List of Papers Browse by Subject Areas Author List
Abstract
Accelerated acquisition of fMRI enables enhanced detection of neurovascular (BOLD) activity in the brain, but image reconstruction becomes challenging with high k-space undersampling: Task-evoked BOLD signals are small in magnitude, which traditional anatomical MRI reconstruction methods fail to recover, as they favor spatial accuracy over temporal fidelity. We present DD-INR, a \emph{D}ynamics-\emph{D}riven \emph{I}mplicit \emph{N}eural \emph{R}epresentation framework tailored for accelerated fMRI that benefits from incoherent time-varying sampling and a tailored spatio-temporal prior, outperforming traditional methods, demonstrated in simulation and in-vivo acquisition, both in terms of image quality and retrieval of activation patterns.
DD-INR achieves this by splitting the fMRI data into a static background and a temporally varying dynamic component, representing only the dynamics with a dedicated INR, thereby focusing the model’s capacity on activation-relevant changes while remaining compact. In general, DD-INR provides a promising framework for accelerated fMRI reconstruction, with the potential to improve the sensitivity and robustness of fMRI studies within practical scan time limits.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2792_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/JoosenLi/DD-INR
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiQia_DDINR_MICCAI2026,
author = { Li, Qiaoxin AND Pan, Caini AND Comby, Pierre-Antoine AND Giliyar Radhakrishna, Chaithya AND Ciuciu, Philippe},
title = { { DD-INR: Dynamics-Driven Implicit Neural Representation for Accelerated Whole-Brain Functional MRI Reconstruction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes an INR-based reconstruction method for accelerated task fMRI. The signal is decomposed into a static background, estimated from temporally aggregated measurements, and a dynamic component represented by an INR with spatial or spatiotemporal regularization. The goal is to focus the reconstruction on the weak task-related BOLD fluctuations rather than the time-invariant anatomical background.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper addresses an important problem in accelerated fMRI. The basic modeling choice is appropriate for this setting: the dominant signal is largely static, whereas the quantity of interest is a much smaller dynamic component. Estimating the background separately and reserving the INR for the dynamic residual is therefore a reasonable design. The method is organized clearly at a high level, and the regularization imposed on the dynamic term is appropriate for task fMRI. The simulation results are encouraging, and within the comparison set included in the paper, the proposed method improves the reported activation maps and time-series metrics.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper gives the impression of introducing a new reconstruction framework for accelerated fMRI, but the underlying ideas are closer to existing work than the manuscript acknowledges. The separation of slowly varying background structure from a smaller dynamic component is already established in accelerated fMRI and dynamic MRI reconstruction. Petrov et al. used a low-rank plus sparse model for accelerated fMRI, including prospective experiments, and Chiew et al. incorporated task-related temporal information directly into highly accelerated fMRI reconstruction through a constrained low-rank model. More broadly, INR-based reconstruction for dynamic MRI is already an active direction. I would therefore encourage the authors to position the paper more narrowly, as an INR-based implementation of a background-plus-dynamic strategy for accelerated fMRI.
Several key quantities are introduced with less precision than one would expect in a methodological paper. The notation around (f_\theta) is not defined carefully enough when first introduced, and the construction of the background estimate (\hat{x}{bg}) through (A{1:T}^{\dagger}y_{1:T}) remains largely intuitive. The idea itself is acceptable, but the operator, the temporal aggregation step, and the underlying assumptions should be stated more explicitly.
The numerical results are favorable, but the qualitative evidence is more limited than the wording suggests. In Fig. 3, the recovered curves are improved relative to the noisier baselines, but they still do not follow the expected response especially closely in an absolute sense. I would therefore encourage the authors to moderate the stronger claims about the accurate recovery of amplitude and phase.
The paper would benefit from a more direct discussion of downstream fMRI inference. In this area, image quality alone is not the decisive criterion. Prior work has already noted that nonlinear reconstruction can affect temporal correlations, effective degrees of freedom, and statistical sensitivity. That point deserves more careful treatment here.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
I think the paper is built around a relevant idea, and I agree with the decision to distinguish the dominant background from the much weaker dynamic signal. My main recommendation is to sharpen how the work is presented. In my view, the contribution is best described as an INR-based implementation of a background-plus-dynamic reconstruction strategy for accelerated fMRI, rather than as a fundamentally new reconstruction framework. The paper would be stronger with a more precise formulation of the method, especially the notation around (f_\theta), the definition of (\hat{x}_{bg}), and the optimization problem being solved. I would also encourage a more direct discussion of the closest fMRI-specific prior work and a more careful treatment of how the reconstruction may affect downstream statistical analysis.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a worthwhile problem and is based on a coherent modeling idea. However, I do not think the current version makes a sufficiently strong case for acceptance as a methodological contribution. The main reason is that the conceptual ingredients of the method are closer to existing fMRI and dynamic MRI reconstruction work than the manuscript presently indicates. In addition, the formulation needs to be stated with greater precision, some of the claims are stronger than the qualitative evidence supports, and the in vivo validation is limited.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper introduces an INR-based framework for fMRI reconstruction from accelerated acquisition. There are two main contributions. First, the paper introduces an additive model for fMRI data, which consists of the static background component plus a temporally varying dynamic component. By explicitly modeling the background as static, i.e., constant across dynamics, INR only need to focus on modeling the temporally varying component. Second, two regularization methods are proposed. The authors demonstrate that the two regularization methods results in improved agreement of the ROI-averaged time series with ground truth despite the fact that little or no temporal model/prior is imposed.
little model/prior is imposed in the temporal domain
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Heuristically, modeling fMRI data as a static background component plus a temporally varying dynamic component makes a lot of sense, similar to residual learning, or the low-rank plus sparse (L+S) reconstruction approach. The authors argue that such a decomposition allows the INR to focus its capacity on modeling the dynamic component as opposed to repeatedly learning the static background component. 2.The authors provide a strong justification for their choice to do Fourier feature embedding separately in space and time. Specifically, their justification for this is two-fold, i.e., reducing number of total parameters and the ability to allocate different bandwidths for spatial versus temporal features. 3.The authors perform extensive evaluation on both simulated data where a ground truth is available, and on data acquired in vivo. This includes quantitative (Table 1) and qualitative (Figs. 2 and 3) evaluations against ground truth when using the simulated data, and also qualitative evaluation when using in vivo data. 4.Finally, it is non-trivial that the proposed regularization works well. The proposed regularization does NOT explicitly assume a hand-crafted model for the temporal signals across dynamics. It only encourages neighboring voxels to have similar temporal patterns.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper is already very clear and well-written although additional details can be incorporated during revision to further improve the clarity and highlight the contribution of this work. 1.For readers not familiar with INR, it might be unclear (until later in the paper) that the method does not require training on a carefully curated dataset but instead works on a single subject. Consider clarifying or maybe briefly highlighting this in the abstract and introduction so that the readers can better appreciate this work. 2.At the bottom of page 3, it is unclear how \hat{x}{bg} is estimated. Later in the paper, the authors explain that \hat{x}{bg} is obtained from temporally aggregated measurement. Is the (k, t)-space data temporally aggregated first into 3D k-space without the temporal dimension? Or is it the case that a least-squares problem is solved assuming the underlying image to be reconstructed is temporally static? 3.At the top of page 4, the authors presents two regularizers in (3) but only introduces the latter one, i.e., “spatially regularize the temporal derivative”. It might be helpful to at least briefly describe the former one as well. 4.Why does the image contrast differ across different methods in Fig. 2? Is it because of scaling during visualization? 5.In Table 1, it might be helpful to also compute and report tSNR. It helps to make the claim that the proposed method improves temporal stability. 6.In Fig. 4, the reconstructed image for PnP looks over regularized in the spatial domain. Is the regularization parameter for PnP properly tuned? 7.In the discussion, the authors attempt to explain the different behaviors of spatiotemporal vs spatial regularizations. However, the discussion does NOT explain why R_{xyz} results in more temporal smoothing whereas R_{xyz, t} results in less temporal smoothing in Fig. 3.It might be helpful to include more in-depth discussion about this behavior in the discussion section.
There is one technical question though. When using INR for 3D+t MRI reconstruction, it is reasonable (and in fact common) to assume that the underlying 3D+t data satisfies certain regularity condition in both space and time, e.g., spatiotemporally continuous. In the current formulation (static background + dynamic), does the regularity condition still holds after removing the static background. There is no issue with regularity in the temporal domain because temporally a constant is removed and adding/subtracting a constant doesn’t change the regularity of the underlying function. However, there might be an issue in the spatial domain. Say the static background reconstruction is not perfect and there are artifact in the estimated static background \hat{x}{bg} which introduces spatial discontinuities. In this case, even if it is assumed that the 3D+t fMRI data is continuous, the dynamic component to be modeled using INR is NOT spatially continuous because of the spatial discontinuity in \hat{x}{bg}.
One small limitation of the work is that joint reconstruction is performed across all dynamics, making it hard to apply this method for real time fMRI applications.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
There are some typos and aesthetic issues that should be addressed during revision. 1.At the bottom of page 3, there is a \hat missing from \hat{x}_{bg}. 2.At the bottom of page 5, there is a subscript missing (F_2 instead of F2). 3.Fig. 2.The color bar is hard to see. The black background of the text box is aesthetically unpleasing. I am also have trouble finding where the blue arrows are.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method is sound and the evaluation is extensive with both simulation and in vivo experiments. The contribution of decomposing fMRI data into a static background component plus a temporally varying dynamic component makes sense and is novel. The proposed regularization methods are also sound. A score of 5 is given instead of 6 because both contributions are quite standard.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
DD-INR is an INR framework for accelerated 3D+time fMRI reconstruction. The key design is a signal factorization into a static background (reconstructed via conjugate gradient from temporally aggregated k-space) and a dynamic component modeled by a SIREN with decoupled spatial/temporal Fourier feature embeddings. The dynamic component is regularized using TV on the spatial gradient of the temporal derivative, computed via autodifferentiation. Evaluated on simulated (SNAKE) and in vivo (single subject, visual block design) data against NUFFT, CG, CS, PnP, and a generic INR baseline.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The problem framing is sharp and well-motivated: BOLD signals are 2–5% intensity fluctuations superimposed on a strong static background, and dedicating INR capacity exclusively to the dynamic residual is a principled design choice. The use of autodifferentiation to compute the temporal derivative for regularization is elegant and technically clean. The comparison against PnP-fMRI (a strong domain-specific baseline) strengthens the evaluation. Both simulation with ground-truth activation maps and in vivo results are provided, with appropriate fMRI-specific metrics (F2-score, GLM z-score maps, BOLD time-series correlation). The performance gap over baselines is large and consistent — Corr of 0.909 vs. 0.84 for PnP and 0.40 for generic INR is a meaningful improvement.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Single subject in vivo evaluation. The in vivo experiment involves one healthy volunteer with a single task paradigm (visual block design). This is insufficient to assess generalizability across subjects, tasks, motion levels, or scanner variability. The simulation compensates partially but uses a simplified phantom (white/gray matter/CSF only), which may not capture the full complexity of real fMRI noise. 2.Static background assumption is fragile. The background is estimated once from temporally aggregated measurements and frozen throughout reconstruction. This is violated by head motion, physiological noise (cardiac/respiratory fluctuations), and slow scanner drifts — all common in fMRI. The discussion acknowledges motion but dismisses it without quantitative analysis. No experiment tests robustness under varying motion levels. 3.Reconstruction time is prohibitive for practical use. ~2 hours per scan on an RTX 6000 GPU is noted as “comparable to traditional methods” — but this is not accurate for clinical or large-scale research use. Standard CS fMRI reconstruction is typically minutes, not hours. This severely limits the practical appeal of the method, and no analysis of the computational bottleneck or potential speedups is provided beyond attributing cost to NUFFT. 4.Hyperparameter sensitivity is uncharacterized. The regularization weights (λ = 10⁻² and 2×10⁻³), warm-up schedule, Fourier embedding dimensions (Bxyz ∈ R²⁵⁶ˣ³, Bt ∈ R⁶⁴), and network width (512) are all fixed without ablation. Given the sensitivity of SIREN-based methods to initialization and frequency parameters, and the critical role of λ in balancing data fidelity vs. regularization for weak BOLD signals, this is a significant gap. 5.No resting-state fMRI evaluation. The method is evaluated only on task-based fMRI with a simple block design — the easiest possible activation pattern to recover. Resting-state fMRI (functional connectivity, default mode network) involves weaker, less structured temporal fluctuations and represents the majority of clinical and research fMRI use. No discussion of applicability to this setting is provided.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The problem formulation is well-motivated and technically sound. The performance gains over all baselines — including a domain-specific PnP method — are substantial and consistent across simulation and in vivo settings. The core idea of dynamics-centric INR factorization is a genuine contribution to accelerated fMRI reconstruction. The main weaknesses are limited in vivo evaluation scope, uncharacterized hyperparameter sensitivity, and the 2-hour reconstruction time which limits practical utility. These are addressable in a revision but would benefit from author response on the robustness and scalability points.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The reviewers see the paper as a well-motivated idea: separating static background from weak dynamic BOLD signals lets the INR focus on what matters, and this yields strong, consistent gains with convincing fMRI-specific evaluation.
There were concerns about validation (single subject, simple task), assumptions like static background may break under motion, reconstruction is slow, and downstream statistical reliability and robustness remain underexplored.
The paper would raise good discussion in MICCAI. The authors should discuss and address them in the final version.