Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Reconstructing 3D Computed Tomography (CT) images from a few X-ray projections is a highly ill-posed inverse problem due to the loss of volumetric information. We propose PhyDiCT a training-free framework that integrates a differentiable Physics-based forward model, grounded in the Beer–Lambert law, with a text-conditioned Diffusion as a strong prior to reconstruct 3D lung CT images. We refer to our approach as training-free since the prior model is used without fine-tuning, and our goal is to steer the denoising procedure to generate samples consistent with X-ray observations. We guide the diffusion generation using Split Gibbs sampling to jointly optimize for projection fidelity (reward) and consistency with prior knowledge. Also, we introduce a test-time refinement step that enhances image realism and anatomical coherence. We extensively evaluate our method on publicly available 3D CT datasets using both perceptual and semantic metrics, demonstrating that it surpasses existing plug-and-play diffusion and fully trained reconstruction approaches. Our findings highlight that combining a strong generative prior with the underlying physics of image formation substantially improves reconstruction quality, e.g., 7.5% improvement on SSIM compared to full training methods. Code will be released at https://github.com/batmanlab/PhyDiCT.git.



Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1942_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/batmanlab/PhyDiCT

Link to the Dataset(s)

N/A

BibTex

@InProceedings{DaiWei_PhyDiCT_MICCAI2026,
        author = { Dai, Weicheng AND Ghosh, Shantanu AND Batmanghelich, Kayhan},
        title = { { PhyDiCT: Plug-and-Play CT Reconstruction from Sparse X-Rays via Differentiable Rendering and Strong Priors } },
        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 a guided DDIM approach for CT reconstruction based on the text-guided generative prior MedSyn, a model trained to generate images using additional contextual information such as reports from X-ray images. In addition, MedSyn operates directly on fully 3D volumes without requiring a pre-trained autoencoder.

    The data-guidance term comprises a standard least-squares term, corresponding to the negative log-likelihood under a Gaussian noise model, together with an LPIPS term that evaluates a perceptual discrepancy between the expected projection and the measurement. An additional “Extra Compute” refinement step is also introduced, consisting of re-noising the image and applying the denoiser again at the corresponding noise schedule.

    The proposed framework is evaluated on simulated data using two physical models: one close to the forward model used for reconstruction, and one more realistic. The results are compared with those of other reconstruction approaches.

  • 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 idea of image reconstruction with a text-guided diffusion prior is interesting and potentially relevant for radiological applications. In particular, incorporating contextual information through text could be useful in settings where side information is available and clinically meaningful.

    That said, MedSyn, on which the proposed reconstruction framework is based, dates from 2024.Therefore, I would not be surprised if similar ideas have already been implemented and published, and the paper should clarify more explicitly what is genuinely new here relative to prior work.

  • 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 framework uses a projection-space LPIPS term in the data fidelity term J. However, LPIPS(x,y) is, to my knowledge, a bounded function of x, and therefore is not coercive, so that \int exp(-J(y,x)) dx = +infty. Therefore, it is mathematically incorrect to formulate the proposed framework as a sampler from a posterior of the form p(x)exp(-J(y,x)). The inclusion of LPIPS should instead be presented as an additional guidance term incorporated into the diffusion sampler.

    • Using LPIPS to guide the reconstruction in a DPS framework seems new to me. However, its usefulness is difficult to assess without the corresponding ablation study, i. e. , reconstruction without LPIPS in J.

    • It would be useful to include an example where the context c carries meaningful information for the reconstruction, for example c = “the patient has a lung tumour”, and then compare reconstructions with and without c.

    • To me, the proximal step (cf. SGS part) used to update z by minimising l(z) = J + |z-x0|2 does not appear to be new; see, for example, Chung et al. (2023) and Zhu et al. (2023).

    • The EC step also does not seem new to me, as many similar ideas based on Langevin dynamics have already been proposed, e. g. Zhang et al.

    • Please check these references and cite them if needed.

    • To me, this work is about CBCT (cf. Figure 1(b)), and the paper should state this more clearly.

    • The title is misleading: (1) to me the method is DPS and not plug and play and (2) it does not mention the text guidance.

    Minor:

    • Beginning of Section 2: I believe the posterior should be written as p(x|Y,c) \propto p(x|c) exp(-J), i. e. , there is a c missing.

    • I think it would be less confusing if (3) used notation similar to that in (4), i. e. I(\mu)[r]. Also, I believe r should be replaced by p.

    • Still in (3) and (4), the source location s is missing on the left-hand side. Consider the following notation: I_s(\mu)[p] = . . . . and P_s(x)[p] = . . . . This notation would also be compatible with (5) by replacing s with \omega_k.

    • Below (4), the lambda_m are not “intersection points” but the coordinates of the intersection points along the ray.

    • I believe (4) should be dropped, as it is well known in the literature that (3) is discretised for reconstruction purposes.

    • Page 5: the paper states x_t-1 = DDIM(z_t,x_t). Please define in (2) the DDIM update as a function of (x0)_t and x_t so that DDIM(z_t,x_t) can be understood.

    • The EC step may introduce hallucinations, although they could be mitigated with the context c. This could be discussed in the paper.

    H. Chung, S. Lee, and J. C. Ye, “Decomposed diffusion sampler for accelerating large-scale inverse problems,” arXiv preprint arXiv:2303.05754, 2023.

    Y. Zhu, K. Zhang, J. Liang, J. Cao, B. Wen, R. Timofte, and L. Van Gool, “Denoising diffusion models for plug-and-play image restoration,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 1219–1229.

    B. Zhang et al. , “Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing”, 2025

  • 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

    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?

    • The paper proposes some interesting ideas, in particular the use of text-guided diffusion models for reconstruction, although it is difficult to assess their practical impact from the presented results.

    • A major issue to me is that the paper builds part of its methodological narrative around LPIPS being treated as part of a negative log-posterior term, which appears mathematically incorrect (see above). This component should instead be presented more carefully as a penalty or guidance term.

    • The effect of LPIPS should be investigated through an ablation study. Several parts of the framework may not be new (above comments) and the paper should be positioned more carefully with respect to the relevant prior literature.

  • Reviewer confidence

    Very confident (4)

  • [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.

    It turned out that my initial comment regarding the coercivity of LPIPS was not relevant. Therefore I no longer have concerns about the mathematical formulation of the proposed framework.



Review #2

  • Please describe the contribution of the paper

    This paper proposes PhyDiCT, a plug-and-play framework for reconstructing 3D lung CT volumes from sparse 2D X-ray projections. The method combines three key components: (1) a physics-based differentiable forward model grounded in the Beer–Lambert law to simulate X-ray projections from candidate CT volumes, (2) a pre-trained text-conditioned 3D diffusion model as a strong anatomical prior, and (3) an iterative guidance strategy based on Split Gibbs sampling to balance projection fidelity with prior realism during inference. In addition, the paper introduces a test-time refinement step (EC) to further improve image realism and anatomical consistency. Experiments on public CT datasets with synthetically generated X-rays show improvements over both fully trained reconstruction baselines and prior plug-and-play diffusion methods.

  • 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.

    Overall, I found the paper clear, technically interesting, and generally well organized. The motivation is strong: sparse-view CT reconstruction is an important and highly ill-posed inverse problem, and the paper addresses it using a modern combination of generative priors and physics constraints. The presentation is easy to follow. The method section also uses a reasonable mathematical framework. Formulating the reconstruction problem as posterior inference with a prior term and an observation-consistency term is principled, and grounding the reward function in a differentiable Beer–Lambert forward model is stronger than relying purely on heuristic image-space losses. On the empirical side, the paper evaluates against multiple baselines, includes both fully trained and plug-and-play methods, tests across two synthetic projection pipelines (DiffVox and DeepDRR), and provides ablations on the number of views and the EC refinement step. This gives the experimental section reasonable breadth and shows care in benchmarking.

  • 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.

    My main concerns are as follows.

    1.Novelty. While the overall system is well engineered, the technical novelty appears somewhat diffuse. Many of the core ingredients are individually established: differentiable rendering for tomography, diffusion priors for inverse problems, text-conditioned generative models, variable-splitting / Gibbs-style iterative optimization, and post-hoc refinement steps. The contribution seems to lie primarily in combining these components effectively for sparse-view lung CT reconstruction, rather than introducing a fundamentally new reconstruction algorithm or inference method. This can still be valuable, especially in an applied medical imaging setting, but the paper would benefit from better formulating the key contributions and make the novelty claim more convincing.

    2.The evaluation setup is essentially CT → synthetic X-ray → reconstruct CT. While understandable due to limited public paired datasets, this closed-loop synthetic benchmark may be much easier than real clinical settings, where scanner noise, scatter, calibration mismatch, and domain shift are present. Therefore, the reported gains may not fully reflect real-world performance.

    3.Question: In section 2, it mentions “Instead of performing a full and costly sampling procedure, we take a few gradient steps (two in all experiments) on ℓ(z) to obtain an optimized z_(t−1).” I would be interested in understanding how using only two gradient steps is justified in practice, and whether there are any guarantees or empirical evidence regarding convergence or stability of this update procedure？

  • 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

    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?

    This paper addresses an important sparse-view CT reconstruction problem with a well-motivated framework that combines physics-based differentiable rendering and a strong pretrained diffusion prior. The method is clearly presented, and experiments are reasonably comprehensive with multiple baselines and ablations, showing promising results. However, the novelty is moderate since much of the contribution comes from integrating existing components rather than introducing a fundamentally new method. In addition, parts of the optimization procedure appear heuristic, and the evaluation relies on a synthetic pipeline (CT → simulated X-ray → reconstruct CT), which limits confidence in real-world applicability.

  • 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

    The paper provides a physics grounded test-time method of adapting a prior diffusion model for sparse view Xrays to CT reconstructions. In addition they also provide a novel reward model relying on physical process of CT volume construction that is used in this process.

  • 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 motivation of the paper is clear, sparse Xray to CT is not fully solved problem and is of definite interest to this community.
    • methodological contributions is solid - formulation of the reward model using Beer–Lambert law is very intuitive and solid.
    • The evaluations and ablations cover a lot of ground and the results show a definite advantage of the method over other SOTA.
    • The results are discussed well.
  • 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.

    Just minor comments

    • Missing a limitations section, it is worth discussing the trade off between the lower CLIP T2I score and in what scenarios full training methods would be better.
    • The explanation of the EC is not very clear, the writing should be improved on how does proposed method improve manifold alignment.
  • 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

    -

  • 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?

    Very grounded motivation and the proposed methods align with the motivation. The methods provided are novel for the MICCAI conference and will be of interest to the readers. Finally the results are a step improvement over sota and the analysis of the parameters and discussion of the results are more or less complete outside of minor nits.

  • 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 all reviewers (R1, R2, R3) for their informative feedback and valuable suggestions. We are encouraged by the positive comments acknowledging our contributions in CT reconstruction with a text-guided prior and a physics-based reward function (R1, R2, R3), as well as the solid experimental demonstrations and result discussions (R2, R3). We address the key concerns below:

1) R1 & R2–Out-of-domain (noised) observation & LPIPS incorporation Thank R1 for pointing LPIPS is not coercive, you are correct. We found that observation methods such as DeepDRR produce realistic noise patterns and beam hardening effect, which are not differentiable. Therefore, the discrepancy introduces domain shifts with respect to our rendering method in Eq.4.Thus, LPIPS term is used as an additional regularization of guidance for perceptual consistency to address out-of-domain observation shifts, where the diffusion trajectory is guided partially by the perceptual gradient. We will add the ablation study without LPIPS term, and revise LPIPS incorporation to make the distinction clearer.

2) R1–Novelty of SGS and EC & R2–Novelty Claims & R3–Explanation of EC We do not claim SGS, nor SDEdit, as our contribution. Instead, we claim the following contributions:1) Incorporating a physical 3D-to-2D projection into the reward function with both MSE and LPIPS regularization (Tab. 1, two rendering settings); 2) Training-free lung 3D CT reconstruction from sparse X-ray (from reward function) and a text-guided prior; 3) A post-hoc EC step to remove artifacts (Fig. 2 last 2 columns, Tab. 1 “w/o EC”). We acknowledge the high-level resemblance between our EC step and DAPS (Zhang et al.), in that both involve adding noise. However, the motivation and mechanism differ substantially. 2.1) Motivation: DAPS aims to solve complicated nonlinear inverse problems through posterior exploration and Langevin-style sampling. However, our EC goal is artifact suppression and manifold alignment after posterior guidance. 2.2) Mechanism: DAPS introduces Decoupled Noise Annealing at each timestep together with posterior guidance. However, our EC is a post-hoc artifact-correction stage, where we add noise through the DDPM forward process and apply denoising steps only (Eqs. 1–2), without reward guidance, after the inverse problem has been solved. The goal is to produce visually cleaner reconstructions (Fig. 2 last 2 columns) and aligning with the data manifold closely. We also theoretically proved the manifold alignment (EC drift is within an upper-bound) but did not include it due to page limit.

3) R1–Studying the effect of text We show quantitative results in Tab. 1 under “ours w/o text.” Removing text slightly degrades CLIP-T2I performance (as expected without text guidance), while focusing more on image quality (CLIP-I2I improves). The remaining metrics stay largely on par with main results.

4) R1–Notation & minor clarification We will provide more details of CBCT, and we will modify the notation. We cited DAPS, and we will cite the other two. We refer to our model as plug-and-play because it requires no training on the target dataset. The text guidance is intended as a strong prior.

5) R2–Two optimization steps per iteration The gradient steps can be viewed as proximal steps for sampler. SGS formulated this procedure as posterior sampling. Therefore, the optimization step is posterior sampling, and two steps follow prior methods such as DAPS and PnP-DM (Tab. 1).

6) R3–Trade off CLIP-T2I & Application scenario of full-training methods The lower CLIP-T2I score in Tab. 1 can be explained by the in-domain training of X2CT and Xray2CTPA, where all other methods adopt a prior model (MedSyn) trained on a different dataset, and it is used as a plug-in module without fine-tuning. If sufficient in-domain paired data are available, full-training is preferred. But for good performance across out-of-distribution observations / datasets, posterior sampling methods are preferred.



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.

    This paper received mixed scores 3, 4, 5.Concerns include the presentation clarity and method’s generalizability, as well as a potential incorrect formulation (raised by reviewer 1). AC invites the authors for a rebuttal to clear the concerns raised by all reviewers.

  • 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.

    After the authors’ rebuttal, R2 clarified that their main concern was a misunderstanding. All reviewers agree to the acceptance.



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 rebuttal and reviewer discussion resolve the main mathematical concern, and Reviewer 1 explicitly updated to Accept after recognizing that the LPIPS-related coercivity concern was not valid. Given the strong motivation, solid physics-guided diffusion formulation, and generally positive reviewer balance after rebuttal, the remaining concerns are mainly about positioning, notation, and real-world synthetic-to-clinical generalizability rather than core validity.



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.

    This paper addresses the important problem of sparse-view CT reconstruction by combining physics-based differentiable rendering with pretrained diffusion priors. The paper received consistently positive reviews, with all three reviewers recommending Accept. Reviewers appreciated the clear motivation, solid technical design, and comprehensive experimental evaluation with strong baselines and ablations. The proposed framework was considered relevant and likely to be of interest to the MICCAI community. The main concerns relate to the moderate level of novelty, as the method mainly integrates existing components, as well as the reliance on a synthetic evaluation pipeline, which limits evidence of real-world applicability. Some optimization choices also appear somewhat heuristic. Overall, I recommend acceptance. Despite the incremental nature of some components, the work is technically sound, well presented, and demonstrates meaningful improvements over prior methods on an important reconstruction task.



back to top