List of Papers Browse by Subject Areas Author List
Abstract
Idiopathic pulmonary fibrosis (IPF) exhibits variable CT progression, from subtle interstitial changes to overt architectural distortion. Longitudinal imaging data is often sparse and confounded by non-rigid deformation, making accurate delineation of disease progression from single timepoint data a fundamental clinical imperative. We investigate single-scan IPF progression in thoracic CT by synthesising a future scan conditioned on the baseline scan, time horizon and patient covariates. To preserve fine lung structure under limited disease-specific longitudinal data, we decouple a transferable thoracic generative/longitudinal prior from disease-specific adaptation. We first learn a detail-preserving latent representation using patch-based autoencoding with overlap-average stitching.
We then pretrain a covariate-conditioned latent diffusion model and longitudinal ControlNet on a large-scale thoracic CT screening dataset (>30,000 scans; >12,000 patients), and finally adapt the model to IPF progression synthesis with lung-focused patch-decoded supervision. Across internal IPF cohorts and a fully held-out external cohort, the screening-pretrained configuration achieves the strongest in-lung fidelity and lowest feature-space distance, and significantly outperforms training the progression adaptor without longitudinal pretraining. These results support thoracic diffusion pretraining as a reusable generative foundation for data-limited longitudinal disease modelling.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4966_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/niccolo246/baseline-to-future-ct
Link to the Dataset(s)
N/A
BibTex
@InProceedings{McCNic_From_MICCAI2026,
author = { McConnell, Niccolò AND Chen, Yunjie AND Wang, Shanshan AND Yamada, Daisuke AND Egashira, Ryoko AND Szmul, Adam AND Aslani, Shahab AND McCabe, John AND Yang, Tianqi AND McGregor, Helen AND Vasudev, Pardeep AND Azimbagirad, Mehran AND Jones, Mark AND Vanaudenaerde, Bart AND van Moorsel, Coline AND Wuyts, Wim AND Consortium, SUMMIT AND Mogulkoc, Nesrin AND Taylor, Paul AND Janes, Sam M. AND Alexander, Daniel C. AND Jacob, Joseph},
title = { { From Baseline to Future CT: Large-Scale Diffusion Pretraining for Single-Scan IPF Progression Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a four-stage generative framework for predicting disease progression in idiopathic pulmonary fibrosis (IPF) from a single baseline thoracic CT scan. The central design decision is to separate a transferable thoracic generative and longitudinal prior, learned from a large-scale lung cancer screening dataset (31.110 scans, 12.443 patients), from IPF-specific fine-tuning on a much smaller cohort (629 training patients, 986 scan pairs).
Stage 1 trains a patch-based 3D convolutional autoencoder on screening data, mapping 256³ volumes to 4-channel 64³ latents using lung-biased patch supervision and an autoencoder loss combining pixel fidelity, SSIM, gradient, KL regularization, and a 3D PatchGAN adversarial term. Stage 2 pretrains a covariate-conditioned latent diffusion model on full-volume screening latents, conditioning on age, sex, height, BMI, pack-years, and lung volume. Stage 3 pretrains a ControlNet adaptor on longitudinal screening pairs, injecting the baseline latent and time interval as control signal. Stage 4 fine-tunes this adaptor on IPF pairs with additional change-consistency regularizers and patch-decoded voxel-level supervision to enforce local parenchymal fidelity. At inference, DDIM sampling is combined with latent-average stabilization and overlap-average stitching for full-resolution 256³ output.
The central claim is that longitudinal pretraining on screening data is the dominant factor in enabling stable high-capacity 3D diffusion progression modelling, which would otherwise be intractable on cohort-scale IPF data alone.
- 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 pretraining strategy is the most clearly supported contribution. The decision to learn a thoracic prior from a large-scale screening dataset, where repeat imaging is available and acquisition variability is broad, and then adapt to a data-limited disease cohort, is a reasonable approach to a real bottleneck. The Scratch ablation, which retains the screening-pretrained autoencoder and diffusion backbone but trains the ControlNet exclusively on IPF data, degrades consistently across all reported metrics on both the internal and external test sets. The primary comparison is statistically significant after Holm correction on both sets, which makes the benefit of longitudinal pretraining the most credibly established result in the paper.
The evaluation design is a further strength. The paper reports both an internal test set and a fully held-out external site, giving at least a preliminary view of behavior under site shift. Statistical testing uses paired one-sided Wilcoxon signed-rank tests with Holm correction, which is appropriate given the directional hypothesis and the multiple metrics being tested. The AE ceiling (deterministic reconstruction of the ground-truth follow-up through the frozen autoencoder) contextualizes the headroom remaining in the progression modelling component: the ceiling is at 0.981 SSIM versus the proposed model at 0.947, indicating that the compression step itself is not the bottleneck.
Supplementing pixel-level metrics with a perceptual distance based on an open-source thoracic CT foundation model encoder is also appropriate. For IPF, where the clinically relevant features are local and spatially subtle, perceptual alignment in an encoded feature space is a more meaningful quality measure than global SSIM alone.
- 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 evaluation is restricted to image fidelity metrics, with no assessment of clinical utility. SSIM, MSE, PSNR, and ViT feature distance measure how closely a synthesized scan resembles a ground-truth follow-up, but they do not address whether the predicted progression is clinically informative. For IPF, the clinically relevant question is whether the model can predict the trajectory of functional decline with any accuracy. The paper acknowledges the absence of expert-annotation-free voxel-level progression measures as an active area of research, but that acknowledgment does not substitute for a downstream evaluation, even a surrogate one. Without some correlation between synthesized image changes and functional endpoints, the clinical value of the approach remains asserted rather than demonstrated.
The external test set contains 17 pairs. This is too small to draw reliable conclusions about generalization under site shift. The results show the same metric ordering as internal testing, which is encouraging, but with this sample size the standard deviations dominate and the evaluation cannot support strong generalization claims. The paper itself flags this as a limitation, but it is worth noting that “transfer feasibility” is the most that can be responsibly claimed, not stability across sites.
The comparison with prior work is narrow. The most directly comparable published method for single-scan IPF follow-up synthesis is Jiang et al. (2025) [6], published in Medical Image Analysis. The paper excludes it on the grounds that it uses text-based lung-function conditioning and 2.5D slice-based synthesis, which differ from the proposed setting. The exclusion is not unreasonable on technical grounds, but the clinical task is identical, and the paper would benefit from a more explicit discussion of what the input difference implies for the comparison. As it stands, the only comparison methods are an in-house latent U-Net and latent cVAE, both matched to the same latent resolution, which is a weaker competitive baseline than the published literature.
No ethics statement appears in the paper. The study uses multi-center patient imaging data collected across several countries. IRB approval or equivalent ethics review is not mentioned anywhere in the text. This is likely a consequence of anonymization requirements, but it should be added to the camera-ready version.
- 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 paper addresses a real and practically constrained problem: large-capacity 3D generative modelling for a disease with limited longitudinal imaging data. The proposed pretraining-to-adaptation strategy is principled, the ablation evidence directly supports the main claim, and the results hold on an external test set. The reported weaknesses are real, but they do not undermine the core technical finding. A weak accept reflects a solid contribution to the generative modelling for medical imaging literature, with the expectation that clinical grounding and broader benchmarking would strengthen the work in a follow-up.
- 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 initial review raised four concerns: limited clinical-utility assessment, the small external cohort (n=17), the omitted comparison with Jiang et al. [6], and the missing ethics statement. The rebuttal addresses all four. Clinical scope is honestly bounded as a methodological CT-synthesis study, the external cohort is reframed as supporting transfer feasibility, the Jiang et al. exclusion is defended on input-modality and conditioning differences I find reasonable, and ethics is confirmed.
The meta-reviewer’s additional asks on patch-wise justification and missing implementation and cohort details are also addressed, the latter with a full hyperparameter and architecture dump.
I recommend accepting it.
Review #2
- Please describe the contribution of the paper
This paper presents pre-training and fine-tuning strategies for 3D patch-wise generative models to predict future CT scans from baseline images. The large-scale pre-training improves future CT prediction, suggesting that the model serves as an effective predictor for IPF progression. Furthermore, the patch-wise training approach effectively preserves lung details. The evaluation, which includes external validation and appropriate metrics, is a notable strength.
- 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 assumption that generative pre-training can enhance longitudinal prediction performance is compelling.
2.Given that PSNR, SSIM, and MSE have inherent limitations in evaluating generative models, the inclusion of ViT distance and Airway error is a significant addition.
3.The study benefits significantly from a substantial pre-training dataset and rigorous external validation, which enhance the generalizability of the proposed model.
- 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.Lack of Evidence for Patch-wise Training Benefits The rationale behind using patch-wise training to preserve details is not sufficiently justified. The authors did not provide comparative results between whole-image training and patch-size ablations. Since other architectures like U-Net and CVAE can also be implemented in a patch-wise manner but may not show similar detail preservation, the authors should clarify if whole-image experiments were conducted for comparison.
2.Insufficient Qualitative Evidence and Performance Issues The qualitative evaluation is limited, with only two results provided. It is difficult to assess a generative model’s performance based on such a small sample. Specifically:In Fig. 2A, the Ground Truth (GT) at t=1 shows honeycombing patterns, which the proposed model fails to predict accurately.In Fig. 2B, the GT at t=2 exhibits ground-glass opacities, yet the model does not reflect these features well.Additionally, the authors should provide results for the ‘from-scratch’ training setup to demonstrate the actual gain from pre-training.
3.Limited Ablation of Baseline Models The ablation study is relatively weak. Given the recent advancements in generative modeling, why did the authors limit their comparison to only two models (U-Net and CVAE)? A broader range of diffusion-based baselines should be considered.
4.Missing Experimental Details The manuscript lacks essential experimental details such as batch size, learning rate, GPU specifications, and Hounsfield Unit (HU) windowing used in pre-processing. Furthermore, the absence of source code makes the figures and results difficult to reproduce.
- 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 authors claim contributions in three areas: pre-training, longitudinal synthesis, and the preservation of lung details.
While the first contribution is interesting, it is a well-established fact that larger datasets improve generative performance. Although the quantitative results are promising, it remains unclear whether these results stem from general high-fidelity generation (typical of diffusion models) or a genuine capability to model follow-up progression.
There are insufficient qualitative examples to support the model’s efficacy.
While patch-wise encoding can preserve details, the compared latent U-Net and CVAE models also operate under patch-wise settings. Without a direct comparison between patch-wise and full-image experiments, the third contribution is not convincingly validated.
Although this study has potential, it is not yet ready for publication in its current form.
- Reviewer confidence
Somewhat confident (2)
- [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.
First, regarding the classification of this paper as a methodological work: the MICCAI Call for Papers explicitly states that “manuscripts should clearly demonstrate innovations and contributions over state-of-the-art methodologies.” However, the ablation studies for your proposed method remain insufficient to justify this. Even if closely related prior works are scarce, simply comparing your model against basic architectures like a U-Net or VAE does not constitute a plausible ablation study to verify its superior performance over existing methods.
Second, although the authors state that the patch-wise approach is not a contribution, Section 2.1 is dedicated to it, suggesting it is a component of the proposed methodology. I remain skeptical about the “frequency-preserving” claims attributed to your model, especially since the ablation models utilizing the same patch-based framework do not seem to exhibit or address this property.
Lastly, the qualitative evidence remains weak. The specific details and outcomes of the senior thoracic radiologist’s review are not sufficiently clarified or demonstrated in the text.
Review #3
- Please describe the contribution of the paper
This paper proposes a four-stage 3D generative framework for synthesizing future thoracic CT scans from a single baseline scan in patients with idiopathic pulmonary fibrosis (IPF). The core idea is to decouple a transferable thoracic generative prior from disease-specific adaptation to overcome the limited size of longitudinal IPF cohorts. The authors propose a multistage approach, leveraging a pretrained ControlNet adaptor on longitudinal screening pairs that is then finetuned on IPF-specific pairs. At inference, follow-up latents are sampled via DDIM, averaged across m independent samples for variance reduction, and decoded with overlap-average stitching. The model is evaluated against a latent U-Net regressor and latent conditional VAE on internal IPF cohorts and a held-out external site, with ablations systematically isolating the contribution of longitudinal pretraining.
- 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) Authors attempt to addresses a genuine and difficult clinical problem, predicting IPF progression from a single baseline CT. Furthermore, their methodoly attempt to do this by decoupling strategy (large-scale thoracic pretraining followed by limited-data IPF adaptation) is well-motivated and addresses a real bottleneck in longitudinal disease modelling. (ii) Operating in a 256³ latent space rather than 2.5D slice-based synthesis better captures volumetric fibrotic structures (traction bronchiectasis, honeycombing) that are clinically meaningful in IPF. The patch-based autoencoding with overlap-average stitching is a practical and technically sound solution for making high-resolution 3D synthesis tractable. (iii) The paper includes internal and external held-out cohort evaluation (site-shift), a systematic ablation study isolating the contribution of longitudinal pretraining (Proposed vs. Scratch), and appropriate statistical testing (paired one-sided Wilcoxon signed-rank tests with Holm correction). The AE ceiling also provides a useful upper bound for contextualizing the progression modelling gap. (iv) Training the ControlNet adaptor from scratch on cohort-scale IPF data alone is insufficient for stable high-capacity 3D diffusion modelling. This is consistently demonstrated across all metrics on both the internal and external test sets (all Holm-corrected p < 0.001 internally; p < 0.05 externally).
- 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.
(i) Reported metrics (SSIM, MSE, PSNR, ViT feature distance, airway volume error) are generic image fidelity measurements. There is no radiologist assessment of whether the synthesized progression patterns (e.g., extent of honeycombing, ground-glass opacity change, reticulation) are clinically plausible or directionally accurate relative to ground truth. This limits the ability to assess whether the model is learning meaningful IPF-specific progression rather than generic texture fidelity. [Minor] (ii) The external cohort comprises only 17 baseline–follow-up pairs. While the site-shift result is encouraging, this sample size does not provide sufficient statistical power to draw robust conclusions about deployment generalization. The authors acknowledge this, but it remains a meaningful limitation for a key claim of the paper. [Minor]
(iii) Jiang et al. [6] is the most directly comparable single-scan IPF progression prediction method. The authors cite differences in input assumptions and generation regime to justify omitting a quantitative comparison, but at minimum a qualitative comparison or discussion of performance relative to reported numbers in [6] would substantially strengthen the experimental positioning. (iv) Key details necessary to reproduce the work are absent. For example, the depth and width of the 3D U-Net denoiser and ControlNet, the number of training steps/epochs for each stage, and the optimizer choice and learning rate schedule. Furthermore, the value of m (number of latent samples averaged at inference for stabilization) is missing. [Major] (v) The results tables report aggregate means and standard deviations across all test pairs without stratifying by disease severity or progression phenotype (e.g., stable vs. rapidly progressive IPF trajectories, baseline fibrosis extent, FVC decline category). Without this information, it is unclear whether the reported performance is driven primarily by stable IPF cases, where the follow-up scan closely resembles the baseline, or whether the model generalizes to the more clinically challenging rapidly progressive phenotype. This is particularly relevant for the external cohort, where patient demographics and disease stage distributions relative to the internal cohort are unreported. Stratified results or at minimum a table of cohort characteristics (age, sex, baseline FVC % predicted, median follow-up interval) would allow readers to contextualize the clinical scope and potential bias of the reported metrics
- 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
(i) Authors need to include hyperparameter values for training and fine-tuning of the ControlNet on both the large-scale screening dataset (Stage 3) and the IPF-specific dataset (Stage 4). Although certain values such as loss weights (λ) would be released with the code, either a table or one-line description of evaluated ranges would be beneficial. Learning rates, batch sizes, and number of training steps per stage should also be included in the methodology as either a short sentence or table. (ii) The paper claims to outperform related methods but does not include a direct quantitative or qualitative comparison with Jiang et al. [6], which is the most relevant prior work for single-scan IPF progression prediction. Authors should clarify why this comparison is infeasible and ideally provide at least an indirect performance contextualization. (iii) Authors should provide a cohort characteristics table (age, sex, baseline FVC % predicted, median follow-up interval, distribution of stable vs. progressive trajectories) for both the internal and external test sets. Without this, it is difficult to assess whether reported metrics are inflated by an over-representation of stable IPF cases or whether the model generalizes to rapidly progressive phenotypes, which are more clinically consequential.
- 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?
Lack of hyperparameter description in methodology, along with missing demographic information regarding the datasets makes the results difficult to interpret alone. Methodologically, ablation experiment alongside comparison to a model trained from scratch and U-Net comparison are satisfactory and paper is relatively interesting.
- 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.
The methodology proposed in this work is innovative and addresses a current critical problem in medical imaging regarding generating synthetic future for 3D volumes when follow-ups are limited. Authors address the limited comparison provided in the original manuscript and provide a rational for it, although this should still be included in discussion section.
Authors have addressed missing experimental parameters to ensure reproducibility for the given methodology, which was a major limitation of the original manuscript.
Author Feedback
We thank the Reviewers and Chairs.
Clinical scope/metrics (R2-R4). We agree that radiologist scoring and downstream endpoints are important for assessing clinical utility. This is a methodological study of longitudinal CT synthesis: generated follow-up CT is evaluated against observed follow-up in-lung, structurally and in thoracic feature space. We do not claim deployment-level utility. The Discussion confirms our focus on in-lung similarity rather than an IPF-specific endpoint, for which annotation-free CT-based standards remain underdeveloped. CT foundation-model distance and airway error assess thoracic feature alignment and airway structure, complementing SSIM/MSE/PSNR.
Progression vs generic generation (R3). R3 questioned whether gains reflect progression modelling or generic high-fidelity generation typical of diffusion. Proposed vs Scratch ablation isolates this: both use the same screening-pretrained autoencoder (AE) and diffusion (LDM) backbone, differing only in ControlNet (CN) longitudinal pretraining; thus latent image quality and diffusion capacity are controlled. Scratch degrades across all metrics, including airway error (Tables 1-2), while Proposed significantly improves pre-specified fidelity/perceptual metrics. This suggests gains arise from a transferable longitudinal adaptor, not merely a high-capacity diffusion backbone.
Patch-wise design (R3). We agree that patch/full-volume ablations would characterise this design. However, the main isolated claim is longitudinal pretraining, not universal superiority of patch-wise training. Patch-wise autoencoding and overlap-average stitching make 256^3 synthesis tractable while preserving a 64^3 latent grid. Direct 256^3 AE training at comparable capacity exceeded A6000 memory, while reducing capacity would confound comparison; stronger compression attenuated high-frequency detail in senior board-certified radiologist review. The high AE ceiling in Tables 1-2 indicates that the representation preserves follow-up detail, so the remaining gap is progression modelling rather than decoding. All comparison models use the same latent representation.
Qualitative evidence (R3,R4). We agree that Fig. 2 cannot exhaustively establish clinical plausibility. The cases are illustrative, and evidence rests primarily on aggregate paired evaluation across all test pairs. In these examples, senior thoracic radiologist review identified local fibrotic progression patterns, including traction bronchiectasis in zoomed regions, which were absent or over-smoothed in baseline outputs. Predicting exact emergence/extent of subtle patterns from one scan remains challenging; definitive assessment requires clinical endpoint validation.
Prior work (R2,R4). Jiang et al. is the closest clinical-task prior, but direct comparison is not straightforward because inputs, datasets, and generation regimes differ. Jiang et al. use OSIC data, lung-function-test guidance via CLIP text encoding, and 2.5D slice generation; our framework uses baseline CT, time horizon, and covariates for full-volume 3D synthesis in a 64^3 latent space. Reported numbers are not directly comparable. Without identical conditioning, data, and implementation, comparison would conflate formulation with performance. We therefore used baselines with identical inputs/conditioning (latent U-Net/cVAE), providing a controlled evaluation.
Reproducibility/cohort/ethics (R2-R4). Scans were clipped to [-1200,800]; AdamW; initial LRs AE/LDM/CN/IPF-finetune=1e-4/3e-5/6e-5/4e-6; effective-batch =6/16/4/4; epochs=120/100/50/30.AE channels=(64,128,256); LDM/CN channels=(128,256,512), 2 ResBlocks, condition channels=5. Diffusion used T=1000, DDIM=300, m=5 on Nvidia A6000.Configs/code will be released upon acceptance. Scans were de-identified and used under relevant data-governance/ethics approvals. Details were omitted for anonymisation. The external cohort is modest, supporting transfer feasibility, not deployment generalisation.
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.
The reviewers acknowledge a four-stage generative framework with strong evidence supporting the benefit of large-scale longitudinal pretraining and generally sound evaluation design. However, concerns remain regarding limited clinical validation, small external cohort, insufficient qualitative evidence, and incomplete experimental details. For rebuttal, authors should clarify clinical relevance, justify patch-wise design, expand comparisons to prior work, and provide missing implementation and cohort details to strengthen credibility.
- 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 satisfactorily addressed the major concerns, and the paper now presents a solid and well-supported contribution to longitudinal 3D medical image generation.
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.
Reviewers appreciated the generative diffusion based approach for pretraining to longitudinally predict IPF. However, limited baselines and ablations is still considered a weakness. Authors must clearly explain the limitations of the study and rationale for the specific set of experiments performed.
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.
After rebuttal, the post-rebuttal opinions lean positive, with two reviewers recommending acceptance and one maintaining rejection. Reviewer #2 found that the rebuttal adequately addressed concerns about clinical scope, the small external cohort, comparison with Jiang et al., and the ethics statement. Reviewer #3 also recommends acceptance, noting that the method is innovative for future 3D CT synthesis under limited follow-up data and that missing experimental details were clarified. However, Reviewer #4 remains negative due to insufficient comparison with stronger state-of-the-art methods, limited ablation support for the patch-wise design and frequency-preserving claims, and weak qualitative/radiologist evidence.
Overall, the final version should clearly discuss the remaining limitations regarding clinical validation, small external cohort size, qualitative evidence, and incomplete comparison to stronger generative baselines. The authors should provide the full open-source implementation to improve the reproducibility.
