MHG-FM jointly generates structural and functional brain connectivity via multimodal hypergraph flow matching, beating several baselines on HCP-YA with about 8x faster sampling
Related research and updates
Synopsis
The work proposes a Multimodal Hypergraph Flow Matching (MHG-FM) framework that builds modality-specific hypergraphs, learns higher-order representations with Hypergraph Neural Network (HGNN) encoders, performs bidirectional cross-modal fusion via Dual Cross-Attention (DCA), and maps fused representations into a compact latent space with a variational autoencoder, where conditional flow matching enables joint structural connectivity (SC) and functional connectivity (FC) generation and cross-modal translation; on the Human Connectome Project Young Adult (HCP-YA) dataset it outperforms several state-of-the-art baselines in reconstruction quality, topology preservation, distributional similarity, and SC-FC coupling, while sampling about 8x faster than a matched diffusion backbone.
Fig. 2 : Ground truth versus generated SC and FC matrices, averaged across all n = 70 n{=}70 test subjects, for MHG-FM (top) and MHG-DiT (bottom). Columns: SC ground truth/generated/residual, FC ground truth/generated/residual.
arXivInterpretation
MHG-FM is introduced for joint SC and FC generation and bidirectional cross-modal translation, rather than generating the two connectivities independently. Existing approaches typically use pairwise graphs that capture only dyadic interactions and often generate SC and FC independently, limiting preservation of higher-order structure-function relationships; this work models higher-order interactions with hypergraphs and generates jointly. At the abstract level, experiments are reported on the HCP-YA dataset, with advantages over several state-of-the-art baselines in reconstruction quality, topology preservation, distributional similarity, and SC-FC coupling; specific metric values are not given in the abstract.
Modality-specific hypergraphs with HGNN encoders learn higher-order representations, and Dual Cross-Attention performs bidirectional cross-modal fusion. It combines higher-order relational modeling with bidirectional cross-modal fusion for the complementary SC-FC modality pair, instead of relying only on pairwise graph structure. The abstract describes the method components (hypergraph construction, HGNN encoders, DCA fusion) but provides no quantitative ablation or component-contribution results.
A variational autoencoder maps fused representations to a compact latent space, and conditional flow matching enables connectivity synthesis and multimodal translation via latent transport. Generation and translation are unified in one latent transport process, and the abstract states sampling is about 8x faster than a matched diffusion backbone. The abstract reports an approximately 8x speedup as a relative comparison; absolute sampling times and hardware settings are not given.
Perspective
The work targets neuroimaging research settings that need large-scale paired SC-FC data, especially studies related to neuropsychiatric disorders; its value lies in supplementing scarce data through generation and cross-modal translation, and in supporting sampling-heavy use cases with about 8x faster sampling. The intended audience is researchers working on brain connectomics and multimodal neuroimaging modeling, under settings comparable to HCP-YA in data and task.
The abstract does not give specific values for each evaluation dimension, the HCP-YA sample size, the baseline list, or the statistical testing approach, nor does it state the measurement conditions for the approximately 8x speedup; the hypergraph construction, the independent contributions of DCA and flow matching, and generalization across datasets or populations all need confirmation in the full text. This document is abstract-level material without figures or experimental details, so the above remains open questions to verify.