pith. sign in

arxiv: 2605.30387 · v1 · pith:TOGYMRJMnew · submitted 2026-05-28 · 💻 cs.LG · cs.AI· cs.CV· eess.SP

Functional MRI Time Series Generation via Wavelet-Based Image Transform and Spectral Flow Matching for Brain Disorder Identification

Hwa Hui Tew , Junn Yong Loo , Fang Yu Leong , Julia K. Lau , Ding Fan , Hernando Ombao , Rapha\"el C.-W. Phan , Chee Pin Tan
show 1 more author
Chee-Ming Ting
This is my paper

Pith reviewed 2026-06-29 09:12 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CVeess.SP
keywords fMRI generationflow matchingwavelet transformdiscrete cosine transformBOLD signalsbrain disorder classificationgenerative modelsspectral flow matching
0
0 comments X

The pith

Cascading wavelet and cosine transforms with spectral flow matching generates synthetic fMRI signals that improve brain network classification.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

fMRI data is scarce because scanning is expensive, which limits training of models that identify brain disorders from BOLD time series. The paper proposes Dual-Spectral Flow Matching, a method that first decomposes signals with a discrete wavelet transform to capture multi-scale transients, then projects them into discrete cosine transform space for compact low-frequency representation across regions and time. A flow matching model is trained on these representations to produce class-conditioned frequency maps. Inverse transforms recover time-domain signals that retain physiological structure. The resulting synthetic data yields higher accuracy in downstream classification of brain networks than earlier generative approaches.

Core claim

The central claim is that converting BOLD signals to a dual wavelet-DCT frequency representation lets a spectral flow matching model learn structured, class-specific priors; inverting the generated maps then produces time-domain signals that preserve key physiological dynamics and measurably raise performance on fMRI-based brain disorder classification tasks.

What carries the argument

Dual-Spectral Flow Matching (DSFM), the cascade of discrete wavelet transform for multi-scale variations followed by discrete cosine transform for energy compaction, then spectral flow matching to generate class-conditioned cosine-frequency maps.

If this is right

  • Class-conditioned generation supplies disorder-specific synthetic samples without violating frequency structure.
  • The dual-transform priors reduce the non-stationarity problem that defeats many standard generative models on BOLD data.
  • Reconstructed signals can be used directly as data augmentation for any downstream network classifier.
  • The same pipeline supports generation at multiple scales because the wavelet step already encodes transient and multi-resolution features.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same dual-transform route could be tested on EEG or MEG time series where similar non-stationarity appears.
  • If the frequency priors hold, the generated data might also support simulation studies of how network connectivity changes under different disorder subtypes.
  • A natural next measurement is whether the synthetic series match real data on standard physiological statistics such as power spectra in specific bands.
  • Privacy-sensitive clinical datasets could be expanded by releasing only the trained flow model rather than raw scans.

Load-bearing premise

The generated cosine-frequency maps, once inverted through DCT and DWT, remain free of artifacts that would make the reconstructed BOLD signals less useful or less realistic than real data for classification.

What would settle it

An experiment in which adding the generated samples to real training data produces no gain in classification accuracy on held-out real fMRI, or in which quantitative metrics detect systematic non-physiological artifacts in the reconstructed time series.

Figures

Figures reproduced from arXiv: 2605.30387 by Chee-Ming Ting, Chee Pin Tan, Ding Fan, Fang Yu Leong, Hernando Ombao, Hwa Hui Tew, Julia K. Lau, Junn Yong Loo, Rapha\"el C.-W. Phan.

Figure 1
Figure 1. Figure 1: The pipeline of DSFM. ROI-based BOLD time series are first extracted, followed by DWT [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Original (Rows 1&3) vs. synthetic BOLD signals (Rows 2 &4) and generated normalized [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: We plot the 2D t-SNE embedding of HC and MDD synthetic data generated by our method (top left). Then, we compare with the distributions using Jensen-Shannon Divergence and probability density functions (top right and bottom). Classification Score. To validate the fidelity of the generated samples, we evaluate the classifica￾tion performance of BrainNetCNN (Kawahara et al., 2017), comparing DSFM to GAN and … view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of the average resting-state hemodynamic response function (rsHRF) and [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: a) Group-averaged connectivity patterns of real and synthetic HC/MDD connectivity pat [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of univariate and multivariate spectral representations: ImagenTime/T2I-Diff [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Frequency-specific functional connectivity (FC) matrices for healthy controls (HC) and [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
read the original abstract

Functional Magnetic Resonance Imaging (fMRI) provides non-invasive access to dynamic brain activity by measuring blood oxygen level-dependent (BOLD) signals over time. However, the resource-intensive nature of fMRI acquisition limits the availability of high-fidelity samples required for data-driven brain analysis models. While modern generative models can synthesize fMRI data, they often remain challenging in replicating their inherent non-stationarity, intricate spatiotemporal dynamics, and physiological variations of raw BOLD signals. To address these challenges, we propose Dual-Spectral Flow Matching (DSFM), a novel fMRI generative framework that cascades dual frequency representation of BOLD signals with spectral flow matching. Specifically, our framework first converts BOLD signals into a wavelet decomposition map via a discrete wavelet transform (DWT) to capture globalized transient and multi-scale variations, and projects into the discrete cosine transform (DCT) space across brain regions and time to exploit localized energy compaction of low-frequency dominant BOLD coefficients. Subsequently, a spectral flow matching model is trained to generate class-conditioned cosine-frequency representation. The generated samples are reconstructed through inverse DCT and inverse DWT operations to recover physiologically plausible time-domain BOLD signals. This dual-transform approach imposes structured frequency priors and preserves key physiological brain dynamics. Ultimately, we demonstrate the efficacy of our approach through improved downstream fMRI-based brain network classification. The code is available at https://github.com/htew0001/DSFM.git .

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces Dual-Spectral Flow Matching (DSFM), a generative model for fMRI BOLD time series. It first applies discrete wavelet transform (DWT) to capture multi-scale transients and global variations, then projects into discrete cosine transform (DCT) space for low-frequency energy compaction across regions and time. A class-conditioned spectral flow matching model is trained in this domain; generated cosine-frequency samples are inverted via inverse DCT and inverse DWT to produce time-domain signals asserted to be physiologically plausible. The framework is evaluated by showing improved performance on downstream fMRI-based brain network classification for disorder identification. Code is released at the cited GitHub repository.

Significance. If the dual-transform pipeline produces signals that retain BOLD non-stationarity, low-frequency dominance, and multi-scale structure without introducing classification artifacts, the method could meaningfully address data scarcity in neuroimaging. Public code availability is a concrete strength that enables direct reproducibility checks.

major comments (2)
  1. [Abstract] Abstract: The central claim that the approach yields 'improved downstream fMRI-based brain network classification' is asserted without any reported accuracy/F1 values, baseline comparisons (e.g., against vanilla diffusion or VAE generators), ablation results, or statistical tests. This absence makes it impossible to determine whether gains are attributable to the dual-transform priors or to generic data augmentation.
  2. [Abstract (reconstruction paragraph)] Abstract (reconstruction paragraph): The assertion that inverse DCT + inverse DWT 'recover physiologically plausible time-domain BOLD signals' that 'preserve key physiological brain dynamics' lacks any quantitative fidelity checks (power-spectrum match in the 0.01–0.1 Hz band, autocorrelation structure, or variance bounds typical of real BOLD). Because downstream classification improvements could arise from any augmented data rather than faithful samples, this validation step is load-bearing for the paper's efficacy claim.
minor comments (1)
  1. [Abstract] The acronym 'DSFM' is introduced after the full name; ensure the abbreviation is defined on first use and used consistently in all subsequent sections.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment point-by-point below, with plans to revise the abstract for greater specificity and to add supporting quantitative details where appropriate.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that the approach yields 'improved downstream fMRI-based brain network classification' is asserted without any reported accuracy/F1 values, baseline comparisons (e.g., against vanilla diffusion or VAE generators), ablation results, or statistical tests. This absence makes it impossible to determine whether gains are attributable to the dual-transform priors or to generic data augmentation.

    Authors: The experimental section of the manuscript (Section 4) reports the requested details, including accuracy and F1 scores, comparisons against vanilla diffusion and VAE generators, ablation studies isolating the dual-transform components, and statistical significance tests. To make the abstract self-contained, we will revise it to briefly cite key quantitative gains (e.g., relative accuracy improvement) and reference the baselines and ablations. revision: yes

  2. Referee: [Abstract (reconstruction paragraph)] Abstract (reconstruction paragraph): The assertion that inverse DCT + inverse DWT 'recover physiologically plausible time-domain BOLD signals' that 'preserve key physiological brain dynamics' lacks any quantitative fidelity checks (power-spectrum match in the 0.01–0.1 Hz band, autocorrelation structure, or variance bounds typical of real BOLD). Because downstream classification improvements could arise from any augmented data rather than faithful samples, this validation step is load-bearing for the paper's efficacy claim.

    Authors: We agree that explicit fidelity metrics strengthen the physiological-plausibility claim and help distinguish the method from generic augmentation. In the revised manuscript we will add a dedicated paragraph (or subsection) reporting power-spectrum correlation in the 0.01–0.1 Hz band, autocorrelation-function similarity, and variance bounds relative to real BOLD signals. revision: yes

Circularity Check

0 steps flagged

No circularity: forward generative pipeline with independent downstream validation

full rationale

The paper describes a forward pipeline (DWT to capture multi-scale transients, DCT for low-frequency compaction, spectral flow matching on the cosine-frequency domain, then inverse transforms) that generates class-conditioned samples. No equations, fitted parameters, or self-citations are shown that reduce the claimed preservation of BOLD dynamics or the classification improvement to a tautology or input by construction. The method is presented as a novel composition of standard transforms plus flow matching, with efficacy asserted via external downstream task performance rather than any self-referential fit or uniqueness theorem. This is the common case of a self-contained methodological proposal.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Review performed on abstract only; no explicit free parameters, axioms, or invented entities are stated. The framework implicitly assumes that BOLD signals admit useful multi-scale wavelet and low-frequency DCT representations and that flow matching can learn their joint distribution without additional regularization details.

axioms (2)
  • domain assumption BOLD signals exhibit globalized transient and multi-scale variations that are captured by discrete wavelet transform.
    Invoked in the first step of the pipeline description.
  • domain assumption Low-frequency dominant BOLD coefficients exhibit localized energy compaction under discrete cosine transform.
    Invoked when projecting into DCT space.

pith-pipeline@v0.9.1-grok · 5833 in / 1357 out tokens · 22198 ms · 2026-06-29T09:12:41.820655+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

6 extracted references · 3 canonical work pages · 2 internal anchors

  1. [1]

    A Fourier Space Perspective on Diffu- sion Models,

    URLhttps://arxiv.org/abs/2505.11278. Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Communications of the ACM, 63(11):139–144, 2020. Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Im- proved training o...

    work page arXiv 2020

  2. [2]

    DiffWave: A Versatile Diffusion Model for Audio Synthesis

    URLhttps://proceedings.neurips.cc/paper_files/paper/2020/ file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf. Emiel Hoogeboom and Tim Salimans. Blurring diffusion models. InThe Eleventh International Conference on Learning Representations, 2023. URLhttps://openreview.net/forum? id=OjDkC57x5sz. 11 Published as a conference paper at ICLR 2026 Yang Hu, Xiao Wan...

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  3. [3]

    Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

    URLhttps://openreview.net/forum?id=vDoAA8xKXL. Fuad Noman, Sin-Yee Yap, Rapha¨el C.-W. Phan, Hernando Ombao, and Chee-Ming Ting. Graph autoencoder-based embedded learning in dynamic brain networks for autism spectrum disorder identification. In2022 IEEE International Conference on Image Processing (ICIP), pp. 2891– 2895, 2022. Fuad Noman, Chee-Ming Ting, ...

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  4. [4]

    Flow Matching, Fourier 1.00±0.20 107.90±14.35 0.46±0.02 0.05±0.00

  5. [5]

    Flow Matching, Wavelet 2.19±0.34 41.88±3.37 0.29±0.18 0.04±0.00

  6. [6]

    Table 8 presents the complete ablation results on the MDD dataset with additional experiments on different spectral representations and type of generative models

    Diffusion, Wavelet 3.57±0.68 44.55±3.83 0.48±0.02 0.04±0.00 DSFM (Ours) 0.10±0.01 18.20±1.41 0.17±0.05 0.04±0.00 E.2 FULLRESULTS OFABLATIONSTUDIES. Table 8 presents the complete ablation results on the MDD dataset with additional experiments on different spectral representations and type of generative models. In comparison, the Fourier repre- sentation ac...

    2062