List of Papers Browse by Subject Areas Author List
Abstract
Fluorescence molecular tomography (FMT) is a non-invasive optical imaging modality for three-dimensional molecular visualization, yet its reconstruction is fundamentally limited by the severe ill-posedness of the inverse problem, resulting in poor spatial resolution and localization instability. While diffusion models improve detail recovery, their dependence on small-step Gaussian approximations necessitates long stochastic denoising processes, leading to high computational cost and reconstruction instability under few-step sampling. We present TCDFlow-FMT, a fast deterministic flow-matching framework that reformulates FMT reconstruction as continuous-time transport between noise and fluorescent sources distribution. By replacing stochastic denoising with ODE-based deterministic flow, our approach achieves reproducible and efficient reconstruction in only a few steps. We introduce a spatially guided flow Transformer that integrates multi-view fluorescence measurements into velocity field prediction via structure alignment and spatial modulation, resolving the global–local mismatch between diffuse measurements and sparse sources. To accelerate inference, we propose a Gibbs-based geometric consistency reweighting strategy that implicitly approximates optimal transport and straightens noisy coupling paths into stable deterministic trajectories. Phantom and in vivo experiments demonstrate that TCDFlow-FMT achieves efficient few-step reconstruction with improved spatial localization and quantitative accuracy. Code will be available at: https://github.com/Qianqian-Xue/TCDFlow-FMT.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0784_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Qianqian-Xue/TCDFlow-FMT
Link to the Dataset(s)
N/A
BibTex
@InProceedings{XueQia_Beyond_MICCAI2026,
author = { Xue, Qianqian AND Liu, Xingyu AND Shan, Hongming AND Zhang, Peng AND Wang, Wenjian},
title = { { Beyond Stochastic Diffusion: Trajectory-Consistent Deterministic Flow Matching for Fluorescence Molecular Tomography } },
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 authors present a flow matching based approach to compute to solve the inverse problem in Fluorescent Molecular Tomography (FMT). They use a transformer based architecture to compute the velocity vector at each point of the flow process. Main novelties are the architecture, and the application. The method is applied to recover the shape of real and simulated glass cylinders in a mouse and a phantom.
- 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.
-
I think it makes sense to use such a generative model for an inverse problem like this.
-
The paper includes an ablation study.
-
The paper is generally well written.
-
- 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.
-
One of the main selling points is the superiority with respect to inference speed compared to a diffusion model. However, the required computation time is not actually measured and compared to the diffusion model baseline.
-
The authors make multiple statements like “method resolves the . . . generative uncertainty of previous models”. I don’t think this is a fair statement for two reasons. (i) The variance in the prediction of a generative model can be a good thing, as it is a way to account for the inherent uncertainty of the model. (ii) The flow based model is a generative model as well, right? Depending on the random sample chosen as a starting point, we should get non deterministic outputs, right? It would actually be interesting to see the amount of variantion (captured uncertainty) the model produces.
-
I think the choice of training and evaluation data is a bit limited. Would it make sense to use other shapes, but just cylinders?
-
The authors don’t mention publishing the code or data.
Monor Points:
-
The relevant papers on flow matching should really be cited in the introduction.
-
In Section 2: “. . . the fluorescent sources distribution p_data(x). . . “ Should this not be p_data(x|Phi)?
-
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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?
I think this is a decent paper, but has some substantial flaws in evaluation and theoretical framing, regarding the role of generative models.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
My concerns have been addressed and I am happy for the paper to be published. However, there is one point where I would like to insist the authors reconsider how the formulate the motivation for their method. In their rebuttal the authors say that one feature is that there is no noise injected during inference (unlike in diffusion models). I would like to insist that this is really a feature of a generative model. The in this way, that the diffusion model can describe the aleotoric uncertainty of the problem. For every input there is no unique target output but multiple outputs are possible. The fact that the model produces probabilistic outputs (by adding noise during inference) is thus a feature, I believe.
Review #2
- Please describe the contribution of the paper
This paper proposes TCDFlow-FMT, a deterministic flow-matching framework for fluorescence molecular tomography (FMT) reconstruction. To tackle the challenges of non-unique and unstable solutions for sparse fluorophores, this paper incorporates two case-specific designs into the standard flow matching pipeline. The spatially guided flow Transformer (SGFT) conditions velocity field estimation using extracted structural priors from multi-channel fluorescence measurements. Geometric consistency reweighting (GCR) adapts weighted conditional FM with custom objective terms to enforce near-straight deterministic transport trajectories. The model is trained with synthetic data and shows improved reconstruction and localization in a controlled phantom experiment and in one mouse model.
- 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.The paper adapts flow matching with reasonable and effective designs to improve efficiency and stability, specifically for FMT reconstruction. 2.In SGFT, the proposed dual conditioning mechanism is intuitive for handling the mismatch between diffuse measurements and sparse fluorophore distributions, which also does not introduce heavy computation overhead compared to standard flow matching conditioning. 3.The use of core and guide kernels for the multi-scale potential fields in GCR is well-motivated for fast convergence and reconstruction precision of the sparse fluorescent signals. 4.The tailored energy function in GCR, combining Dice-based alignment and sparsity regularization, is appropriate for the reconstruction objective and reflects domain-specific optimization. 5.The method demonstrates consistent improvements in proposed evaluations of controlled settings compared to classical and learning-based baselines. 6.The ablation study covers both key designs in SGFT (sequence and vector features) and GCR, which shows a good balance between efficiency and reconstruction accuracy.
- 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.The writing of the method part tends to entangle existing methods with newly proposed task-specific designs, making it hard to clearly identify the novelty of the proposed work. For example, adaptive layer normalization is a common operation in diffusion models. In GCR, latent feature alignment and noise scaling are also common tricks for diffusion models. If the designs are inspired by existing work and tailored for specific applications, the original works should be properly cited. 2.The key limitation is the evaluation data size. The physical phantom experiment only has one phantom with 2 fluorescent tubes at 3 separations. And the in vivo experiment only includes one mouse with one glass tube. The small evaluation size makes quantitative results unconvincing. 3.The unrealistic fluorescence distribution (cylinder glass tube) does not reflect real biological scenarios. The experiment does not evaluate performance under different fluorophore configurations, such as irregular shapes and changes in optical properties.
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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.
(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?
This paper presents a reasonable and well-grounded adaptation of flow matching for FMT reconstruction, with task-specific designs such as SGFT and GCR. The use of deterministic flow matching to improve efficiency and stability over diffusion-based approaches is meaningful. The introduction of model design is easy to follow and well-motivated. The ablation study also provides insight into the role of key module designs. However, the extremely small evaluation dataset size and unrealistic phantom fluorescent signal configurations are the major concerns that prevent the paper from receiving a positive rating. Such a limited (one phantom and one mouse) and highly simplified setup (cylindrical glass tube) does not reflect realistic biological complexity and is insufficient to demonstrate robustness or generalization. Although this paper proposes a promising method, a more rigorous and realistic evaluation is necessary to provide sufficient support to meet the MICCAI standard.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The rebuttal did not solve the concerns about the incremental novelty and extremely limited experiments. The results from only a few synthetic volumes and one in vivo sample are not convincing.
Review #3
- Please describe the contribution of the paper
They constructed the first deterministic flow-matching framework, which can stably reconstruct fluorescence molecular tomography.
- 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.
By using Gibbs-energy-based weights in the objective function, the method can learn stable and shorter trajectories.
- 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 performance improvement is limited compared with some SOTA methods.
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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
The paper is well-written, but the citation number is different from the order in the paper. It is better to align both of them.
- 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?
To resolve unstable reconstruction and/or inefficient computation, they constructed the framework, which consists of a structure alignment encoder and a spatial modulation decoder, and the performance was verified by the experiments.
- 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
We thank the reviewers for recognizing our novelty (R1, R3), SGFT/GCR designs (R2, R3), and reproducibility (R1, R2, R3), and address the main questions(Q) below. (1) <R1 Q1: Inference speed and computation time> To clarify, the inference speed advantage essentially refers to a significant reduction in NFE (core metric for generative inference speed). As stated in the paper, diffusion baseline is constrained by its mechanism to require “thousands of denoising steps”, whereas TCDFlow-FMT achieves efficient reconstruction in “only a few steps” (25). According to actual times, under identical environments, TCDFlow-FMT is approximately 13.36s faster than diffusion. (2) <R1 Q2: Generative uncertainty> Addressing your concern, we clarify that “resolves generative uncertainty” refers not to diversity from initial samples, but to uncertainty from diffusion models injecting noise z at denoising step. For FMT, such uncontrolled intra-step stochasticity undermines reproducibility. In contrast, TCDFlow-FMT’s deterministic ODE introduces no noise once the initial sample is fixed, ensuring reproducible reconstruction. We will strictly refine the relevant details and investigate the uncertainty arising from random initial samples in future work. (3) <R2 Q1: Clarification of task-specific novelty and citations> We thank R2 for the comments and will add relevant citations, revising the corresponding details. To delineate existing research from our contributions, the designs for FMT corresponding to the specific parts raised by R2 are explained as follows. First, the original AdaLN (DiT, Peebles & Xie, 2023) performs globally uniform modulation for low-dimensional conditions, failing to exploit spatial information in fluorescence signals. Guided by zseq, our spatial AdaLN extends modulation into a patch-wise spatial form, assigning independent weights to different patches for precise localization. Second, while our x0 treatment is inspired by EDM (Karras et al., 2022), it differs fundamentally. EDM uses linear scaling for numerical stability during training, whereas we achieve geometric alignment between noise and the target potential field via dynamic non-linear projections. (4) <R1 Q3 & R2 Q2&Q3: Evaluation data size, fluorophore shapes, and realistic scenarios> We thank the reviewers for the comments. While current sample sizes are limited, two-target phantom and single-target in vivo setups are conventional FMT protocols for verifying algorithmic feasibility. Moreover, implanting fluorescent tubes in living mice is a widely adopted in vivo validation method that ensures high experimental controllability and provides precise ground truth to effectively evaluate algorithmic performance (Cai et al., IEEE TCI, 2026; Chen et al., Opt. Express, 2023; Zhang et al., IEEE TCI, 2025). Importantly, the core contribution of our work lies in overcoming the uncertainty of diffusion model via a fast deterministic reconstruction strategy, thereby improving both image reconstruction efficiency and accuracy. Our ablation, phantom, and in vivo experimental results consistently validate this conclusion. We also agree that larger sample sizes and complex, irregular fluorophore configurations are crucial for evaluating model generalization, and we will explore these in our future work. (5) <R3 Q1: performance improvement> FMT reconstruction is a highly ill-posed inverse problem, and existing methods already demonstrate strong competitiveness; thus, the performance improvements over certain SOTA methods in some scenarios are limited. Our main contribution lies not only in improving quality, but more importantly, in proposing a fast deterministic reconstruction method to effectively mitigate the uncertainty in diffusion models and improve reconstruction efficiency. (6) <R1&R3: Code and citation> We thank the reviewers for pointing out citation and language issues, which will be addressed. Code will be publicly available upon acceptance.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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.
Please provide justifications and clarifications for the concerns raised by the reviewers.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
This paper proposes TCDFlow-FMT, adapting flow matching for FMT reconstruction via two modules, SGFT and GCR.
The reviewers broadly agree that the paper is clearly written and the proposed modules are well-motivated. R1 raises concerns about the evaluation of inference speed and the framing around generative uncertainty. R2 raises the more fundamental concern that the task-specific designs are not clearly differentiated from existing techniques, and that the evaluation dataset is too small and unrealistic to be convincing. R3 is largely positive but notes limited performance improvement over some baselines.
After considering the reviews and the rebuttal, SGFT and GCR read more as engineering tricks than as designs derived from an identified bottleneck specific to FMT. This is not fatal, but it shifts the primary contribution to the evaluation and demonstration of flow matching on this application, a bar that the current experiments do not meet. One phantom with cylindrical glass tubes and one mouse is insufficient to demonstrate robustness or generalization to realistic biological scenarios.
Thus, this paper is not recommended for acceptance in its current form.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The post-rebuttal reviews remain largely consistent with the initial reviews. Some reviewers acknowledge the proposed framework for addressing unstable reconstruction and computational inefficiency, while others continue to raise concerns about limited novelty, insufficient experimental validation, and the positioning of the method. However, the majority of reviewers view the paper positively. Considering the overall discussion, I recommend acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
I read the paper myself and think it deserves green light. Indeed, the methodology of flow matching is somewhat overrated (not for fluorescent microscopy, though); but the use of OT and the use of X-ray as the ground truth appealed to me: the authors did everything they could for this underrepresented imaging modality, including phantom and in-vivo shots.
I do want to stress that the authors should soften some claims w.r.t. “significant superiority” of their method, given such a small and limited data that nobody else will be able to use to repeat the experiments.
Also, definitely amend the description of inference routine, following the post-rebuttal comment by R1.
Still, rare and interesting work – should be a part of MICCAI.