{"id":"f1961acc-31a5-438e-ab0a-ca0675113f28","arxiv_id":"2509.03141","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"TADM-3D predicts future brain MRIs from a baseline scan and time gap by diffusing the scan-to-scan residual, with brain-age regularization and bidirectional training, and reports gains over prior 3D baselines.","lead":"A new 3D diffusion model forecasts how a person's brain MRI will change over a chosen number of years by learning the difference between scans, then adding that difference to today's scan. It adds a brain-age check and trains the model to predict both future and past scans, reporting better volume predictions in Alzheimer's-relevant regions than prior baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BAE guidance relies on a frozen network fed by a jointly trained encoder and applied to synthetic scans; the paper never specifies the differentiable path or addresses miscalibration, so the core novelty may not work as claimed.","rationale":"The reader's weakest assumption identifiesthe BAE signal as the key uncertainty. My reading of the paper confirms that the BAE loss is a core novelty (it drives the temporal-awareness claim and the ablation shows a large drop without it), yet the text contains conflicting details about what is frozen and what is trained: Sec. 3.4 says BAE is fixed, Sec. 4.2 says the encoder is trained end-to-end. Because the BAE's input is the encoder's latent, a shifting encoder invalidates the BAE's calibration. Moreover, Algorithm 1's 'Generate predicted residual' step is unspecified; the differentiability of L_BAE depends entirely on this choice, and the paper never explains how gradients flow through the DDPM sampling. These are internal inconsistencies, not just disagreements with external consensus. They are load-bearing because if the BAE is miscalibrated on synthetic scans, Eq. (7) is not a valid measure of temporal consistency, and the improved performance attributed to the BAE may instead come from the residual-prediction formulation or the 3D architecture. The proposed concrete test directly measures the BAE's calibration on synthetic outputs and would settle whether the loss is meaningful. I do not escalate the verdict to REJECT because the paper is an empirical study and the results could still be reproducible; the lack of code prevents verification, but the concern warrants the CONDITIONAL verdict the reader gave. No ad hominem is intended; the issue is in the technical description.","tokens_in":13868,"tokens_out":5035,"duration_ms":60254,"concrete_test":"Using the trained TADM-3D and the frozen BAE, take a held-out set of test subjects and compute the BAE's predicted age gap bΔ for (a) real baseline/follow-up pairs and (b) real baseline / synthetic follow-up pairs produced by the model. Compare the mean absolute error |bΔ − Δ| between (a) and (b). If the error on synthetic pairs is substantially larger (e.g., >2x) or systematically biased (e.g., mean error shifts sign), the BAE is miscalibrated on generated data, and L_BAE is not providing the temporal guidance claimed. Additionally, report the gradient norm of L_BAE with respect to the DDPM parameters to confirm a non-trivial signal flows.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty is the BAE temporal regularizer (Eq. 7). Its soundness requires that the frozen BAE, trained on real scans, provides an accurate, differentiable estimate of the age gap between the baseline and the model-generated follow-up. Two unaddressed issues undermine this. First, Sec. 3.4 states the BAE is 'pre-trained on the training set, and its parameters remain fixed', but Sec. 4.2 states the encoder Φ is 'trained together with the diffusion model in an end-to-end manner'. Since the BAE consumes features from Φ (bΔ = Ψ(Φ(ÎTb)) − Ψ(Φ(ITa))), the BAE's input distribution drifts during training; its age predictions may become miscalibrated. Second, Algorithm 1 says 'Generate predicted residual bIΔ from the DDPM' without specifying whether this is a one-step denoising estimate (x̂0 from the current timestep), full reverse sampling, or DDIM. L_BAE requires a differentiable path through this generation; different choices yield very different gradient signals, and the one-step estimate from a partly noised residual may be unreliable for age prediction. The BAE is also applied to synthetic follow-ups (ÎTb = ITa + bIΔ) that differ in distribution from the real scans it was trained on. If the BAE is miscalibrated on such outputs, Eq. (7) does not measure temporal consistency but instead rewards age-typical artifacts. Since the ablation (Table 3) attributes a large performance drop to removing the BAE, the central claim depends on this mechanism working as described. The paper's Limitations section does not address this.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TADM-3D, a 3D denoising diffusion model that predicts brain MRI progression by learning voxel-wise residuals between baseline and follow-up scans. It conditions the diffusion process on the age gap between scans, patient metadata, and a baseline latent encoding, and introduces two temporal regularizers: a Brain-Age Estimator (BAE) loss (Eq. 7) that penalizes deviations between the predicted and true age gap, and a Back-In-Time Regularisation (BITR) that alternates forward and backward prediction during training. The method is evaluated on the OASIS-3 internal test set and on an external NACC test set, reporting MSE/SSIM and five regional volume errors against DaniNet, CounterSynth, and BrLP. Ablations isolate the contributions of age-gap conditioning, BAE, BITR, patient metadata, and 2.5D versus native 3D generation.","tokens_in":14207,"tokens_out":9078,"duration_ms":101075,"significance":"If the technical pipeline is made precise, the contribution is potentially valuable: learning 3D residual progression conditioned on age difference rather than absolute age is a sensible design, the external NACC evaluation is good practice, and the ablation study in Table 3 is informative. The main strengths are the 3D architecture, the residual formulation, and the explicit attempt to regularize temporal consistency with an age-estimation signal. However, the quantitative support is currently weakened by the absence of significance testing, by ambiguous implementation of the two central regularizers (BAE and BITR), and by some overstatement of external generalisation. These issues are fixable in revision, but they are load-bearing for the paper's main claims.","major_comments":[{"comment":"The BAE regularizer is the central novelty, but Eq. (7) is not well-defined as written. L_BAE requires a differentiable estimate of the generated residual ÎΔ; Algorithm 1 says only “Generate predicted residual bIΔ from the DDPM”, leaving open whether this is one-step denoising (x0 estimate), full reverse sampling, or DDIM. These choices give very different gradient signals. Additionally, Eq. (7) uses bΔ = Ψ(Φ(ÎTb)) − Ψ(Φ(ITa)), while §4.2 states Φ is trained end-to-end and §3.4 states the BAE parameters are frozen. If Φ is shared, the frozen BAE receives a drifting feature distribution and may be miscalibrated on synthetic scans. Since Table 3 shows removing BAE raises MSE from 0.004 to 0.009, this must be resolved: specify the differentiable path, clarify whether Φ is shared, and justify or calibrate the BAE on generated images.","section":"§3.4, Eq. (7), §4.2, Algorithm 1"},{"comment":"The BITR description and pseudocode are inconsistent. The text says the roles of ITa and ITb are randomly swapped, but Algorithm 1 keeps za = Φ(ITa) and computes ÎTb = ITa + ÎΔ, merely flipping “sign in Δ, A, and D”. A sign-flipped baseline age A and cognitive-status D do not implement a role swap; backward training should condition on Φ(ITb) and target residual ITa − ITb. As written, the pseudocode does not realize the proposed regularizer. Because Table 3 attributes a large drop to removing BITR (MSE 0.004 vs 0.008), this must be corrected or explicitly implemented as described.","section":"§3.5, Algorithm 1"},{"comment":"No statistical significance testing is provided, and several headline differences are within one standard deviation: internal MSE 0.004±0.001 vs BrLP 0.005±0.002; SSIM 0.902±0.014 vs 0.887±0.017. Paired per-subject tests (e.g., Wilcoxon signed-rank or bootstrap confidence intervals on paired differences) are required to support the claim that TADM-3D overcomes previous methods. The same applies to the ablation rows and to the cognitive-conditioning experiment in Table 4.","section":"Tables 1–3"},{"comment":"The external generalisation claim is overstated. On the NACC set, BrLP achieves higher SSIM (0.909±0.023 vs 0.902±0.017) and lower lateral-ventricle volume error (0.213±0.350 vs 0.235±0.200); TADM-3D is best on MSE and on four of five regions. The text should qualify the external performance explicitly and should not present the external results as uniform superiority without significance testing.","section":"§4.4, Table 2"}],"minor_comments":[{"comment":"“reducing the MSE and increasing the SSIM by +0.001 and +0.15” should read “reducing MSE by 0.001 and increasing SSIM by 0.015” (Table 1: 0.004 vs 0.005; 0.902 vs 0.887).","section":"§4.4"},{"comment":"The TADM-3D w/o BITR row reports “874±0.021” for SSIM; this should be “0.874±0.021”.","section":"Table 3"},{"comment":"The text says comparisons include SADM [8], but SADM results are absent from Tables 1–2. If SADM requires a longitudinal sequence and is therefore not directly comparable, state this explicitly.","section":"§4.4"},{"comment":"The notation B(0.5) is not defined (presumably Bernoulli). Also, “flip sign in D” is meaningless if D is categorical cognitive status; the pseudocode should be cleaned up.","section":"Algorithm 1"},{"comment":"No weight or scaling is specified between L_DM and L_BAE. Please report whether the two losses are weighted or normalized, since their scales differ substantially.","section":"Eq. (8)"},{"comment":"The cognitive-conditioning experiment would benefit from specifying the number of subjects, the paired setup, and significance testing; the current aggregate mean±std does not by itself support the conclusion that the model is not biased toward healthy ageing.","section":"§4.6, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of the journal and builds on a legitimate prior MICCAI paper. I believe the method is promising, but the two central regularizers (BAE and BITR) are underspecified and the statistical support for the headline claims is currently weak. I would be willing to review a revised version that resolves these points and, preferably, releases code."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: a legitimate, well-scoped extension of the authors' own MICCAI 2024 paper, now in 3D, with a bidirectional training regularizer (BITR), cognitive-status conditioning, and external validation. Not a breakthrough, but a competent empirical study that deserves peer review. The caveat: the central BAE loss has an underspecified differentiable path, and the reported gains are not tested for significance.\n\nWhat is actually new: the 3D residual diffusion conditioned on age gap, BITR, cognitive-status conditioning, and the NACC external test. The paper is honest that the core age-gap conditioning and BAE loss carry over from their earlier work. The ablations in Table 3 are clean and show each component earns its keep, especially BITR and BAE. External validation on NACC is good practice, and the limitations section honestly flags long-horizon degradation.\n\nThe soft spots: first, no significance testing anywhere. On the internal test, TADM-3D's MSE is 0.004±0.001 vs BrLP's 0.005±0.002 and SSIM 0.902±0.014 vs 0.887±0.017 — those overlap within one standard deviation. The hippocampus error difference is larger and suggestive, but still without paired error bars. Second, the BAE mechanism is not fully pinned down. The BAE is frozen and pre-trained on real scans, while the encoder it consumes features from is trained end-to-end; the input distribution to BAE drifts during training, and it is applied to synthetic follow-ups that may be out-of-distribution. The paper never says whether the 'predicted residual' used for L_BAE comes from one-step denoising or full reverse sampling. If BAE is miscalibrated on generated images, Eq. (7) rewards age-typical artifacts rather than patient-specific progression. This is a real concern, though not a demonstrated failure — the ablation shows BAE helps, so empirically it does something useful. Third, Algorithm 1's swap step says 'flip sign in A, D' but not in the scans; that looks like a typo. Fourth, it's not stated whether the baselines were retrained on the same data splits.\n\nWho is this for: people working on brain progression simulation or synthetic control arms. It's a modest incremental contribution with honest reporting and one genuinely useful idea (BITR), plus a cautionary lesson about using frozen age predictors on generated images. I'd send it to referees.","headline":"Legitimate incremental extension of the authors' own 2D work, with a clean ablation and external validation; the central BAE loss is underspecified and the headline metric gaps are within noise, but the paper deserves serious review.","tokens_in":14826,"tokens_out":2324,"would_cite":true,"duration_ms":24787,"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":"TADM-3D predicts future brain MRIs as a baseline plus a diffusion-generated age-gap residual, and reports lower errors than DaniNet, CounterSynth, and BrLP on OASIS-3 and most NACC regions.","keywords":["brain MRI","disease progression modeling","diffusion models","temporal regularization","age-gap conditioning","Alzheimer's disease","medical image synthesis","longitudinal imaging"],"falsifier":"For a fixed set of test subjects, run the BAE on real follow-up scans and on TADM-3D's predicted follow-ups from the same baselines, and compare the BAE's estimated age gaps against the true interval; if the BAE error on synthetic outputs is systematically different from its error on real scans, the L_BAE guidance is miscalibrated and the reported temporal gains need re-interpretation. Also, re-train with L_BAE computed through one-step DDIM denoising instead of whatever generation path Algorithm 1 uses, and check whether the MSE, SSIM, and volumetric gains persist.","tokens_in":13665,"feed_emoji":"🧠","tokens_out":4822,"duration_ms":50201,"temperature":0.7,"pith_summary":"The paper tries to establish that a 3D diffusion model that learns the residual intensity difference between a baseline and a follow-up brain MRI, rather than synthesizing the follow-up from scratch, can predict Alzheimer's-relevant brain progression more accurately than existing GAN- and diffusion-based simulators. The key temporal trick is conditioning on the age gap between scans instead of absolute age, and using a pre-trained brain-age estimator to penalize generated scans whose estimated age gap deviates from the requested one. A second trick trains the model backward as often as forward so it learns the direction of time. If correct, TADM-3D gives clinicians a sharper forecast of regional atrophy, especially hippocampus volume, and generalizes to an out-of-distribution cohort.","feed_headline":"Diffusion model forecasts brain MRIs from a single baseline scan","feed_subtitle":"TADM-3D lowers hippocampus volume error ~40% on OASIS-3 and holds up on NACC external data.","key_machinery":"The central object is the predicted residual image IΔa,b = ITb − ITa, generated by a DDPM and added to the baseline. The load-bearing identity is that temporal progression can be learned as an intensity residual, so the model never has to regenerate whole-brain anatomy. Temporal control comes from conditioning on the age gap Δa,b rather than target age, and from the BAE loss L_BAE = (Δ̂a,b − Δa,b)², where Δ̂a,b is estimated by the frozen brain-age estimator. BITR makes time reversible during training by swapping baseline and follow-up at each step with probability 0.5. The residual formulation also keeps predictions anchored to the patient's own baseline anatomy.","core_discovery":"TADM-3D predicts the follow-up MRI as baseline plus a generated residual. The residual is produced by a DDPM conditioned on the baseline's latent encoding, the requested age gap, baseline age, and cognitive status. A frozen brain-age estimator maps baseline and predicted scans through a shared encoder into an estimated age gap, and the squared deviation from the true gap is added to the standard denoising loss. During training the roles of baseline and follow-up are swapped with probability 0.5, so the same model must predict forward and backward in time. On the OASIS-3 internal test set the method reaches MSE 0.004 and SSIM 0.902, beating BrLP (0.005 and 0.887) and reducing hippocampus volu","pith_inferences":["The age-gap conditioning trick is a candidate remedy for age-imbalanced longitudinal datasets; a direct test would train on a deliberately age-skewed split and measure whether performance degrades less than with age-conditioned baselines.","BITR is a generic temporal regularizer that could be grafted onto other conditional generative pipelines; a cheap test is adding it to a latent diffusion baseline and checking temporal-consistency metrics without sacrificing image quality.","Because the BAE is frozen while the encoder it consumes is trained end-to-end, the BAE's input distribution drifts during training; monitoring BAE age-gap error on generated images would reveal whether L_BAE remains a trustworthy guide.","The residual formulation suggests a natural extension beyond T1-weighted brain MRI, e.g., predicting change in PET or other modalities as a residual shift, since static anatomy does not need to be regenerated."],"forward_implications":["Forecasting a future MRI at arbitrary time intervals from a single baseline scan becomes an input parameter, because the model conditions on the age gap rather than a fixed follow-up schedule.","Regional volumetric forecasts in AD-relevant structures improve, with hippocampus error roughly 40% lower than the strongest baseline on the internal test set.","The model transfers to an out-of-distribution cohort: on NACC it leads on MSE and 4 of 5 regional volume metrics.","Long-term prediction is the weak point: accuracy visibly drops for intervals beyond about 6.5 years, as the paper's own limitation example shows.","Cognitive-status conditioning carries real signal: feeding the wrong cognitive label degrades predictions, most of all in the hippocampus."],"supporting_citations":[{"why":"DaniNet is a 2D GAN baseline for brain scan simulation that TADM-3D must beat on image and volumetric metrics.","marker":"[2]"},{"why":"CounterSynth is a morphologically constrained 3D GAN baseline, providing the main 3D counterfactual comparison.","marker":"[3]"},{"why":"BrLP is the latent diffusion baseline that is the closest competitor on both internal and external test sets.","marker":"[11]"},{"why":"OASIS-3 supplies the longitudinal T1w MRI training and internal test data with cognitive status labels.","marker":"[14]"},{"why":"The DDPM framework provides the denoising diffusion machinery used to generate residual images.","marker":"[21]"},{"why":"The brain-age prediction approach underlies the Brain-Age Estimator used for the temporal-consistency loss.","marker":"[23]"},{"why":"The NACC dataset is the external evaluation set used to test out-of-distribution generalization.","marker":"[24]"},{"why":"SynthSeg is used to segment brain regions and compute the volumetric errors that drive the main comparisons.","marker":"[27]"},{"why":"The protocol for deliberately mislabeling cognitive status comes from this work, and is used to test whether the model tracks disease-specific progression.","marker":"[28]"}],"fun_headline_variants":["Brain MRI forecast: one scan, future predicted with 40% better accuracy","Diffusion model predicts brain aging, learns by reversing time","TADM-3D: bidirectional training sharpens brain progression predictions","From baseline to future: AI model cuts hippocampus error by 40%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the pre-trained brain-age estimator gives a trustworthy age-difference score when applied to synthetic diffusion outputs, and that the training loop can compute that score through a differentiable denoised residual—neither of which the paper verifies.","fun_headline_variants_meta":{"raw":{"variants":["Brain MRI forecast: one scan, future predicted with 40% better accuracy","Diffusion model predicts brain aging, learns by reversing time","TADM-3D: bidirectional training sharpens brain progression predictions","From baseline to future: AI model cuts hippocampus error by 40%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000518,"raw_usage":{"total_tokens":2415,"prompt_tokens":876,"completion_tokens":1539,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":1462}},"tokens_in":620,"tokens_out":1539,"duration_ms":10719,"temperature":1.0,"reasoning_tokens":1462,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:06:58.586623+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a fixed set of test subjects, run the BAE on real follow-up scans and on TADM-3D's predicted follow-ups from the same baselines, and compare the BAE's estimated age gaps against the true interval; if the BAE error on synthetic outputs is systematically different from its error on real scans, the L_BAE guidance is miscalibrated and the reported temporal gains need re-interpretation. Also, re-train with L_BAE computed through one-step DDIM denoising instead of whatever generation path Algorithm 1 uses, and check whether the MSE, SSIM, and volumetric gains persist.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DaniNet is a 2D GAN baseline for brain scan simulation that TADM-3D must beat on image and volumetric metrics."},{"cited_title":"Pombo, R","cited_arxiv_id":null,"evidence_quote":"CounterSynth is a morphologically constrained 3D GAN baseline, providing the main 3D counterfactual comparison."},{"cited_title":"Puglisi, D","cited_arxiv_id":null,"evidence_quote":"BrLP is the latent diffusion baseline that is the closest competitor on both internal and external test sets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"OASIS-3 supplies the longitudinal T1w MRI training and internal test data with cognitive status labels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The brain-age prediction approach underlies the Brain-Age Estimator used for the temporal-consistency loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The NACC dataset is the external evaluation set used to test out-of-distribution generalization."}],"review_version":1}