Hi-SMGNN: Hierarchical GNN for Glioma Genotype

• Hi-SMGNN is a hierarchical graph learning framework that integrates dual connectomes from diffusion and structural MRI to predict glioma genotype.
• It employs a multimodal interaction module with cross‐modal attention, personalized modular partitioning, and multiscale feature fusion to capture brain connectivity across scales.
• Experiments on the UCSF-PDGM dataset show the model improves accuracy and F1 score by approximately 3% and 6% over leading competitive methods.

Searching arXiv for the Hi-SMGNN paper and closely related context. Tool call: arxiv_search(query="(Tang et al., 13 Aug 2025) OR \"Hierarchical Brain Structure Modeling for Predicting Genotype of Glioma\" OR Hi-SMGNN", max_results=5, sort_by="relevance") Hi-SMGNN, short for Hierarchical Structure-Morphological Graph Neural Network, is a hierarchical graph learning framework for non-invasive prediction of glioma genotype, specifically Isocitrate DeHydrogenase (IDH) mutation status, from multi-modal MRI-derived brain connectomes. It was introduced to address the limitations of functional MRI, described as noise and limited availability, by jointly modeling structural connectomes from diffusion MRI and morphological connectomes from structural MRI within a hierarchical, multimodal, and multiscale architecture. The framework is organized around three principal components: a Multimodal Interaction Module (MIM), a Multiscale Feature Fusion mechanism (MFF), and Personalized Modular Partitioning (PMP), with the stated aim of improving robustness, effectiveness, individual specificity, and interpretability in IDH mutation prediction on the UCSF-PDGM dataset (Tang et al., 13 Aug 2025).

1. Problem setting and conceptual basis

IDH mutation status is presented as a crucial biomarker for glioma prognosis, and Hi-SMGNN is designed to predict that genotype non-invasively from brain connectomic data rather than from functional MRI alone (Tang et al., 13 Aug 2025). The stated motivation is that structural and morphological connectomes offer a non-invasive alternative, while existing approaches often ignore the brain's hierarchical organisation and multiscale interactions.

The framework is explicitly grounded in the proposition that the brain is organized hierarchically, from regions to modules to global organization, and that glioma-related effects can alter both white matter microstructure and gray matter morphology at different scales. In that formulation, a model restricted to only local or only global structure is treated as incomplete. Hi-SMGNN therefore operates simultaneously at the regional, modular, and global levels.

A recurrent misconception is to treat Hi-SMGNN as merely a multimodal fusion network. In the reported formulation, multimodal fusion is only one part of the system. The method also imposes a hierarchical decomposition of brain organization and introduces patient-specific module learning. This distinguishes it from approaches that concatenate modalities without an explicit multiscale or modular mechanism.

2. Connectomic representation and hierarchical organization

Hi-SMGNN takes as input two connectomes. The Structural Connectome (SC) is built from dMRI and FA images, with edges quantifying white matter tract strength, described as fiber integrity, between brain regions. The Morphological Connectome (MC, also denoted R2SN) is constructed by computing interregional similarities, for example Pearson correlation, between radiomic features derived from structural MRI (T1c), thereby encoding morphometric similarity between regions (Tang et al., 13 Aug 2025).

Each connectome is represented as a graph

$G = (V, X, A),$

where $V$ denotes brain regions, $X$ denotes node features initialized from adjacency matrix rows, and $A$ denotes the adjacency matrix of connectivity.

The hierarchical organization is defined across three scales:

• Regional level: subgraphs centered at each node with local neighbors.
• Modular level: groups of nodes with strong intra-connectivity and functional or structural similarity, learned adaptively for each subject.
• Global level: a summary of fused information from all modules and regions.

The rationale given for this hierarchy is that brain function and pathology operate across scales. According to the reported description, hierarchy enables capture of fine-grained and large-scale patterns, multimodal interactions at different organizational levels, and suppression of noise and redundancy in high-dimensional biomedical data. This suggests that the architecture is designed not only for classification accuracy but also for a structured decomposition of disease-related signal across levels of organization.

3. Multimodal Interaction Module

The Multimodal Interaction Module is the component responsible for integrating SC and MC so as to model fine-grained cross-modal brain alterations at different scales. Its two stated mechanisms are a Siamese GNN encoder and cross-modal attention (Tang et al., 13 Aug 2025).

For a subgraph $j$, either regional or modular, the SC and MC inputs are processed by shared-parameter GNN encoders: $$Z_{S}^{j} = \Phi(G_{S}^{j}, \theta), \qquad Z_{Mor}^{j} = \Phi(G_{Mor}^{j}, \theta).$$

Cross-modal attention is then defined by using the morphological representation as query and the structural representation as key and value: $$Q_{j} = W_{Q} Z_{Mor}^{j}, \qquad K_{j} = W_{K} Z_{S}^{j}, \qquad V_{j} = W_{V} Z_{S}^{j},$$

$$Z_{j} = \text{softmax}\left( \frac{Q_{j}K_{j}^T}{\sqrt{d_k}} \right) V_{j}.$$

The reported interpretation is that SC provides the structural scaffold for MC, so the more stable structural modality guides attention over the more variable morphological modality. In operational terms, the module produces regional interaction representations $z_r^i$ and modular interaction representations $z_m^k$. The framework attributes a central role to this design in extracting interaction features, and the ablation study reports that removing MIM reduces performance (Tang et al., 13 Aug 2025).

This formulation should not be confused with symmetric late fusion. The attention mechanism is directional: structural information guides morphological attention. A plausible implication is that the model encodes an explicit prior about the relative stability of the two connectomes.

4. Personalized Modular Partitioning

Personalized Modular Partitioning is introduced to reflect individualized modular organization of the brain, under the premise that tumor impact on connectivity and morphology can vary across patients. Its purpose is stated as increasing both specificity and interpretability (Tang et al., 13 Aug 2025).

The module takes regional interaction embeddings $V$0 together with the MC graph and produces a soft module assignment matrix: $V$1 where thresholding with $V$2 is used to sharpen the soft assignment.

Rather than optimizing modules from MC topology alone, Hi-SMGNN computes modularity using both SC and MC. The modularity-based objective is

$V$3

where $V$4 is the modularized adjacency matrix, $V$5 is the node degree vector from $V$6, $V$7 is total edge weight, and $V$8 indexes the connectome type, SC or MC. The total training objective combines modularity optimization with classification: $V$9

The reported significance of PMP is twofold. First, it replaces fixed modular divisions with per-patient learned partitions. Second, because the partitions are constrained by modularity, the learned modules are intended to align with brain topology. The paper description links this directly to neuroscientific interpretability. It also notes that removing PMP in ablation reduces performance, while the full model gives the highest metrics (Tang et al., 13 Aug 2025).

A common misunderstanding would be to view PMP as a generic clustering layer. In the reported framework, it is a supervised, task-coupled, modularity-regularized partitioning mechanism that uses multimodal topology.

5. Multiscale feature fusion and global representation

The Multiscale Feature Fusion mechanism addresses the challenge that direct concatenation of regional and modular representations leads to redundancy, particularly because regional features are described as noisier and higher-dimensional (Tang et al., 13 Aug 2025). MFF is therefore framed as a denoising and compression mechanism as much as a fusion operator.

The procedure first maps modular representation $X$0 to the regional scale: $X$1

A soft-thresholding operation is then applied to each regional feature: $X$2

The purified regional representation $X$3 is concatenated with $X$4 and averaged to form the global interaction representation: $X$5

This mechanism is described as actively suppressing redundancy. In the interpretation provided with the model, the thresholding step removes features below threshold and thereby reduces irrelevant or noisy regional components before global summarization. The ablation study attributes an important role to MFF, reporting that removing it lowers performance and that the full model confirms multiscale fusion effectively reduces redundant information (Tang et al., 13 Aug 2025).

6. Empirical results, robustness, and interpretability

The experimental evaluation is reported on the UCSF-PDGM cohort, consisting of 103 IDH-mutant and 392 IDH-wildtype cases, using T1c and FA MRI (Tang et al., 13 Aug 2025). Baseline comparisons include MLP, GraphSAGE, DiffPool in single-modality and multimodal variants, as well as recent multi-modal graph frameworks Cross-GNN, Joint-GCN, and sf-GBT. The evaluation metrics are Accuracy (ACC) and F1-score.

Method group	ACC (\%)	F1 (\%)
Unimodal GNNs	66.67–73.74	up to 35.29
Multimodal GNNs	72.73–81.82	up to 66.67
Hi-SMGNN	84.85	72.73

The paper states that Hi-SMGNN outperforms all baselines and state-of-the-art competitors, improving accuracy by approximately $X$6 and F1 by approximately $X$7 over the next best model, sf-GBT (Tang et al., 13 Aug 2025). Within the study, the full model yields the highest metrics, and removing MIM, MFF, or PMP reduces performance.

The reported robustness claim is that the model performs consistently better across various splits, with the benefit attributed to adaptive modules and denoising or fusion mechanisms. The interpretability claim is that personalized modular partitions link model features to recognizable brain modules, while the modularity-based loss guides these modules to align with brain topology. This does not imply mechanistic biological validation in itself; rather, the paper presents interpretability in the sense of mapping learned representations onto modular brain organization.

More broadly, the framework is positioned as a hierarchical, multimodal, and multiscale alternative to approaches that either rely on noisier functional imaging or disregard the interaction between structural and morphological connectomes. Its reported contribution lies in combining regional-to-modular modeling, cross-modal attention, redundancy-suppressing fusion, and individualized module learning within a single graph neural network architecture for IDH mutation prediction (Tang et al., 13 Aug 2025).