Jul 2026· Medical Image Anal.· Vol 113, pp.
104218
· 0 citations· 67 references
MedicineComputer Science
TL;DR
Findings indicate that anatomically grounded node-level multimodal fusion can improve classification performance while preserving neurobiological interpretability, thereby providing a principled framework for multimodal schizophrenia classification and biomarker discovery.
Abstract
We propose TB-GCAN, a tri-branch cross-attention graph neural network for schizophrenia classification using multimodal MRI, including sMRI, fMRI, and DTI. Built on a multi-site dataset of 1191 samples from seven scanning sites, the model exploits atlas-defined one-to-one anatomical correspondence across modalities to enable node-level cross-attention during intermediate representation learning. In 7-site leave-one-site-out evaluation, TB-GCAN achieved 84.63% accuracy and outperformed GAT, GCN, CNN, SVM, and MMGNN in the tri-modal setting. Attention-based region ranking highlighted biologically plausible schizophrenia-related regions, and downstream analyses linked the learned imaging representations to PANSS dimensions and transcriptional programs. Unlike generic multimodal GNNs that learn cross-modal relations from data, TB-GCAN directly leverages atlas-aligned regional correspondence to perform anatomically constrained node-level interaction. These findings indicate that anatomically grounded node-level multimodal fusion can improve classification performance while preserving neurobiological interpretability, thereby providing a principled framework for multimodal schizophrenia classification and biomarker discovery.
MOTIVATION
Hypergraph-based models for brain disorder prediction mainly adopt imaging-derived hypergraphs as propagation backbones. However, the entanglement of topology construction and feature propagation leaves regional representations weakly constrained by underlying biological organization, making them vulnerable to subject-specific variation and noise, particularly in heterogeneous multimodal settings.
RESULTS
We present BriGHT, a Brain transcriptome-reGularized Hypergraph framework for mulTimodal disorder prediction. BriGHT employs a transcriptome-derived structural reference as a soft anchoring prior to regularize neuroimaging ROI embeddings, stabilizing representation geometry while preserving disease-relevant subject-specific variation. BriGHT further incorporates a reliability-aware fusion module to estimate subject-specific modality reliability from prediction confidence, cross-modal consistency, and decision certainty, enabling adaptive integration under heterogeneous modality quality. Experiments on three neuroimaging cohorts (ADNI, ADHD-200, REST-meta-MDD) and four modalities (VBM, fMRI, FDG, AV45) demonstrate that BriGHT consistently outperforms competing graph/hypergraph learning methods across six brain disorder prediction tasks. Perturbation analyses show that BriGHT benefits from the spatial correspondence between transcriptomic modules and imaging ROIs, rather than from arbitrary hypergraph regularization alone. Ablation and meta-analytic interpretability analyses support the contribution of transcriptomic anchoring and adaptive fusion to robust and biologically meaningful brain disorder prediction.
AVAILABILITY
The software is publicly available at: https://github.com/Yaolab-fantastic/BriGHT.
SUPPLEMENTARY INFORMATION
Supplementary data are available at Bioinformatics online.
Neurodegenerative diseases such as Alzheimer's disease (AD) and Parkinson's disease (PD) are characterised by progressive, stage-dependent disruptions in brain connectivity. Multimodal neuroimaging data, particularly functional MRI (fMRI) and diffusion tensor imaging (DTI), provide complementary perspectives on functional and structural brain organisation. However, most existing graph-based approaches compress whole-brain networks into a single global representation, limiting their capacity to model hierarchical connectivity patterns and to deliver interpretable insights for biomarker discovery. In this work, we propose HiBrain, a hierarchical prototype-based framework for multimodal brain network analysis. HiBrain explicitly represents brain graphs at node-, graph-, and stage-level abstractions, while preserving the distinct structural and functional connectivity characteristics throughout the hierarchy. The proposed framework progressively abstracts representative local connectivity patterns into global network representations and disease-stage–specific prototypes, enabling both accurate stage-aware classification and principled interpretability. Experiments on multimodal AD and PD datasets demonstrate consistent improvements over state-of-the-art baselines in multi-stage classification tasks. Moreover, prototype-driven visualisations of connectivity difference matrices and biomarker subgraphs reveal clear and stage-specific brain network signatures, highlighting the interpretability and scientific utility of the proposed framework for neurodegenerative disease analysis. The source code is available at https://github.com/yangkf825/HiBrain.
Jing Ren, Kefan Yang, Le Linh Dan Nguyen et al.· Proceedings of the 32nd ACM...· 1 citation
The rapid growth of multimodal neuroimaging research has produced fragmented literature that limits systematic characterization of cross-modal relationships and disease-specific knowledge structures. To address this, we constructed a multimodal neuroimaging knowledge graph from 1838 peer-reviewed studies (2016–2026) spanning fMRI, EEG, fNIRS, and PET, using an LLM-based extraction and retrieval-augmented semantic merging pipeline. The resulting graph comprised 4190 nodes and 7007 edges, exhibiting a scale-free topology with a dominant connected component covering 76.6% of nodes. Alzheimer’s disease, the hippocampus, and fMRI/PET emerged as the most central hubs linking disease, anatomical, and methodological dimensions. Louvain community detection identified 25 functional modules, with seven major communities—centered on Alzheimer’s biomarker integration, molecular/fluid imaging, and psychiatric functional connectivity—forming the field’s core structure. Cross-modal analysis revealed the strongest coupling between fMRI and PET, indicating high methodological convergence. At the disease level, Alzheimer’s disease displayed a mature, hierarchically organized biomarker system, whereas major depressive disorder and chronic pain showed diffuse, less consolidated knowledge structures. These results reveal pronounced disparities across neuroimaging research domains and demonstrate that LLM-augmented knowledge graphs can systematically uncover latent structural organization relevant to multimodal integration and biomarker discovery.
Zhenni Liu, Hanzhen Ouyang, Xuan-Zi Liu et al.· Bioengineering· 0 citations
Accurate classification of brain disorders from neuroimaging data remains challenging because of substantial inter-subject heterogeneity and the complex multi-scale patterns present in functional connectivity and morphological representations. To address these challenges, we propose HyperAMS-Net, a deep learning framework for brain disorder classification using neuroimaging representations derived from resting-state functional MRI or structural MRI. HyperAMS-Net integrates adaptive multi-scale convolution, hypergraph attention, spatial-channel attention, and adaptive feature fusion. Specifically, adaptive multi-scale convolution learns data-driven weights over multiple receptive fields to capture complementary patterns at different scales. Hypergraph attention models higher-order dependencies among learned feature representations through node--hyperedge--node message passing, while spatial-channel attention enhances discriminative feature learning. Adaptive feature fusion further aggregates complementary information across parallel network branches. HyperAMS-Net is evaluated on three benchmark datasets spanning distinct brain disorders: ABIDE for autism spectrum disorder, REST-meta-MDD for major depressive disorder, and ADNI for Alzheimer's disease, using 5-fold stratified cross-validation. HyperAMS-Net achieves state-of-the-art performance across all evaluated datasets, attaining the highest accuracy and AUC among the compared methods. Ablation studies further demonstrate the contribution of each proposed component, with the largest performance degradation observed when hypergraph attention is removed.
GraM-Diff is proposed, a unified classifier-guided Graph-Mamba diffusion framework for EEG synthesis that embeds Graph Convolutional Networks within a diffusion U-Net to model inter-electrode connectivity and Bidirectional Mamba state-space blocks for linear-complexity long-range temporal modeling.
M. Tanveer, A. Rana, Sanskriti Jain et al.· 0 citations
Background/Objectives: Deep learning has achieved remarkable success in medical image analysis; however, limited model interpretability remains a major barrier to its clinical translation. MRI-negative temporal lobe epilepsy (TLE) is characterized by the absence of readily identifiable structural abnormalities on conventional MRI. Methods: We propose a hierarchical, multiscale 3D residual network (H-MSResNet) combined with layer-wise relevance propagation (LRP). The study included structural T1-weighted MRIs from 101 patients with MRI-negative TLE and 101 healthy controls. Model classification performance was evaluated using a fivefold cross-validation approach. Subsequently, group-level LRP analysis was integrated with a standard brain atlas to quantify the anatomical regions contributing to the model’s decisions. Results: H-MSResNet achieved an average classification accuracy of 76.27% and a best single-fold accuracy of 82.50%, with higher accuracy, specificity, and F1 score but lower sensitivity and AUC than the two comparison models. Group-level, LRP-based analysis combined with a standard brain atlas revealed that the model’s decision-making primarily focused on structures related to the temporal lobe and limbic system, including regions such as the hippocampus, parahippocampal gyrus, and amygdala. Population-level attribution also showed interhemispheric differences across several regions. Conclusions: Structural MRIs of MRI-negative TLE contain latent discriminative information that can be recognized by deep learning models. The H-MSResNet and LRP framework provides viable, explainable methodological support for the computer-aided diagnosis and brain region analysis of MRI-negative epilepsy.
He Wang, Yi-Lin Jiang, Kai-Yue Wu et al.· Diagnostics· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
