Brain-HGCN: Dual-Modal Graph Networks

• Brain-HGCN is a family of brain-specific heterogeneous graph neural networks that integrate dual-modal imaging data to capture both functional and structural connectivity.
• It preserves within-modality and cross-modality relations using homo-meta-paths and hetero-meta-paths alongside an attention backbone for effective message passing.
• The approach incorporates modality-aware pooling, dynamic graph augmentation, and rigorous evaluation on ADNI-3, achieving diagnostic accuracies above 93%.

Brain-HGCN is a designation used in arXiv literature for several brain-specific graph neural architectures, but in the most explicit form it refers to a heterogeneous graph neural network that fuses resting-state fMRI functional connectivity and DTI structural connectivity for mild cognitive impairment diagnosis. In that formulation, the model preserves within-modality and cross-modality relations through homo-meta-paths and hetero-meta-paths, uses a Heterogeneous Graph Attention Network backbone, introduces modality-aware heterogeneous pooling, and applies a heterogeneous graph data augmentation strategy to address sample imbalance on ADNI-3 (Yin et al., 2024). The same label, or closely related mappings to it, has also been used for Hodge-Laplacian spectral models, dynamic prototype-based structural graph construction from T1-MRI, hierarchical signed pooling for functional networks, hypergraph autoencoding, and hyperbolic graph learning, so the term is best understood as a family of brain-network GNN formulations rather than a single universally standardized architecture (Huang et al., 2023).

1. Definition and terminological scope

In the dual-modal MCI study, Brain-HGCN denotes the complete pipeline consisting of heterogeneous graph construction with four meta-paths, HAN-based convolutions with hierarchical attention, heterogeneous pooling, dynamic FC augmentation, and MLP classification (Yin et al., 2024). The paper states that the backbone is referred to as Brain-HAN, while the overall method is a brain-specific heterogeneous graph convolutional/neural network because it defines multiple node and edge types per brain modality, performs relation-specific transformations and attentional message passing, incorporates heterogeneous pooling tailored to brain modalities, and includes biologically constrained heterogeneous graph augmentation.

The term is not unique to that paper. In adjacent literature, “Brain-HGCN” or an explicit mapping to it appears in at least five additional senses: a Hodge-Laplacian heterogeneous graph convolutional neural network operating jointly on node- and edge-level functional signals (Huang et al., 2023); a dynamic structural brain network construction method from T1-MRI called Hierarchical Prototype Embedding GCN, described as synonymous with Brain-HGCN (Leng et al., 2023); a hierarchical signed graph representation learning framework for functional networks (Tang et al., 2022); a hypergraph convolutional autoencoder for multi-view hyperconnectomes (Banka et al., 2020); and a hyperbolic graph convolutional network in Lorentz space for signed fMRI graphs (Jia et al., 18 Sep 2025). A recurrent source of confusion is that HGCN may mean heterogeneous, hierarchical, hypergraph, Hodge-Laplacian, or hyperbolic depending on the paper.

This nomenclatural multiplicity suggests that Brain-HGCN is not a single canonical model class. A plausible implication is that any technical discussion of Brain-HGCN must be anchored to the specific paper and operator family under consideration.

2. Dual-modal heterogeneous graph formulation for MCI diagnosis

The dual-modal Brain-HGCN formulation is motivated by the observation that mild cognitive impairment is associated with subtle disruptions in both brain function and structure, and that structure and function are coupled but can also decouple in transmodal cortex (Yin et al., 2024). Resting-state fMRI provides blood oxygen level dependency information from which functional connectivity is derived, while DTI provides white matter structure information and structural connectivity. The model is designed to avoid homogeneous fusion that collapses this heterogeneity.

The ADNI-3 cohort contains 409 subjects with dual-modal imaging: 100 MCI, 48 EMCI, 21 LMCI, and 240 NC. rs-fMRI preprocessing uses DPABI, discards the first 10 volumes, performs motion correction, spatial normalization, and spatial smoothing, and parcellates the brain with the AAL atlas into $N = 90$ ROIs. For each ROI, the mean BOLD time series $x_i \in \mathbb{R}^{T}$ is extracted, and functional connectivity is computed by Pearson correlation,

$$C_{ij} = corr(x_i, x_j) = \frac{\sum_{t=1}^{T}(x_i(t) - \mu_i)(x_j(t) - \mu_j)}{\sigma_i \sigma_j}.$$

Edges with $|C_{ij}| < 0.2$ are set to $0$, the FC matrix is normalized to obtain $A_f \in \mathbb{R}^{90 \times 90}$, and node features for fMRI are the mean time series $X_f \in \mathbb{R}^{90 \times D}$.

DTI preprocessing uses PANDA, computes diffusion metrics and FA maps, parcellates them into the same 90 AAL ROIs, performs deterministic tractography, and extracts radiomic features from FA maps per ROI. Structural connectivity attributes are derived from tractography, with edges satisfying $S_{ij} < 5 \rightarrow 0$, where $S_{ij}$ is the number of reconstructed streamlines between $i$ and $x_i \in \mathbb{R}^{T}$0 or a related tract-strength measure. The normalized structural adjacency is $x_i \in \mathbb{R}^{T}$1, and node features are FA radiomics $x_i \in \mathbb{R}^{T}$2.

The heterogeneous graph has two node types, $x_i \in \mathbb{R}^{T}$3 and $x_i \in \mathbb{R}^{T}$4, each with 90 nodes. Homo-meta-paths preserve within-modality relations:

• $x_i \in \mathbb{R}^{T}$5 via $x_i \in \mathbb{R}^{T}$6
• $x_i \in \mathbb{R}^{T}$7 via $x_i \in \mathbb{R}^{T}$8

Hetero-meta-paths model cross-modal relations:

• $x_i \in \mathbb{R}^{T}$9
• $$C_{ij} = corr(x_i, x_j) = \frac{\sum_{t=1}^{T}(x_i(t) - \mu_i)(x_j(t) - \mu_j)}{\sigma_i \sigma_j}.$$0

At node level, cross-modal structure-function coupling is defined by similarity between connectivity rows,

$$C_{ij} = corr(x_i, x_j) = \frac{\sum_{t=1}^{T}(x_i(t) - \mu_i)(x_j(t) - \mu_j)}{\sigma_i \sigma_j}.$$1

implemented as cosine similarity, and for each fMRI node only the TopK DTI nodes with largest $$C_{ij} = corr(x_i, x_j) = \frac{\sum_{t=1}^{T}(x_i(t) - \mu_i)(x_j(t) - \mu_j)}{\sigma_i \sigma_j}.$$2 are retained, with $$C_{ij} = corr(x_i, x_j) = \frac{\sum_{t=1}^{T}(x_i(t) - \mu_i)(x_j(t) - \mu_j)}{\sigma_i \sigma_j}.$$3. At community level, shared closed induced subgraphs with three nodes forming a triangle in both $$C_{ij} = corr(x_i, x_j) = \frac{\sum_{t=1}^{T}(x_i(t) - \mu_i)(x_j(t) - \mu_j)}{\sigma_i \sigma_j}.$$4 and $$C_{ij} = corr(x_i, x_j) = \frac{\sum_{t=1}^{T}(x_i(t) - \mu_i)(x_j(t) - \mu_j)}{\sigma_i \sigma_j}.$$5 induce binary hetero edges. These are combined as

$$C_{ij} = corr(x_i, x_j) = \frac{\sum_{t=1}^{T}(x_i(t) - \mu_i)(x_j(t) - \mu_j)}{\sigma_i \sigma_j}.$$6

The subject-level heterogeneous graph is then

$$C_{ij} = corr(x_i, x_j) = \frac{\sum_{t=1}^{T}(x_i(t) - \mu_i)(x_j(t) - \mu_j)}{\sigma_i \sigma_j}.$$7

The central design choice is explicit retention of within-modality and cross-modality structure. This suggests that the model treats structure-function coupling and decoupling as graph relations to be learned rather than nuisances to be averaged away.

3. Attention backbone, pooling, augmentation, and objective

The backbone of Brain-HGCN is HAN, which uses node-level attention within each meta-path and semantic-level attention across meta-paths (Yin et al., 2024). For node $$C_{ij} = corr(x_i, x_j) = \frac{\sum_{t=1}^{T}(x_i(t) - \mu_i)(x_j(t) - \mu_j)}{\sigma_i \sigma_j}.$$8, layer $$C_{ij} = corr(x_i, x_j) = \frac{\sum_{t=1}^{T}(x_i(t) - \mu_i)(x_j(t) - \mu_j)}{\sigma_i \sigma_j}.$$9, and meta-path $|C_{ij}| < 0.2$0, the model computes

$|C_{ij}| < 0.2$1

and then aggregates meta-path-specific embeddings as

$|C_{ij}| < 0.2$2

Here $|C_{ij}| < 0.2$3 are node-level attention weights and $|C_{ij}| < 0.2$4 are semantic-level weights. The paper states that balancing homo- and hetero-meta-paths is handled by semantic-level attention, which automatically assigns higher weights to more informative relations.

A distinguishing component is heterogeneous graph pooling. Separate pooling scores are learned for each node type,

$|C_{ij}| < 0.2$5

followed by dense compression and softmax to obtain $|C_{ij}| < 0.2$6 and $|C_{ij}| < 0.2$7. Zero-padded block selectors $|C_{ij}| < 0.2$8 and $|C_{ij}| < 0.2$9 update adjacency and features in block form. The paper attributes three properties to this design: it respects node-type boundaries, retains hetero edges via cross-block terms, and allows different pooling ratios per modality.

To address class imbalance, the model introduces a biologically constrained heterogeneous graph data augmentation strategy that perturbs FC while keeping SC fixed as a structural anchor. Dynamic FC is reconstructed by sliding windows and triad-based summarization across windows. For each edge $0$0, the global dynamic FC score is computed as

$0$1

and $0$2 is thresholded at $0$3. The augmented FC defines $0$4, while $0$5 remains fixed and $0$6 are recomputed. Original $0$7 and augmented $0$8 are fed as paired samples during training.

The classifier reads out graph embeddings by concatenating global mean and max of node features at each of three backbone stages, concatenates across stages, and feeds the result to an MLP. The primary loss is binary cross-entropy,

$0$9

and the total loss is $A_f \in \mathbb{R}^{90 \times 90}$0. Training uses Adam with initial learning rate $A_f \in \mathbb{R}^{90 \times 90}$1, weight decay $A_f \in \mathbb{R}^{90 \times 90}$2, cosine annealing, batch size 32, and 200 epochs. The architecture has 3 HAN layers, 3 heterogeneous pooling layers, and 3 readout layers, with 8 attention heads, hidden dimension 128, pooling ratio 0.8, dropout 0.45, and PairNorm.

4. Empirical results, ablations, and anatomical interpretation

On ADNI-3, the dual-modal Brain-HGCN reports mean 5-fold cross-validation performance of ACC 93.3%, SEN 93.5%, SPE 93.3%, and AUC 0.961 for NC vs. MCI (Yin et al., 2024). Additional tasks report ACC 92.7%, SEN 90.6%, SPE 96.9%, and AUC 0.927 for NC vs. EMCI; ACC 94.4%, SEN 92.1%, SPE 97.0%, and AUC 0.972 for NC vs. LMCI; and ACC 92.4%, SEN 85.7%, SPE 96.0%, and AUC 0.952 for EMCI vs. LMCI. The paper states that single-modality fMRI SOTA methods achieve ACC 82.0–92.2% and dual-modality methods 82.6–93.0%.

Ablation results on NC vs. MCI isolate the contribution of each component. HAN on node-level hetero-meta-path only yields ACC 0.80, SEN 0.46, SPE 0.94, and AUC 0.66. Adding heterogeneous pooling increases these to ACC 0.86, SEN 0.66, SPE 0.98, and AUC 0.76. Adding community-level hetero edges yields ACC 0.90, SEN 0.77, SPE 0.98, and AUC 0.84. Adding augmentation reaches ACC 0.93, SEN 0.93, SPE 0.93, and AUC 0.96. The paper summarizes these effects as follows: pooling boosts ACC by 11.5% and SEN by 20.1% on average; community-level hetero-meta-path further improves ACC by 7.3% and SEN by 18.1%; augmentation yields balanced SEN/SPE, both greater than 90%, and the best AUC.

Interpretive analyses emphasize structure-function decoupling in transmodal cortex. Visualization within default mode and fronto-parietal networks shows substantial decoupling between SC and FC when only homo-meta-paths are used; adding $A_f \in \mathbb{R}^{90 \times 90}$3 and $A_f \in \mathbb{R}^{90 \times 90}$4 bridges regions separated in SC by leveraging cross-modal edges driven by coupling and shared community structure. Semantic-level attention weights are not reported numerically, but the reported performance gains after adding community-level edges are presented as evidence that hetero paths carry significant diagnostic signal.

The first pooling layer repeatedly selects rs-fMRI ROIs including middle frontal gyrus, superior frontal gyrus, middle temporal gyrus, amygdala, and hippocampus, and DTI ROIs including caudate and pallidum. The paper interprets these selections as consistent with cognitive and limbic alterations in MCI and DTI sensitivity to white matter changes in subcortical circuitry. Reproducibility is supported by explicit graph construction formulas, threshold definitions, TopK selection, triad counting, hyperparameters, and the 5-fold cross-validation protocol, although random seed and hardware are not reported.

5. Other models mapped to the name “Brain-HGCN”

The label “Brain-HGCN” spans several technically distinct graph-learning paradigms across brain imaging studies. The following uses are all explicitly present in the literature synthesized here.

Paper	Meaning of Brain-HGCN	Data and task
(Huang et al., 2023)	HL-HGCNN with Hodge-Laplacian spectral filters and TGPool	ABCD fMRI, general intelligence
(Leng et al., 2023)	DH-ProGCN, described as synonymous with Brain-HGCN	ADNI-1/2 T1-MRI, pMCI vs sMCI
(Tang et al., 2022)	Hierarchical signed graph learning with contrastive augmentation	HCP and OASIS functional networks
(Banka et al., 2020)	Hypergraph convolutional autoencoder on multi-view hyperconnectomes	ADNI GO morphological networks
(Jia et al., 18 Sep 2025)	Hyperbolic graph convolutional network in the Lorentz model	ADHD-200 and ABIDE fMRI
(Yang et al., 2020)	Generic HGCN capsule framework, proposed as a natural mapping to brain graphs	General graph classification

The Hodge-Laplacian formulation jointly models node-level and edge-level functional signals on simplex graphs. It defines $A_f \in \mathbb{R}^{90 \times 90}$5 and $A_f \in \mathbb{R}^{90 \times 90}$6 because $A_f \in \mathbb{R}^{90 \times 90}$7, uses Laguerre polynomial approximations for spectral filters, introduces TGPool as topology-preserving pooling on boundary operators, and reports that HL-HGCNN achieves RMSE $A_f \in \mathbb{R}^{90 \times 90}$8, outperforming GAT, BrainGNN, dGCN, BrainNetCNN, and Hypergraph NN on ABCD intelligence prediction (Huang et al., 2023).

The T1-MRI structural formulation learns subject-specific critical regions as dynamic prototypes from a Convmixer-like 3D CNN with 2048 channels, organizes them hierarchically as $A_f \in \mathbb{R}^{90 \times 90}$9, infers dynamic adjacency by self-attention, and reasons with a two-layer GCN under the total objective $X_f \in \mathbb{R}^{90 \times D}$0. On ADNI-2 after training on ADNI-1, it reports ACC $X_f \in \mathbb{R}^{90 \times D}$1, SEN $X_f \in \mathbb{R}^{90 \times D}$2, SPE $X_f \in \mathbb{R}^{90 \times D}$3, and AUC $X_f \in \mathbb{R}^{90 \times D}$4 for pMCI vs sMCI (Leng et al., 2023).

The signed hierarchical model is explicitly not hyperbolic. It splits functional adjacency into positive and negative channels, computes balanced and unbalanced embeddings under balance theory, performs top-$X_f \in \mathbb{R}^{90 \times D}$5 hierarchical pooling by information scores, and augments data by temporal clamping before an InfoNCE objective. It reports, for example, 81.51$X_f \in \mathbb{R}^{90 \times D}$61.14% accuracy on HCP gender prediction and 77.51$X_f \in \mathbb{R}^{90 \times D}$71.84% accuracy on OASIS AD vs. NC (Tang et al., 2022).

The hypergraph formulation builds view-specific hyperconnectomes using $X_f \in \mathbb{R}^{90 \times D}$8-nearest-neighbor hyperedges over connectivity profiles, stacks incidence matrices across views, encodes them with two hypergraph convolutional layers, reconstructs the incidence matrix, and regularizes latent embeddings adversarially. On ADNI GO morphological networks, it reports 73.69% average classification accuracy for AD vs. MCI, compared with 48.31% for the single-view ACE baseline (Banka et al., 2020).

The hyperbolic formulation explicitly reserves the Brain-HGCN name for a Lorentz-model graph network with learnable curvature, signed top-$X_f \in \mathbb{R}^{90 \times D}$9 sparsification of Pearson-correlation graphs, hyperbolic attention, signed aggregation in tangent space, and graph readout by the Fréchet mean. It reports ACC $S_{ij} < 5 \rightarrow 0$0 and AUC $S_{ij} < 5 \rightarrow 0$1 on ADHD-200, and ACC $S_{ij} < 5 \rightarrow 0$2 and AUC $S_{ij} < 5 \rightarrow 0$3 on ABIDE (Jia et al., 18 Sep 2025).

6. Limitations, misconceptions, and open directions

Several limitations recur across Brain-HGCN formulations. In the dual-modal heterogeneous model, dataset imbalance and modest sample sizes remain challenges, although augmentation is introduced specifically to address the approximate NC:MCI ratio of $S_{ij} < 5 \rightarrow 0$4 (Yin et al., 2024). The paper also notes that thresholds for FC, SC, and dynamic FC, together with the similarity choice for cross-modal coupling, may affect graph topology, and suggests learning thresholds or adopting differentiable sparsification. Community detection is simplified by shared triads, with modularity-based or spectral alternatives presented as future possibilities.

A common misconception is that HGCN always refers to hyperbolic graph convolution. The signed hierarchical paper explicitly distinguishes hierarchical GCN from hyperbolic GCN and states that its model is Euclidean rather than hyperbolic (Tang et al., 2022). Conversely, the 2025 Brain-HGCN paper uses the term in the hyperbolic sense and argues that Euclidean GNNs distort hierarchical organization in functional brain networks (Jia et al., 18 Sep 2025). The term therefore cannot be interpreted geometrically without paper-specific context.

Another recurrent issue is incomplete reporting. The dual-modal heterogeneous paper does not explicitly report parameters, FLOPs, or runtime (Yin et al., 2024). The T1-MRI prototype paper notes a typographical inconsistency in ADNI-2 counts, states that statistical significance testing is not reported, and does not fully enumerate preprocessing details in the main text (Leng et al., 2023). The hypergraph autoencoder does not specify several optimization hyperparameters in the paper text and performs classification post hoc with SVM rather than end-to-end supervision (Banka et al., 2020).

Open directions differ by formulation but converge on several themes: more sophisticated higher-order structure, multimodal integration, dynamic or longitudinal graph modeling, cross-site validation, and stronger robustness analysis. In the Hodge-Laplacian line, higher-dimensional simplices beyond edges are identified as a natural extension (Huang et al., 2023). In the hyperbolic line, dynamic fMRI, multimodal integration, and cross-site harmonization are identified as important next steps (Jia et al., 18 Sep 2025). In the dual-modal heterogeneous line, multi-site validation and learned graph topology remain central unresolved issues (Yin et al., 2024).

Taken together, Brain-HGCN denotes a growing but heterogeneous body of brain-network graph learning research. Its most detailed heterogeneous formulation uses explicit structure-function heterogeneity and meta-path attention for dual-modal MCI diagnosis, while the broader label now encompasses hierarchical, Hodge-Laplacian, hypergraph, and hyperbolic approaches that share a common objective: preserving the nontrivial topology of brain data without reducing it to a single homogeneous graph operator.