acceptodds
Under review as a conference paper at ICLR 2027

Decoding Beyond Hemodynamic Delay: Delay-Coupled Graph Dynamics for fMRI-to-Image Reconstruction

Abstract

Reconstructing visual stimuli from functional magnetic resonance imaging (fMRI) requires extracting stimulus-related information from spatially distributed blood-oxygen-level-dependent (BOLD) responses shaped by hemodynamic delay and temporal dispersion. We propose DCG-CHI, a framework that combines delay-coupled graph representation learning with hemodynamic consistency training. Its Delay-Coupled Graph State-Space Model (DCG-SSM) embeds graph diffusion within delayed state transitions, explicitly coupling propagation scales to temporal lags and combining propagation modes through input-dependent weights. Counterfactual Hemodynamic Invariance (CHI) synthesizes two BOLD views of a shared latent trajectory through a parameterized pair of hemodynamic response kernels. Cross-view agreement, stop-gradient anchoring, and synthetic BOLD re-synthesis encourage the shared encoder to recover consistent representations. Independent semantic and structural readouts condition a pretrained Stable Diffusion 2.1 backbone through cross-attention and a trainable ControlNet branch, respectively. CHI is used only during training. Experiments on NSD and THINGS-fMRI demonstrate competitive structural fidelity and semantic consistency, with SSIM scores of 0.468 and 0.366 and CLIP identification scores of 0.963 and 0.755, respectively. Ablations and controlled comparisons support the contributions of both components, while additional tests show improved reconstruction under hemodynamic response shifts.

Primary area:
applications to neuroscience & cognitive science
Keywords:
fMRI-to-image reconstruction, visual brain decoding, graph state-space models, hemodynamic invariance, latent diffusion models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_PsvdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DCG-CHI, a framework for reconstructing visual stimuli from fMRI signals by explicitly modeling temporal hemodynamic effects and spatial relationships among brain regions. The proposed Delay-Coupled Graph State-Space Model integrates graph diffusion with delayed state transitions to capture spatial and temporal dependencies in BOLD signals. In addition, Counterfactual Hemodynamic Invariance introduces synthetic hemodynamic transformations and consistency objectives to encourage stable representations across different response functions. The learned representations are used to condition a pretrained diffusion model through separate semantic and structural pathways. Experiments on NSD and THINGS-fMRI demonstrate competitive reconstruction performance and improved robustness under simulated hemodynamic variations.

Critical strengths
  1. The paper addresses a meaningful challenge in fMRI-based visual reconstruction. Hemodynamic delay and temporal dispersion introduce uncertainty into the relationship between neural activity and observed BOLD signals, and explicitly considering these effects is a well-motivated research direction.

  2. The proposed combination of delay-coupled graph modeling and hemodynamic consistency learning is technically interesting. Modeling spatial propagation jointly with temporal delays provides a structured alternative to conventional temporal encoding approaches.

  3. The experimental results are encouraging. The evaluation includes multiple reconstruction metrics on NSD and THINGS-fMRI, and the ablation studies provide evidence that both the temporal graph modeling and consistency objectives contribute to the reported performance.

Critical weaknesses
  1. The proposed framework combines several existing modeling ideas, including graph diffusion, state-space modeling, consistency regularization, and pretrained diffusion-based reconstruction. Although the integration is technically reasonable, the paper does not sufficiently clarify which aspects represent a distinct methodological advance over existing spatiotemporal graph models.

  2. The physiological interpretation of the proposed delay-coupled dynamics requires stronger justification. Coupling graph diffusion scales with temporal lags does not necessarily establish that the learned representations correspond to actual neural propagation or hemodynamic processes. The paper would benefit from additional analysis connecting the learned dynamics to established physiological properties.

  3. The robustness evaluation relies partly on parameterized hemodynamic transformations. It remains unclear whether improvements under these synthetic shifts translate to naturally occurring hemodynamic variability across subjects, sessions, and acquisition protocols. Additional experiments using real cross-session or cross-subject variation would provide stronger evidence for the claimed robustness.

Questions for authors
  1. How sensitive is the proposed model to the choice of functional connectivity graph and the number of temporal lags?

  2. Can the authors provide evidence that the learned delay-dependent graph dynamics capture meaningful physiological relationships rather than serving primarily as a flexible temporal feature extractor?

  3. How does the proposed hemodynamic consistency objective perform under naturally occurring response variability, rather than only simulated HRF perturbations?

Reviewer_UCTVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces DCG-CHI, a method for image reconstruction from fMRI BOLD signals, processed as the mean activation time series within cortical ROIs. DCG-SSM, a state-space based model, uses the BOLD time series around the timing of each visual stimuli to address the inherent HRF delay in the BOLD signal. Specifically, it processes these signals on a fixed correlation graph, imposing a prior in which older hidden states receive stronger graph smoothing, while learning temporal delay weights and diffusion rates across regions. The complementary part of the method, CHI, is a training-only component used to augment the training with varying HRF formulations. For each sample, it generates two synthetic observations from the encoded state-space trajectory using two HRF filters, training the model an invariance to different HRF types that are abundant in BOLD recordings. Semantic and structural attention readouts condition a frozen Stable Diffusion 2.1 backbone on these encoded representations to generate the reconstructed image. Similarly to previous work, evaluation on four NSD and three THINGS-fMRI subjects includes reconstruction benchmarks, ablations, synthetic HRF-shift tests, and transfer-learning to new subjects analyses.

Critical strengths
  • A promising approach to image reconstruction from fMRI, rooted in physiological reasoning. Combining BOLD histories with inter-ROI interactions and synthetic-HRF consistency offers a plausible strategy for robust reconstruction. Moreover, most prior work used the temporal BOLD time series trivially, although the HRD is non-linear and highly varied across brain regions and subjects, thus the proposed approach touches upon a physiologically reasonable opportunity to harness neural information for reconstruction.
  • Encouraging robustness results across two datasets. Under two-second synthetic HRF shifts, full CHI improves SSIM over generic consistency by 0.028/0.023 and CLIP identification by 0.054/0.044 on NSD/THINGS-fMRI. Showing robustness across datasets suggests the method consists of promising rinciples, although further evaluations and validations are required.
Critical weaknesses
  1. The comparisons do not adequately establish improvement over existing methods. Brain-IT (Beliy et al. , ICLR 2026) is missing from both related work and evaluation, despite its directly relevant regional-interaction architecture and numerically better NSD results on all eight reported metrics. The MindHier comparison includes only one configuration, omitting another that outperforms DCG-CHI on CLIP identification and both feature-distance metrics. The DREAM results correspond to subject 1, whereas DCG-CHI’s results average four subjects; these should be corrected or clearly distinguished. Figure 4 also limits qualitative comparisons to NeuroPictor and BrainFLORA. Including stronger references, such as Brain-IT or MindEye 2 for NSD and MindHier for THINGS-fMRI, would better establish reconstruction quality.

  2. The experiments do not isolate the benefit of the two proposed mechanisms. Eq. 4 imposes stronger graph smoothing on older hidden states, but the paper does not explain why regional HRF variability supports this rule. To test it directly, the authors should compare against an otherwise identical model that uses the same learned smoothing strength at every lag. The existing graph-only, delay-only, parallel, and serial models change more than this relationship. Gains over the serial model are small—0.001–0.002 in SSIM/CLIP identification—and their reliability across independent training runs is not reported. The serial model should also be included in the timing-shift experiments. For CHI, the paper should establish whether generating observations through synthetic HRFs improves upon simply shifting input sequences in time and applying the same consistency objective. A clearly specified experiment removing the entire CHI pathway is also needed: all C0–C3 variants retain the synthetic-observation reconstruction and variance losses, while the separate “w/o CHI*” configuration is insufficiently described. Finally, the robustness test changes signal filtering as well as timing. Including inputs processed through the same filter without a timing shift would distinguish these effects. The larger robustness gains are promising, but these controls are needed to identify their source.

Questions for authors
  1. Decisive architectural evidence: What supports the imposed prior that graph smoothing should be increased with lag? Please compare Eq. 4 with an otherwise identical recurrence using a learned graph scale independent of lag, retaining the same temporal decay, delay weights, modes, and channel maps. Does the gain persist on clean and shifted inputs, including against the serial control? An additional ablation of learned lag-dependence of the smoothing would further establish this design choice.
  2. Comparative evidence: How does the method compare with contemporary methods with available results and code such as Brain-IT? Furthermore, identify the relevant MindHier configurations, and correct or separate DREAM's subject-1 result. Can the principal comparisons use common test stimuli and metric computation, with training subjects, extra data, repetitions, and generator differences clearly documented? Paired qualitative examples against stronger references would also make Figure 4 more informative (e.g. MindEye 2, Brain-IT).
  3. Decisive CHI evidence and removal clarification: Which parameters and loss terms are removed in “w/o CHI*,” and what are its eight standard reconstruction metrics? Since C0–C3 all retain the HRF cycle, can a fully specified complete-removal control and a direct temporal-shift consistency control establish the benefit of the chosen synthetic HRF construction? What evidence supports the shared pair's variation range?
  4. Interpretation clarification: What biological interpretation, if any, is intended for Figures 5F and 6? Are these should be compared to existing evidence of global HRF and regional HRF respectively? If Figure 6 is intended to validate physiological organization, what evidence separates learned structure from the supplied graph topology and supports correspondence with independent neuroscience measurements? If it is only a parameter diagnostic, that narrower scope should be explicit.
Reviewer_aB7wOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper at hand proposes DCG-CHI for fMRI-to-image reconstruction. The paper first argues that conventional decoders underuse the delayed, spatially distributed structure of BOLD signals. It then introduces DCG-SSM, which couples temporal lags with graph diffusion across brain regions. Next, CHI trains the encoder to produce consistent representations from two synthetic hemodynamic views. Separate semantic and structural readouts condition a frozen Stable Diffusion 2.1 model. The authors evaluate the method on NSD and THINGS-fMRI using controlled ablations, synthetic HRF shifts, regional analyses, and subject transfer. DCG-CHI obtains competitive reconstruction scores and generally outperforms its internal controls. Finally, the paper concludes that delay-coupled encoding and hemodynamic consistency improve reconstruction, especially under the tested temporal shifts.

Critical strengths
  • The delay-coupled graph encoder is a well-motivated contribution. It directly models the interaction between temporal lag and regional propagation rather than treating space and time independently. This gives the architecture a clear connection to the structure of BOLD measurements.

    • line 64: I liked that the authors identify lag-independent graph propagation as the precise modeling gap because it makes the architectural choice testable.
  • The internal controls isolate the two proposed components while keeping the generator fixed. The G variants compare alternative spatial-temporal encoders, while the C variants compare consistency objectives. This design makes the ablation results easier to interpret.

    • line 379: I liked that every model receives the same HRF perturbations because this avoids model-specific input confounds.
Critical weaknesses
  • CHI establishes consistency only under the model’s own synthetic HRF transformations. The training HRF family is shared across ROIs, and the test intervention primarily varies temporal shift. This does not establish invariance to empirical variation across regions, subjects, trials, or acquisition settings. The central claim should therefore be limited to the tested synthetic transformations.

    • line 238: Sharing one HRF family across all ROIs makes the invariance target narrower than the paper’s physiological framing suggests.
    • line 384: Improved performance under selected synthetic shifts does not establish resilience to naturally occurring HRF variation.
  • The improvements supporting the architectural claims are very small and lack statistical support. DCG-SSM exceeds the serial encoder by only 0.001 to 0.002 on key metrics, while full CHI exceeds generic consistency by 0.002 to 0.004. The paper does not report uncertainty across independent training runs, and its seed analysis holds the trained checkpoint fixed. It is therefore unclear whether these gains exceed optimization variability.

    • line 353: The small advantage over serial processing needs uncertainty across training runs before it can establish a benefit from coupling.
    • line 1130: A common generation seed controls sampling noise but not training or checkpoint-selection variability.
    • line 930: External methods retain different evaluation protocols, so Table 1 cannot support a direct ranking.
Questions for authors
  • How many independent training runs were conducted for the main ablations? Paired subject-level results and uncertainty across training seeds could materially change my assessment of the reported gains.
  • Does CHI generalize to independently estimated HRFs that vary across subjects or regions? This would clarify whether it learns physiological invariance or consistency only within its augmentation family.
  • Can the strongest baselines be evaluated under the same splits, preprocessing, and metrics? Otherwise, the comparative claim should rely only on the internal controls.

Comments on the paper

1 comment

Thank you for the additional experiments and clarifications. The graph sensitivity and held-out evaluations address several of my concerns, and I appreciate the more careful physiological interpretation. However, I remain unconvinced that the methodological advance over existing spatiotemporal graph models is substantial enough to support acceptance. I therefore maintain my rating of 2.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.