Spectral-Temporal State Space Modeling on Functional Brain Networks

Jaeyoon Sim1*, Hoseok Lee1*, Jihwan Park1, Seunghun Baek1, Yu Zhang2, Won Hwa Kim1

(* : Equal Contribution)

1POSTECH, 2Stanford University

MICCAI 2026

:france: Strasbourg, France

Paper Poster

Illustration of overall framework (SpecT-Mamba).

Abstract

Resting-state fMRI (rs-fMRI) offers a powerful tool for analyzing functional organization in brain for neurodegenerative and neurodevelopmental disorders. Although graph-based and spatio-temporal models have shown promise, most existing approaches decouple spatial structure from temporal dynamics or rely on predefined temporal windows, limiting the capacity of spatial information to directly modulate temporal processing. To address these limitations, we propose a spectral-temporal state space framework that integrates graph spectral representations with state-space modeling through a spectral-temporal kernel, enabling window-free and frequency-based temporal modeling. Experiments on three rs-fMRI cohorts with 1,926 subjects demonstrate that our method consistently outperforms competitive baselines across diverse conditions, including Parkinson’s disease, autism spectrum disorder, and attention deficit hyperactivity disorder. In addition, we provide interpretable insights into spectral-temporal patterns of brain dysfunction, advancing the characterization and classification of neurodegenerative and neurodevelopmental disorders.

Quantitative Result

Table: Comparison of classification performance between static (top) and spatio-temporal (bottom) methods on the three rs-fMRI brain network benchmarks. The best, second-best, and third-best results are highlighted.

Analysis on the Interpretable Results

Figure: Top: The Grad-CAM results describing the top-10 ROIs with the highest activation (Act.) to classify disease on the PPMI (Left) and ADHD-200 (Right) datasets. The indices align with the index values in the AAL116 atlas, and (L) and (R) denote the left and right hemispheres, respectively. Bottom: Top-10 ROIs with the highest activation for classifying disease on the PPMI (Left) and ADHD-200 (Right). Node color indicates the activation.

Spectral-Temporal Dynamics via Kernels

Figure: Visualization of the learned spectral-temporal kernel weights on the three benchmarks. We fix the current time to the last step and plot the kernel weight as a function of the past index, where spectral indices are grouped into low-, middle-, and high-frequency bands.

Effect of Spectral–Temporal Coupling Parameters

Table: Ablation study on spectral-temporal coupling parameters of SpecT-Mamba.

Conclusion

In this work, we proposed SpecT-Mamba, a spectral-temporal framework that conditions temporal evolution on graph spectral structure. By modeling a multi-scale state-space kernel to modulate signal amplitude and temporal decay, our model captures both short- and long-range dependencies. Thus, SpecT-Mamba learns topology-aware temporal representations in which spectral channels encode dynamics relevant to neurological and neuropsychiatric disorders. Experiments on rs-fMRI datasets show that SpecT-Mamba outperforms baselines while providing insights into spectral-temporal patterns of brain dysfunction.

BibTeX

@inproceedings{sim2026spectral,
      title={Spectral-Temporal State Space Modeling on Functional Brain Networks},
      author={Sim, Jaeyoon and Lee, Hoseok and Park, Jihwan and Baek, Seunghun and Yu Zhang and Kim, Won Hwa},
      booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
      year={2026}
    }