List of Papers Browse by Subject Areas Author List
Abstract
Multiphase contrast-enhanced computed tomography (CT) is clinically significant in providing vascular structure and lesion phase-specific enhancements. Yet, its clinical utility is constrained by intrinsic contrast agent-associated risks (e.g., nephrotoxicity, allergic reactions) and multiphase cumulative radiation exposure. To tackle this, synthesizing contrast-enhanced CT (CECT) using non-contrast CT (NCCT) offers a potential alternative. However, achieving a high-quality synthesis of multiphase CECT remains challenging due to the contrast agent (CA)-induced complex contrast flow dynamics and the specific variations across phases. Therefore, this paper proposes a contrast flow pattern and cross-phase specificity-aware diffusion model for NCCT-to-multiphase CECT synthesis. Specifically, a contrast flow pattern learning mechanism is integrated into the conditional diffusion model, which enables orderly phase transitions while ensuring anatomically and temporally coherent enhancement synthesis. Furthermore, a phase distinction network is introduced to align cross-phase specificity features with the contrast features in synthesized CECT images. Experimental results on multicenter abdomen CT datasets have demonstrated the superiority of our method compared to state-of-the-art methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4773_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Kindyz/CFPS-Diff.git
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZheKai_Contrast_MICCAI2025,
author = { Zheng, Kaiyi AND Huang, Mu AND Li, Xinming AND Ma, Jianhua AND Feng, Qianjin AND Yang, Wei AND Zhong, Liming},
title = { { Contrast Flow Pattern and Cross-Phase Specificity-Aware Diffusion Model for NCCT-to-Multiphase CECT Synthesis } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15963},
month = {September},
page = {89 -- 99}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses the challenging problem of synthesizing multiphase contrast-enhanced CT (CECT) scans from a non-contrast CT (NCCT) input. The core difficulty arises from the complex contrast dynamics and phase-specific variations induced by contrast agents. To tackle this, the authors propose a conditional diffusion model framework augmented with two novel modules: Contrast Flow Pattern Learning (CFPL) and Cross-Phase Specificity Learning (CPSL). CFPL encodes order-aware temporal dependencies using a phase triplet embedding (e_pre, e_post, e_t) to modulate the generation process without accessing image features directly. CPSL introduces a Phase Distinction Network (PDN) that classifies phase identity and enforces cross-phase alignment through a feature consistency loss between the PDN encoder and CFPL decoder.
- 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.
Clinically meaningful task: Synthesizing multiphase CECT from NCCT has practical importance in reducing contrast use and radiation exposure, particularly in oncological imaging.
Structured design: The separation of modeling contrast flow dynamics (CFPL) and phase-specific feature alignment (CPSL) reflects a thoughtful and task-aware modularization.
Order-aware conditioning: Encoding the temporal flow as a triplet (e_pre, e_post, e_t) to guide synthesis via a phase sequence embedding module is a novel perspective for temporal modeling in diffusion-based generation.
Auxiliary supervision for phase discrimination: Using phase classification with class-distance-weighted loss and its connection to the generator via feature alignment introduces cross-task regularization that can be beneficial for preserving phase-specific information.
- 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.
Weaknesses
Limited novelty in components and architectural design
While the CFPL module proposes to modulate features using a phase triplet embedding, this amounts to a feature-free affine transformation, similar to previously established modulation techniques like FiLM or AdaIN.
The CPSL relies on a standard classification network and feature consistency loss, which resemble common strategies in domain adaptation and cross-modal alignment (e.g., StarGAN, MUNIT).
The technical novelty thus lies more in integration than in the architectural originality of individual modules.
Ambiguity in conditioning and supervision mechanisms
The flow embedding module (omega_flow) receives phase and time codes but does not consider the current feature map in predicting modulation parameters. This design may limit context sensitivity and adaptability to anatomical variations.
The training procedure for jointly optimizing diffusion reconstruction, phase classification, and feature alignment is not clearly described, raising questions about training stability, optimization conflicts, and convergence behavior.
Lack of reproducibility and benchmarking transparency
The training and validation datasets are entirely private, with no publicly available counterpart. While the internal dataset includes 547 patients and an external set of 116 cases is used for testing, the lack of dataset accessibility severely limits reproducibility.
There is insufficient detail about the acquisition protocols for the external dataset, leaving the domain shift and generalizability unclear.
Weak qualitative evidence for claimed improvements
Despite the claim that the generated images are anatomically and temporally coherent, the visual examples presented in the paper do not convincingly demonstrate such advantages.
The lack of temporally progressive visualization (e.g., over sequential phases) or side-by-side comparisons highlighting flow consistency undermines the claim of improved contrast dynamics modeling.
Compared to prior works (GAN/VAE-based synthesis), there is no compelling evidence that the diffusion model results in superior visual quality, detail preservation, or realism.
Underwhelming ablation and contribution justification
Although an ablation study is presented, it lacks depth in analyzing how each module (CFPL, CPSL, PDN) contributes specifically to image fidelity or phase distinction. No analysis or visualization of learned embeddings or modulation patterns is provided to support the claim that contrast flow dynamics are effectively captured.
- 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 provide sufficient information for reproducibility.
- 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/2025/en/REVIEWER-GUIDELINES.html
Suggestions for Improvement Provide more compelling qualitative results, ideally including time-series comparisons across phases or contrast flow visualizations, to validate the claim of temporal coherence.
Consider releasing at least a subset of the dataset or replicating results on a public dataset to enable reproducibility and fair comparison.
Clarify and analyze the effectiveness of the flow-aware conditioning — how does it improve performance compared to standard phase token or learned embeddings?
Include more detailed visualization of internal representations (e.g., phase embeddings, attention maps, or intermediate feature activations) to support the interpretability of the contrast dynamics modeling.
- 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 — could be rejected, 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 a diffusion-based method for synthesizing multiphase CECT from NCCT by modeling contrast flow dynamics and phase-specific features. While the task is clinically meaningful and the modular design (CFPL + CPSL) is conceptually well-organized, the technical novelty is only moderate. The core modules rely on variations of existing techniques (e.g., affine modulation, feature consistency losses) and do not introduce fundamentally new architectural paradigms.
The main concerns leading to the weak reject are:
Reproducibility is severely limited by the use of a private dataset without publicly accessible code or benchmark, making it hard to validate the claimed contributions.
The qualitative results are underwhelming and do not clearly support the claim of anatomical or temporal coherence.
Ablation and interpretation of internal mechanisms are insufficient, leaving the reader unconvinced about the necessity and effectiveness of the proposed modules.
Despite the use of diffusion models, no clear evidence is provided that this choice leads to substantial improvements over simpler alternatives.
Overall, while the direction is promising, the experimental validation and qualitative presentation fall short of the standards required for acceptance.
- Reviewer confidence
Very confident (4)
- [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 addresses several points raised in the initial review and provides additional clarification regarding the motivation and functionality of the two main modules: Contrast Flow Pattern Learning (CFPL) and Cross-Phase Specificity Learning (CPSL). The authors explain how these components are intended to reflect the temporal progression and phase-specific characteristics of contrast-enhanced CT imaging.
CFPL is described as a mechanism for incorporating temporal order into the synthesis process through phase triplet embeddings, distinguishing it from static modulation methods like FiLM or AdaIN. CPSL is introduced as a cross-phase alignment strategy that applies multi-scale constraints to preserve diagnostic contrast differences across phases, differentiating it from conventional domain classifiers or disentanglement networks used in prior work.
While these explanations clarify the intended behavior of the proposed modules, the underlying mechanisms—such as affine conditioning and feature alignment—remain variations of well-established techniques in generative modeling and domain adaptation. The core contributions, while task-specific, do not represent significant departures from standard practices. As a result, the technical novelty of the work remains moderate.
Overall, the paper presents a method that is aligned with a meaningful application and is conceptually well-organized. However, the architectural contributions are incremental, and the evaluation and interpretability components could be further strengthened. For these reasons, I maintain my original recommendation.
Review #2
- Please describe the contribution of the paper
The authors propose a diffusion-based model for generating synthetic multi-phase contrast-enhanced CT (ceCT) images from an initial non-contrast CT (ncCT) scan. The model is designed to synthesize any of the three contrast phases—arterial, venous, and delayed—from a single ncCT input. To encourage physiologically plausible outputs, the model incorporates design choices that enforce contrast flow pattern learning and enhance phase-specific features. The evaluation includes standard image quality metrics (MAE, PSNR, SSIM) as well as perceptual and distributional similarity measures (FID and LPIPS).
- 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 is well written and includes a thorough ablation study, as well as clearly reported quantitative and qualitative results. The study design is comprehensive, with the proposed model capable of predicting all three phases of multiphase contrast-enhanced CT (ceCT) images. The modifications made to the diffusion model appear well motivated by the physiological changes that occur during contrast administration. This is achieved by conditioning the prediction of each phase on the preceding phase, while avoiding information leakage from future phases. Additionally, a specificity loss term is introduced to reduce the similarity between the contrast-enhanced phases, thereby promoting phase distinctiveness. Overall, the proposed ideas are novel, well explained, and grounded in domain knowledge.
- 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 my main criticisms of this paper pertains to the problem formulation. Based on the introduction, the generation of synthetic contrast-enhanced CT (ceCT) images is motivated by the claim that “poor visibility of abnormalities and organ contours in NCCT scans increases the diagnostic difficulty for radiologists.” To my understanding, this is intrinsically an ill-posed problem: ceCT images are acquired in clinical practice precisely because non-contrast CT (ncCT) lacks sufficient contrast to visualize certain anatomical structures and pathologies—at least to the human eye. Therefore, the diffusion model is, in essence, tasked with “making up” information that is not directly observable in the input, which draws parallels to CT-to-MR translation. This raises important questions about the distribution of healthy vs. pathological cases in the training set and the extent to which the synthetic images faithfully represent underlying pathologies.
By raising this point, I am not trying to downplay the technical quality of the work presented, rather, I hope to better understand the motivation behind the problem formulation and highlight the types of evaluations that may be necessary to validate such an approach. I also hope to encourage broader discussion around this topic during a potential rebuttal phase.
In relation to this discussion and Figure 2 (showing qualitative results), I am curious how well the reported metrics (e.g., MAE, SSIM, FID) reflect the model’s ability to capture the diagnostic-relevant contrast enhancement patterns, as opposed to simply reproducing globally realistic CT images. It appears that many of the errors are localized around bony structures. Perhaps computing the metrics specifically over soft-tissue regions (e.g., liver, kidneys, pancreas) could provide a more meaningful assessment of the model’s performance.
Alternatively, if the motivation is to support downstream medical image analysis tasks (such as organ segmentation), then a more relevant evaluation would involve demonstrating improved segmentation performance using the synthetic ceCT images. While I don’t expect such results to be included in a potential rebuttal, I believe this should be discussed more explicitly in the discussion section to clarify the envisioned utility of the proposed method.
- 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 provide sufficient information for reproducibility.
- 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/2025/en/REVIEWER-GUIDELINES.html
I have included some of my comments to the authors in the weaknesses section, as they directly relate to key points I wanted to raise. Below are a few additional minor comments/questions for the authors’ consideration: The paper reports the use of input volumes of size 256x256x8. Could the authors clarify whether the diffusion model is 2D or 3D? Additionally, are the results reported for full 3D volumes using a sliding-window inference scheme (and if so, do they apply overlapping windows?), or only for the central 8 slices?
Lastly, the authors state that their sampling strategy enforces eps_pre >= eps_post - why not simply eps_pre > eps_post? And regarding the proposed CDW loss term: given that it does not appear to improve the reported quantitative metrics, could the authors provide further justification for its inclusion, either quantitatively or qualitatively?
- 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?
I believe the methodological design and data flow of the proposed model are well thought out and effectively mirror the physiological dynamics of contrast flow. The reported results support the validity of this design. While I have raised several concerns in the paper weaknesses section - particularly regarding the problem formulation, evaluation scope, and broader implications - I view these more as important points for future discussion rather than fundamental flaws in the current study. Overall, the methodological contribution is solid and well-supported by the experiments, which leads me to recommend acceptance of this paper.
- 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 overall opinion of the paper remains unchanged. I believe the methodology effectively addresses a clinically relevant problem, and I support its acceptance. That said, I feel the rebuttal did not adequately address my major concerns, as it focused almost entirely on Reviewer 1’s comments.
Review #3
- Please describe the contribution of the paper
The authors present a novel framework, for synthesizing multiphase contrast-enhanced CT (CECT) images from non-contrast CT (NCCT) by modeling dynamic contrast flow and cross-phase specificity (CFPS). To address the key challenge of accurately replicating contrast-agent-induced enhancement patterns across arterial, venous, and delayed phases, the authors propose a contrast flow pattern learning module, CFPS-Diff, integrated into a conditional diffusion model. This integration enables smooth, temporally consistent phase transitions during synthesis. Additionally, a cross-phase distinction network is introduced to align phase-specific features with learned contrast information, thereby improving anatomical fidelity and temporal coherence in the synthesized CECT images. CFPS-Diff is validated on multi-center abdominal CT datasets and demonstrates superior performance over existing state-of-the-art image synthesis approaches. The integration of dynamic contrast flow modeling and phase-specific learning constitutes a significant contribution to the field of medical image synthesis.
- 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 major strengths of the paper lie in its novel approach and comprehensive evaluation. The authors propose a contrast flow plattern specific-aware (CFPS-Diff) model integrated with a phase distinction network to tackle the challenge of multiphase CECT synthesis from NCCT images. This is a significant contribution, as accurate modeling of dynamic contrast flow and phase-specific enhancement patterns is crucial for clinical diagnosis using CT imaging. The integration of a contrast flow pattern learning mechanism and a cross-phase specificity module ensures that synthesized images are anatomically and temporally coherent The method is rigorously evaluated using ablation studies and comparisons with multiple state-of-the-art (SOTA) methods (U-Net, Pix2Pix, CycleGAN, TransUNet, and DDPM). The results demonstrate consistent superiority of the proposed approach across various image quality metrics including MAE, SSIM, PSNR, FID, and LPIPS. Furthermore, the experiments span multiple datasets and all three CECT phases (arterial, venous, and delayed), highlighting the generalizability of the method. The enhancement of arterial structures and improved quantitative metrics reinforce the clinical relevance and potential translational impact of this 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.
While authors present strong comparisons to multiple SOTA models, the evaluation is limited to abdominal CT data only. The generalizability of the CFPS-Diff method to other anatomical regions or modalities (e.g., thoracic, brain, or other CT) is not assessed. Furthermore, despite proposing clinically relevant applications, there is no direct clinical validation or feedback from radiologists to support the feasibility or interpretability of the synthesized phases. A discussion of potential failure cases, limitations in contrast modeling, or robustness to variations in acquisition protocols would further strengthen the work.
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
In the Introduction, paragraph 2, line 11, there is a typographical error - “iwhere” should be corrected to “where.” Additionally, in line 12, the term “pronieer” appears unclear; the authors should clarify or correct it - presumably, they intended to write “pioneer.”
Authors stated that CT image datasets were collected from patients at a local hospital but did not confirm whether all procedures adhered to relevant ethical guidelines and regulations. Authors should explicitly state that all experimental protocols were approved by an institutional review board or ethics committee and that informed consent was obtained from all subjects and/or their legal guardians.
Lastly, the authors should consider moving Tables 2 and 3 further down in the manuscript to avoid splitting the conclusion. This would ensure that the conclusion reads as one cohesive paragraph without interruption.
- 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 paper introduces a novel diffusion-based framework for synthesizing multiphase CECT images from NCCT, addressing a clinically relevant challenge. It combines contrast flow pattern learning and cross-phase specificity modeling to ensure anatomically and temporally consistent synthesis. The method is well-validated on multi-center datasets, outperforms existing approaches across key metrics, and demonstrates strong potential for clinical translation. Given its innovation, clinical impact, and strong experimental results, it is well-suited to be accepted.
- Reviewer confidence
Very confident (4)
- [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 for their constructive comments.
Novelty; Design (R1) 1)Although CFPL uses affine transforms like FiLM/AdaIN, its key novelty lies in temporally constrained sampling (e_pre≤e_post) and dynamic phase flow embedding aligned with diffusion steps. Unlike FiLM’s static conditioning, CFPL captures the monotonic contrast enhancement (CE) flow patterns essential for multiphase CT. Given the goal of synthesizing multiphases, AdaIN may cause phase convergence, while FiLM lacks sensitivity to CE flow. 2)CPSL leverages a discriminator to extract multi-scale phase-difference features, injecting them into the decoder of the diffusion model, forcing CFPL to perform CE phase synthesis under non-CE phase conditions (e_pos=0). When CFPL is conditioned on CE phases (e_pos in {1,2,3}), CPSL’s features are aligned with the synthesized features from CFPL. In contrast, StarGAN’s domain classifier focuses only on authenticity and categories of real/synthetic images, and MUNIT’s style-content disentanglement targets modality diversity rather than pathology-aware consistency. CPSL enforces clinically driven inter-phase alignment using multi-scale constraints, preserving diagnostic contrast kinetics instead of arbitrary style transfer. 3) All modules are optimized alternately, freezing others at each step.
Reproducibility and data (R1) Due to anonymity, public link was withheld before; code, dataset subset, and synthetic samples will be released upon acceptance. External set comprises liver metastases cases with rare pathological types (e.g., gallbladder sarcomatoid carcinoma, pancreatoblastoma) not included during training.
Improvements (R1) CA-induced contrast flow progresses as: NCCT (no CE), AP (aorta/renal cortex peak), VP (portal vein/liver parenchyma dominance), DP renal medulla/late hepatic washout). Despite 2D slice limitations, synthetic phases in Fig. 2 (col12) preserve this progression: aortic HU drops from AP→DP (165±57→73±12), while liver CE peaks in VP (137 vs NCCT 98), matching clinical kinetics. Critically, difference maps show absent renal cortex CE in AP w/o CFPL/CPSL (col11, 38±24 vs real 86±30), but recovered in full model (col12, 76±27). Full ablation visuals omitted due to space, but further analysis shows CFPL improves anatomical/CE coherence; CPSL sharpens phase-specific CE. Compared to Pix2pixHD’s bowel misenhancement, Pix2pix/CycleGAN’s NCCT-like outputs, and ResViT’s liver artifacts, CFPS-Diff yields clearer, phase-specific images with significant gains (p<0.05) across multi-organs.
Theoretical basis (R2) NCCT reflects baseline HU and hides latent attenuation cues—e.g., subtle hyperattenuation in vascular-rich or heterogeneous lesions signals their propensity for contrast uptake. Specifically, pathology-driven endothelial disruption permits contrast extravasation; on NCCT, necrotic cores and perilesional edema appear hypoattenuated, correlating with ring or delayed enhancement patterns on CECT. Even without contrast, vessel morphology and feeding branches are partially visible, guiding arterial-phase synthesis along real vascular paths. Our conditional diffusion model learns to align latent cues with phase-specific HU shifts, generating enhancement that mirrors perfusion dynamics.
Other regions (R2) On the CT-2 set (and external set), our model outperforms SwinUNETR with MAE ↓3.49–6.55% (2.79–4.51%), PSNR ↑0.57–3.92% (0.24–2.30%), and SSIM ↑0–5.91% (0.35–3.72%) across lesions and abdominal organs (liver, kidney, spleen, stomach, pancreas), with largest gains in lesions.
Generalizability (R3) Multiphase CT is standard in abdomen due to key CE dynamics, yet thoracic/brain CTs typically use single-phase (CFPS-Diff not applicable). Though cross-region generalization is untested, our method shows robustness to intra-abdominal heterogeneity, evidenced by consistent performance across an external set with rare pathological types. Radiologist validation and modality extension are planned as future work.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This paper presents a novel diffusion-based framework for synthesizing multiphase CECT images from NCCT, addressing a clinically important challenge in contrast-enhanced imaging. The method effectively integrates contrast flow pattern learning and cross-phase specificity modeling to achieve anatomically and temporally consistent image synthesis. The design of the model thoughtfully mirrors the physiological dynamics of contrast flow, and the reported results validate the strength of this approach.
The paper is well-supported by rigorous experimentation on multi-center datasets and consistently outperforms existing methods across key metrics. While some concerns were raised regarding the scope of evaluation and broader implications, these are framed as directions for future exploration rather than limitations of the current study.
Overall, the methodology is innovative, the clinical relevance is clear, and the experimental results are compelling. The paper received positive feedback fro reviewers and recommended for 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.
Reject
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
Theoretically, generating three-phase dynamic CTA from a single non-contrast CT scan is highly challenging. Although the proposed method was evaluated using several image quality metrics, it failed to demonstrate clinically critical aspects, such as the temporal information highlighted by Reviewer 1. It would have been informative to visualize how vascular structures change across the three generated phases; however, these results were not presented. Furthermore, establishing a clearer link between the generated outputs and their diagnostic value would enhance the clinical relevance of the work. Given the current limitations in evaluation and validation, I would adopt a more conservative stance and do not support acceptance in its current form.