spatial, functional, or population domains. Variants include:

• Population-representative connectome/CBT: Nodes are ROIs; edges summarize most typical connectivity across subjects or modalities, yielding a connectional brain template (CBT) or network atlas that acts as a "connectional fingerprint" (Chaari et al., 2022).
• Functional propagation atlas: Nodes are individual neurons; directed, weighted, signed edges encode measured causal influences (including timing and sign) between neurons, integrating synaptic and extrasynaptic effects (Randi et al., 2022).
• Scene decomposition in vision: Nodes are neural atlases representing objects or background planes in dynamic scenes; edges (typically implicit) encode ordering or compositing relationships (Schneider et al., 19 Sep 2025).
• LiDAR mapping: Each node is a neural feature field over a sub-volume; edges correspond to spatial transformations derived from odometry or loop closures, forming a globally elastic yet locally rigid map (Yu et al., 2023).
• 3D brain cytoarchitecture: Nodes are sampled mesh points with attributed features; edges are mesh adjacencies, supporting label propagation and topological regularization via GNNs (Schiffer et al., 2021).

Across these formulations, NAGs permit integration of heterogeneous or multi-view data, leverage neural architectures for end-to-end learning, and support both descriptive analysis and downstream applications.

2. Mathematical and Algorithmic Formulations

The mathematical formalism underlying NAGs varies by application. Representative formulations include:

• Multi-graph integration for CBTs (Chaari et al., 2022):
  • For a cohort of $N$ subjects, each with $n_v$ graph views ($X^{(v)}_s \in \mathbb{R}^{n_r \times n_r}$), graph tensors $\mathcal{T}_s = \{X^{(v)}_s\}_{v=1}^{n_v}$ are aggregated to produce a single template $\hat{T} \in \mathbb{R}^{n_r \times n_r}$.
  • The integration objective minimizes population-wise distances plus a regularizer:

  $$\min_{\hat{T}} \sum_{s=1}^N \sum_{v=1}^{n_v} \mathcal{D}(\hat{T}, X^{(v)}_s) + \mathcal{R}(\hat{T})$$ - Choices of $\mathcal{D}$ (e.g., Frobenius norm) and $\mathcal{R}$ (e.g., sparsity) yield different fusion strategies.

• Deep Graph Normalizer (DGN) (Chaari et al., 2022):

  • Uses multi-layer, edge-conditioned GNNs with learned filters over multi-view adjacency matrices. Outputs per-subject templates via node embedding distances and enforces population centeredness via randomized subject normalization loss.
• Functional signal propagation in C. elegans (Randi et al., 2022):
  • The NAG is a directed, weighted, signed graph $G=(V,E,A,\{\tau_{ij}\})$ where
  • $A_{ij} = \mathrm{sgn}(w_{ij})|w_{ij}|$, with $w_{ij}$ as the mean optogenetically-evoked response.
  • Each edge has an associated temporal kernel $k_{ij}(t)$, fitted as a sum of convolved exponentials, capturing the time course of transmission.
• Neural atlases for dynamic scene decomposition (Schneider et al., 19 Sep 2025):
  • Each node $N_i = (C_i, \alpha_i, f_i, g_i, s_i)$ is a learned atlas (texture and opacity) with warping and pose functions.
  • Nodes are rendered via ray–plane intersections and alpha-compositing; gradients are propagated through all transformations for end-to-end optimization.
• Neural feature volume graphs for LiDAR (Yu et al., 2023):
  • Subvolume SDFs $\sigma(p; \Theta_i)$ serve as nodes, connected by relative pose graph edges.
  • Joint optimization is formulated as a MAP inference task integrating range, semantic, and structural priors.

3. Applications Across Domains

3.1. Network Neuroscience and Population Atlases

In human and animal brain mapping, NAGs realize the construction of CBTs, which encapsulate the topological architecture and discriminative features of neuroimaging-derived connectomes. Advanced methods such as DGN leverage deep geometric learning to derive population-centered templates that simultaneously achieve minimal population-averaged distance, maximize biomarker discriminability, and preserve multiscale topological organization (e.g., hubness, modularity, global efficiency) (Chaari et al., 2022).

3.2. Functional Connectomics in C. elegans

Functional NAGs empirically established via optogenetic perturbations vastly outperform anatomical connectome models in predicting spontaneous neural dynamics, due to their ability to encode directed, signed, and temporally resolved functional relationships, including extrasynaptic and indirect pathways—even where anatomical wiring is absent or ambiguous (Randi et al., 2022).

3.3. Editable Scene Representation in Computer Vision

In 3D scene understanding, NAGs enable hybrid representations unifying explicit 2D atlas-based editing with 3D scene-graph-based ordering and manipulation. Nodes are neural atlases for each object or semantic region, composited according to scene geometry. This supports view-consistent, high-fidelity editing, and outperforms prior models on both automotive and general video datasets in PSNR and perceptual quality (Schneider et al., 19 Sep 2025).

3.4. Large-Scale LiDAR Mapping

NAGs in LiDAR mapping (as in NF-Atlas) combine the scalability of graph-based SLAM (via pose graphs) with high-resolution neural fields. Each submap is optimized locally, while global consistency is enforced only via adjustment of volume origins, enabling efficient loop closure and incremental operation without catastrophic forgetting or retraining (Yu et al., 2023).

3.5. Cytoarchitectonic Brain Mapping

For whole-brain cytoarchitectonic parcellation, NAGs are constructed as attributed 3D surface meshes, where GNNs propagate features extracted from histological sections across the mesh. Integration of anatomical priors and spatial coordinates further improves parcellation accuracy and generalization across brains (Schiffer et al., 2021).

4. Evaluation Metrics and Empirical Results

Evaluation of NAGs is domain-dependent but shares common features:

• Population Atlas Evaluation (Chaari et al., 2022):
  • Centeredness: Mean Frobenius distance between template and test graphs.
  • Biomarker Reproducibility: Overlap of discriminative ROI sets.
  • Node-Level Similarity: KL divergence between graph-theoretic measures (strength, efficiency, participation).
  • Global-Level: Modularity ($Q$), global efficiency ($E_{glob}$).
  • Distance-Based: Normalized Hamming and Jaccard distances.

The DGN model achieves the lowest population-centeredness, highest biomarker reproducibility (14–46% boost), and near-identical node/global topological statistics to ground-truth populations (Chaari et al., 2022).

• C. elegans Atlas Predictiveness (Randi et al., 2022):
  • NAG-based prediction of spontaneous correlation matrices achieves $\rho_{func} \sim 0.55-0.70$, outperforming anatomical connectome models ($\rho_{ana} \sim 0.31$).
• Scene Decomposition (Schneider et al., 19 Sep 2025):
  • NAGs outperform prior approaches by 5–11.5 dB (Waymo, dynamic objects) and 7 dB (DAVIS video) in PSNR, with qualitative edits remaining photorealistic and temporally consistent.
• LiDAR Mapping (Yu et al., 2023):
  • NF-Atlas achieves high geometric accuracy (F1, Chamfer-L1), low memory (26.9–64.3 Mb), and rapid training relative to dense or hashed alternatives.
• Cytoarchitectonic Mapping (Schiffer et al., 2021):
  • GNN-based NAGs yield ~16 percentage points improvement over 2D CNN baselines (macro-F1: 66% vs 50% on test), further increasing to ~80% with integrated priors.

5. Comparison and Integration of NAG Methodologies

A range of strategies have been developed for NAG construction and learning:

Application Area	Node Type	Edge/Graph Structure	Optimization/Training
Population atlas / CBT (Chaari et al., 2022)	ROIs	Population graphs	End-to-end GNN, subject normalization
Functional connectomics (Randi et al., 2022)	Neurons	Functional, causal edges	Experimental/correlation fitting
Scene decomposition (Schneider et al., 19 Sep 2025)	Object neural atlases	Depth ordering (implicit)	Gradient-based end-to-end backprop
LiDAR mapping (Yu et al., 2023)	Neural SDF volumes	Pose graph	MAP bundle adjustment, local MLPs
Cytoarchitecture (Schiffer et al., 2021)	Mesh vertices/patches	Mesh adjacency	Contrastive CNN, GNN, priors integration

Alternative methods for integration (e.g., SNF, netNorm, SCA, MVCF-Net) focus on kernel fusion, clustering, or multi-view diffusion but lack the GNN-based, end-to-end differentiable alignments and median-based aggregation that characterize modern NAG approaches (Chaari et al., 2022). NAGs have demonstrated superior representational power, robustness, and ability to capture multi-scale organizational principles when optimized with contemporary geometric deep learning frameworks and robust statistical objectives.

6. Implications, Limitations, and Future Directions

NAGs serve as a unifying abstraction across neuroscience, computer vision, robotics, and biological mapping. Key implications and emerging guidelines are:

• Functional integration surpasses static anatomy: In neural systems, incorporating causal, temporally resolved, and extrasynaptic factors is critical for accurate prediction and interpretation of network dynamics (Randi et al., 2022).
• Graph-based population atlases enable reproducible biomarker discovery and topological analysis: End-to-end trainability and robust subject-to-population alignment are necessary for consistent, interpretable templates (Chaari et al., 2022).
• Hybrid atlas-graph models offer high-resolution, editable scene representations beyond traditional layers, with explicit manageability of multi-object arrangements and transformations (Schneider et al., 19 Sep 2025).
• Separation of local encoding and global elasticity is essential for scalable, incremental mapping in physical environments: Rigidity at the neural field level, with global corrections via pose graphs, supports robustness to loop closures and avoids catastrophic forgetting (Yu et al., 2023).

Limitations are inherent to data quality, scalability, and domain-specific requirements (e.g., cell identification, volumetric imaging, atlas registration). Prospective extensions include further integration of dynamic factors, improved cross-subject alignment, and development of domain-general NAGs for cross-modal and cross-species mapping.

NAGs are expected to play a central role in integrative neural and spatial modeling, offering a flexible interface for statistics, learning, and interactive manipulation across scientific and engineering disciplines.