{"id":"ede389b9-d9ef-4c1e-a116-1bdca4963420","arxiv_id":"2502.05713","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 3D-VQ-GAN combined with a latent neural ODE generates longitudinal IPF CT scans at arbitrary time points, and codebook-frequency biomarkers from the generated scans achieve C-indices of 0.886 (cross-sectional) and 0.959 (longitudinal) versus 0.929 and 1.0 for real scans.","lead":"This paper trains a two-stage model that compresses 3D lung CT scans into a compact codebook and then uses a neural ODE to predict how those codes change over time, generating synthetic CT scans at arbitrary future or past time points. The authors show the generated scans retain enough prognostic information that imaging biomarkers derived from them predict patient survival nearly as well as biomarkers from real scans.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that the model captures disease progression is not yet supported: no comparison against a trivial 'no-change' baseline (e.g., reusing the most recent scan) is reported, so generated scans may simply inherit input information and the survival C-indices may reflect baseline severity rather…","rationale":"The reader's verdict of CONDITIONAL is appropriate, but the weakest assumption they highlight (shared deterministic ODE dynamics) is not the most load-bearing issue for the central claim. The paper explicitly acknowledges this assumption in Section 5, and even a shared dynamic could still produce useful average progression trajectories from patient-specific initial states. A more fundamental gap is the absence of any trivial baseline comparison. Table 1 only ablates internal architectural choices; no experiment tests whether simply reusing the last scan outperforms or matches the model. Since IPF progression is often radiologically subtle over short intervals, a no-change baseline may achieve high similarity scores. Similarly, the survival validation in Section 4 does not show that the generated future scan adds prognostic information beyond the input scans. If the C-index from generated scans is no better than the C-index from the most recent real scan, then the claimed clinical utility of temporal extrapolation is unsubstantiated. This concern is concrete and addressable: adding a no-change baseline is inexpensive and would either support or refute the central claim. The reader's rationale does mention missing baseline comparisons as one of several weaknesses, so there is partial agreement, but their formal weakest_assumption is different. I keep the verdict at CONDITIONAL because the missing experiments are fixable and the paper does provide qualitative examples that suggest the model can generate plausible fibrotic patterns; however, acceptance should require the baseline comparison and reporting of confidence intervals.","tokens_in":14928,"tokens_out":4459,"duration_ms":47488,"concrete_test":"Run the same extrapolation evaluation with a trivial baseline: for every test patient, set the generated scan at each future time point to the last observed input scan (or its registered version), and compute SSIM/PSNR against ground truth. Also compute the Cox C-index using biomarkers derived from the last observed real scan instead of the generated scan, with bootstrap confidence intervals for both the model and the baseline. If the trivial baseline is within the error bars of the 4D-VQ-GAN results on image metrics and survival C-index, the paper's progression claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Table 1, the ablation compares only architectural variants (ConvGRU vs ODE-ConvGRU, with/without skip connections, masked inputs). All variants share the same two-stage framework; none is a trivial predictor. For extrapolation, the obvious baseline is to output the last observed scan (or its latent embedding) unchanged. IPF changes are often subtle, so such a baseline could achieve SSIM/PSNR close to the reported values (0.489 SSIM, 17.8 PSNR for the proposed model). If so, the quantitative evidence does not demonstrate that the model predicts progression beyond a copy. Similarly, the survival analysis in Section 4 reports C-indices for generated scans (0.886 cross-sectional, 0.959 longitudinal) without comparing to biomarkers derived from the most recent real scan. Because the model uses two input scans and predicts a future scan, a high C-index may be driven by baseline disease severity already present in the inputs, not by generated temporal change. Thus the central claim of 'capturing disease progression' and 'personalised' modelling is not established by the reported experiments. This is the most load-bearing gap because it directly undermines the headline claim, whereas the shared-dynamics limitation acknowledged in Section 5 is a modeling assumption that, even if imperfect, could still yield useful average trajectories.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes 4D-VQ-GAN, a two-stage generative model for synthesizing longitudinal CT volumes in IPF patients. Stage 1 trains a 3D-VQ-GAN to compress CT volumes into a discrete codebook; Stage 2 trains a neural ODE on the quantized latent embeddings, conditioned on two observed scans, to generate volumes at arbitrary time points. The authors evaluate interpolation and extrapolation quality with MSE/SSIM/PSNR, perform an ablation over architectural variants, and validate clinical utility by deriving imaging biomarkers from generated scans and computing C-indices for survival prediction. They report C-indices of 0.886 (cross-sectional) and 0.959 (longitudinal) for generated scans, compared with 0.929 and 1.0 for real scans. The paper is clearly written, but the experimental validation is not sufficient to establish the central claim that the model captures disease progression in a personalized manner.","tokens_in":15173,"tokens_out":5244,"duration_ms":52622,"significance":"The paper addresses a relevant and difficult problem: generating longitudinal 3D medical images from sparse, irregular observations. The two-stage design (VQ-GAN compression + latent ODE) is sensible, and the use of a real longitudinal IPF cohort with survival outcomes is a strength. The codebook visualization and the external validation of the reconstruction stage are also commendable. However, as presented, the experimental validation does not establish the central claim that the model 'captures disease progression' or provides 'personalised' predictions. The absence of trivial baselines, the lack of statistical comparison for survival C-indices, and the untested assumption of a shared dynamics mean the quantitative results could be explained by simpler mechanisms such as copying the last observed scan or reflecting baseline severity. If the required baselines and analyses confirm the added value, the contribution would be significant for the medical imaging community.","major_comments":[{"comment":"The ablation compares only variants of the proposed architecture (ConvGRU vs ODE-ConvGRU, skip connections, masked inputs), with no comparison against trivial baselines such as copying the last observed scan or its latent embedding, or linear interpolation in latent space. Since IPF changes between scans can be subtle, the reported extrapolation SSIM of 0.489 and PSNR of 17.816 may be close to a no-change baseline; without such comparisons, the quantitative results do not demonstrate that the model predicts progression beyond the information already present in the input scans.","section":"Section 4, Table 1"},{"comment":"The claim that generated scans yield C-indices 'comparable' to real scans is not supported by appropriate statistical evidence: no confidence intervals, no significance tests, and the test set is small (44 patients). The longitudinal C-index of 1.0 for real scans suggests overfitting or an overly homogeneous test set, and the cross-sectional difference (0.886 vs 0.929) is not shown to be within sampling variability. Moreover, there is no comparison with biomarkers derived from the most recent real scan, so the high C-index of generated scans may reflect baseline disease severity rather than the model's predicted temporal change.","section":"Section 4, Survival outcome prediction"},{"comment":"The model uses a single deterministic neural ODE f_theta, which the authors acknowledge 'assumes a shared disease dynamic across all patients.' This assumption is not tested, yet the title and introduction claim 'personalised disease progression modelling.' To support this claim, the authors should demonstrate that generated trajectories vary meaningfully across patients (e.g., by testing whether patient-specific initial states lead to distinct predicted changes, or by comparing against a model with patient-specific dynamics).","section":"Section 2, Eq. (6); Section 5"},{"comment":"The quantitative evaluation relies on image-similarity metrics (MSE, SSIM, PSNR) that do not directly measure whether the generated changes correspond to pathological progression. The authors should include clinically meaningful metrics, such as changes in quantitative fibrosis extent or the imaging biomarkers already used in their survival analysis, to show that the generated temporal changes are in the right direction and of plausible magnitude.","section":"Section 3, Evaluation Metrics"}],"minor_comments":[{"comment":"The variant names are confusing; for example, 'ODE encoder (ConvGRU)' and 'ODE encoder (ODE-ConvGRU)' could be more clearly distinguished, and the table should state which components are ablated in each row.","section":"Table 1"},{"comment":"The paper should report 95% confidence intervals for all C-indices (e.g., via bootstrap) and a test of the difference between generated and real biomarkers.","section":"Section 4"},{"comment":"The GAN loss is written for a single discriminator, but the text mentions two discriminators D2d and D3d; the equation should include both terms explicitly.","section":"Appendix D.2, Eq. (3)"},{"comment":"The 3D-ConvGRU is described as the 'encoder of the neural ODE solver', but it is actually the encoder that produces the initial latent state; rephrase to avoid confusion.","section":"Sections 2 and D.3"},{"comment":"Provide more details on the test cohort: number of patients, number of scans per patient, and the distribution of time intervals between scans.","section":"Section 3, Datasets"},{"comment":"The phrase 'visual inspection by me' should be rewritten in a professional style, e.g., 'visual inspection by a trained researcher.'","section":"Appendix A"},{"comment":"Add annotations indicating which generated volumes correspond to which time points and which are interpolated versus extrapolated.","section":"Figure 3"},{"comment":"The reconstruction errors on the external Southampton test set are considerably higher than on the internal test set; this deserves a comment in the main text.","section":"Appendix D.4, Table 3"},{"comment":"The paper would benefit from a comparison with at least one existing longitudinal image synthesis method, such as the latent linear mixed-effects model of Sauty et al. or TR-GAN, to contextualize the results.","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"This manuscript is an early preprint that makes strong claims about personalized disease progression modelling from CT. The core weakness is the lack of baseline comparisons, which is fixable and should be addressed before publication. I would also encourage the authors to consider whether the survival analysis is adequately powered. The paper has merit, but in its current form it is more of a technical report than a validated clinical method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a sensible two-stage pipeline—3D VQ-GAN for compression, latent neural ODE for time—and as far as I know nobody has done exactly that for longitudinal lung CT. The authors use real IPF CTs (681 scans, 219 patients), register the lungs, and then validate generated scans by feeding codebook-frequency biomarkers into a Cox model. That survival check is genuinely external: outcomes are real, the C-index is computed on a test set, and it is not derived from the same loop that generates images. The paper is also honest about its main assumptions, including the deterministic ODE's shared-disease-dynamics limitation and the single-dataset validation. Credit where due.\n\nNow the soft spots. The central claim is under-supported. Table 1 only compares architectural variants; there is no no-change baseline, e.g. outputting the most recent real scan or its latent embedding unchanged. IPF changes are often subtle, and the reported extrapolation numbers are low (PSNR around 17.8 dB, SSIM around 0.49), so a trivial predictor could plausibly match or beat them. That matters because “capturing disease progression” is the headline, not just reconstruction quality. The survival analysis has the same issue: C-indices on generated scans (0.886 cross-sectional, 0.959 longitudinal) are compared with real scans, but not with biomarkers derived from the most recent actual scan. Since the model sees two prior scans, a high C-index could reflect baseline severity already present in the inputs rather than predicted temporal change. The authors do call the C-indices possibly overestimated, but they don't fix the comparison. Minor but relevant: no error bars, no code or data release, and the test set is small enough that a perfect 1.0 C-index appears in the real-scan column. The registration quality check is described as visual inspection by one non-clinical author, which is understandable but not a substitute for radiologist verification.\n\nWho is this for? Researchers working on longitudinal medical image generation. It is a reasonable proof-of-concept and the first instantiation of this particular combination for IPF. It deserves a serious referee rather than desk rejection, but it needs major revision before acceptance: add trivial baselines, report uncertainty, release code or at least precomputed generated scans and biomarker values, and ideally validate on a second cohort. I would not cite the quantitative results yet, but I would keep an eye on a revised version.","headline":"The model combination is new and the survival check is a real external validation, but the paper has not yet shown that its generated scans beat a no-change baseline, so the progression claims need work before publication.","tokens_in":15741,"tokens_out":1947,"would_cite":false,"duration_ms":22989,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Given two CT scans of an IPF patient, 4D-VQ-GAN synthesises realistic CT volumes at any requested time point, and biomarkers from those synthetic scans predict survival nearly as well as biomarkers from real scans.","keywords":["4D image synthesis","VQ-GAN","neural ordinary differential equations","disease progression modelling","CT imaging","idiopathic pulmonary fibrosis","survival analysis"],"falsifier":"A decisive test would be to take two scans from patients with documented different IPF subtypes or rates of decline and compare the model's generated third scans; if the synthetic future lungs are statistically indistinguishable across groups, the patient-specific progression claim fails. Quantitatively, the claim would be falsified if, on a held-out cohort stratified by subtype, the C-index of biomarkers from generated scans drops to chance (0.5) once subtype is accounted for.","tokens_in":14713,"feed_emoji":"🫁","tokens_out":8390,"duration_ms":78186,"temperature":0.7,"pith_summary":"Idiopathic pulmonary fibrosis is a progressive lung disease whose future course is hard to predict, and CT scans that could show that course are rarely taken because they involve radiation. This paper tries to establish that a two-stage generative model, 4D-VQ-GAN, can take two CT scans of the same IPF patient at different times and synthesise realistic CT volumes at any requested time point, producing an individualised trajectory of disease progression. The authors argue the generated scans are clinically meaningful: imaging biomarkers extracted from synthetic follow-up scans predict survival with a C-index of 0.886 for cross-sectional markers and 0.959 for one-year longitudinal changes, close to 0.929 and 1.0 obtained from real scans. If the claim holds, it would give clinicians a way to visualise probable disease evolution from sparse data and plan treatment earlier.","feed_headline":"Two CT scans are enough to synthesise any future IPF scan","feed_subtitle":"Synthetic follow-up CT volumes keep enough survival signal that biomarker risk scores nearly match real scans.","key_machinery":"The central object is the 4D-VQ-GAN, a two-stage architecture that couples a 3D-VQ-GAN with a latent neural ODE. The 3D-VQ-GAN learns a discrete codebook of volumetric image patterns and reconstructs CT volumes from quantised embeddings; the neural ODE $\\mathrm{d}h/\\mathrm{d}t = f_\\theta(h,t)$ then defines how those embeddings drift over time. A 3D-ConvGRU encoder compresses the observed embedding sequence into an initial state $h_{t_0}$, the ODE solver produces hidden states at arbitrary target times, and a convolutional projector converts the difference between successive hidden states into a difference map added to the previous embedding. This difference-plus-skip-connection design is what lets the model generate a continuous trajectory from only two observations.","core_discovery":"The central claim is that continuous disease progression in volumetric CT can be learned in a discrete latent space rather than modelled directly on voxels. A 3D-VQ-GAN first compresses each CT volume into a compact set of codebook indices whose entries correspond to local imaging patterns such as reticulation and ground-glass opacity. A second-stage temporal model, built around a neural ODE $f_\\theta$, evolves the latent embedding over time; given two observed scans, the ODE solver is run from the encoded initial state to produce embeddings at any queried timestep, and skip connections between consecutive embeddings let the model learn only the difference between adjacent time points. The frozen 3D-VQ-GAN decoder then turns those embeddings back into CT volumes. The paper demonstrates interpolation and extrapolation on IPF patients and validates the generated scans through Cox survival analysis, reporting that biomarkers from generated scans achieve a C-index of 0.886 (cross-sectional) and 0.959 (longitudinal change) versus 0.929 and 1.0 for real scans.","pith_inferences":["Beyond the paper, the same two-stage recipe could transfer to other diseases with sparse longitudinal imaging, such as multiple sclerosis or osteoarthritis, because it only requires a learnable discrete codebook and a smooth latent trajectory.","The paper's shared disease dynamic assumption is the obvious pressure point; conditioning the ODE on a per-patient latent code or subtype label would be a natural, untested upgrade that makes the personalised claim more literal.","A sharp testable extension would be to feed generated scans into a separate segmentation or staging model and measure whether performance improves; if the synthetic volumes carry the same prognostic texture as real ones, they could function as data augmentation in rare-disease cohorts."],"forward_implications":["Given two CT scans, the model outputs synthetic scans at any requested interval and horizon, covering both interpolation and extrapolation.","The best extrapolation configuration (ConvGRU encoder with embedding skip connections) reaches MSE 0.019, SSIM 0.489, and PSNR 17.816 on held-out patients.","Survival-relevant texture is preserved: cross-sectional biomarkers from generated third scans give C-index 0.886 versus 0.929 for real scans, and longitudinal biomarkers give 0.959 versus 1.0.","Synthetic longitudinal scans could be used to impute missing visits, augment sparse IPF datasets, and simulate treatment effects by comparing generated trajectories under different assumptions."],"supporting_citations":[{"why":"Supplies the 3D-VQ-GAN architecture and loss functions that the first stage adapts to compress CT volumes.","marker":"[16]"},{"why":"Introduces neural ODE solvers, the mechanism the temporal model uses to evolve latent embeddings continuously.","marker":"[4]"},{"why":"The original VQ-GAN whose encoder-decoder and adversarial loss the 3D version generalises.","marker":"[10]"},{"why":"Defines discrete codebook learning underlying the quantised embeddings that the temporal model operates on.","marker":"[54]"},{"why":"Provides the latent-decoder design of ODE plus difference maps that is adapted for predicting embedding dynamics.","marker":"[40]"},{"why":"Supplies the ConvGRU cell that is extended to 3D and used as the temporal encoder.","marker":"[1]"},{"why":"Cox proportional hazards model used to compute C-indices for survival validation.","marker":"[6]"},{"why":"CorrField registration method that aligns longitudinal lung scans so that only disease-related change remains.","marker":"[18]"},{"why":"Provides the faster CorrField implementation actually used to preprocess the CT series.","marker":"[17]"},{"why":"Establishes the latent-ODE formulation for irregularly sampled time series that the temporal model follows.","marker":"[45]"}],"fun_headline_variants":["Future IPF scans from just two CTs","AI predicts IPF progression with 4D VQ-GAN","Synthetic CTs match real scans for survival prediction","Two time points, all future IPF scans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"One fixed mathematical rule is assumed to describe how every patient's lungs change over time, and the paper itself says this is unlikely for a disease with distinct subtypes and progression patterns.","fun_headline_variants_meta":{"raw":{"variants":["Future IPF scans from just two CTs","AI predicts IPF progression with 4D VQ-GAN","Synthetic CTs match real scans for survival prediction","Two time points, all future IPF scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000794,"raw_usage":{"total_tokens":3539,"prompt_tokens":1029,"completion_tokens":2510,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":2446}},"tokens_in":645,"tokens_out":2510,"duration_ms":19183,"temperature":1.0,"reasoning_tokens":2446,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T18:14:40.361748+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test would be to take two scans from patients with documented different IPF subtypes or rates of decline and compare the model's generated third scans; if the synthetic future lungs are statistically indistinguishable across groups, the patient-specific progression claim fails. Quantitatively, the claim would be falsified if, on a held-out cohort stratified by subtype, the C-index of biomarkers from generated scans drops to chance (0.5) once subtype is accounted for.","supporting_citations":[{"cited_title":"Long video generation with time-agnostic VQGAN and time-sensitive transformer","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D-VQ-GAN architecture and loss functions that the first stage adapts to compress CT volumes."},{"cited_title":"Neural ordinary differential equations","cited_arxiv_id":null,"evidence_quote":"Introduces neural ODE solvers, the mechanism the temporal model uses to evolve latent embeddings continuously."},{"cited_title":"Taming transformers for high-resolution image synthesis","cited_arxiv_id":null,"evidence_quote":"The original VQ-GAN whose encoder-decoder and adversarial loss the 3D version generalises."},{"cited_title":"Neural discrete representation learning","cited_arxiv_id":null,"evidence_quote":"Defines discrete codebook learning underlying the quantised embeddings that the temporal model operates on."},{"cited_title":"Vid-ODE: Continuous-time video generation with neural ordinary differential equation","cited_arxiv_id":null,"evidence_quote":"Provides the latent-decoder design of ODE plus difference maps that is adapted for predicting embedding dynamics."},{"cited_title":"Regression models and life-tables","cited_arxiv_id":null,"evidence_quote":"Cox proportional hazards model used to compute C-indices for survival validation."},{"cited_title":"Estimating large lung motion in COPD patients by symmetric regularised correspondence fields","cited_arxiv_id":null,"evidence_quote":"CorrField registration method that aligns longitudinal lung scans so that only disease-related change remains."},{"cited_title":"GraphRegNet: Deep graph regularisation networks on sparse keypoints for dense registration of 3D lung CTs","cited_arxiv_id":null,"evidence_quote":"Provides the faster CorrField implementation actually used to preprocess the CT series."},{"cited_title":"Latent ordinary differential equations for irregularly- sampled time series","cited_arxiv_id":null,"evidence_quote":"Establishes the latent-ODE formulation for irregularly sampled time series that the temporal model follows."}],"review_version":1}