{"id":"771195a1-8bea-4d56-99b4-820b3806b3e9","arxiv_id":"2607.27537","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ProgFormer, a hierarchical voxel-space diffusion transformer with coarse-to-fine attention, improves longitudinal brain MRI prediction over latent and direct volumetric baselines on ADNI, AIBL, and OASIS.","lead":"This paper introduces ProgFormer, a computer model that predicts a person's future brain MRI from their past scans. It combines whole-brain and fine-detail pathways and reports better accuracy than previous methods on three Alzheimer's disease datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA hinges on undisclosed region weights λ_r in Eq. (21); ablation shows removing them drops PSNR more than the margin over baselines.","rationale":"I read the paper as proposing a hierarchy plus a region-weighted loss to improve longitudinal MRI prediction. The ablations support the importance of the two pathways, which is a genuine strength. However, the recorded SOTA margins are small, and the only ablation that removes a component whose parameters are entirely undisclosed produces a performance drop larger than the SOTA margin. Thus the most load-bearing assumption is that the λ_r weights are well-chosen and transferable; the paper gives no evidence for this. A simple equal-weight retraining experiment would directly test whether the claim survives without the hand-chosen weights. If it does not, the paper's contribution is the weighting scheme, not the architecture, and the claim requires major revision. If it does, then the concern is resolved and the paper's framing is acceptable. I agree with the reader's assessment that the paper should be conditional pending this disclosure.","tokens_in":13620,"tokens_out":7296,"duration_ms":77977,"concrete_test":"Run the ADNI pairwise experiment with all λ_r = 1 (equal region weights) and otherwise the same training pipeline and hyperparameters; compare PSNR/SSIM against CounterSynth and BrLP in Table 1. If the PSNR margin is absent or reversed, the SOTA claim depends on the undisclosed weights and the paper should be revised to provide them or temper its claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim rests on the weighted flow-matching objective (Eq. 21). Section 3.7 states that each SynthSeg label r is assigned a predefined weight λ_r, but the values are never reported. The ablation in Table 2 ('w/o Region Weights') shows the full model loses 1.64 dB PSNR (26.39→24.75) and 5.96 SSIM points in the pairwise ADNI setting, and 3.45 dB (27.90→24.45) and 7.36 SSIM points in the trajectory setting. Meanwhile, ProgFormer-S's PSNR margin over the strongest baseline (CounterSynth) is only 0.20–0.28 dB across cohorts. If the region weights are removed, ProgFormer-S falls below CounterSynth on ADNI PSNR and loses its SSIM lead. The SOTA claim is therefore contingent on an unreported, manually chosen weighting scheme. This is not a minor implementation detail: the ablation indicates the main result is not an inherent property of the hierarchical architecture. Without the λ_r values (or code), the results are not reproducible, and if the weights were tuned on these benchmarks, they may not transfer. The absence of significance tests — the pairwise PSNR differences are well within one standard deviation — further weakens the claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ProgFormer, a hierarchical voxel-space diffusion transformer for longitudinal brain MRI prediction. The model combines a coarse patch-level pathway that performs spatial and causal temporal attention across observed scans with a fine voxel-level pathway that refines local structure under coarse-to-fine cross-attention. Both pathways jointly estimate a velocity field via conditional flow matching, and the future scan is obtained by integrating this field with Euler steps, directly in voxel space and without a separately learned autoencoder. The method is evaluated on ADNI, AIBL, and OASIS under pairwise and trajectory settings, reporting the best PSNR on all three cohorts for the sequential variant, with competitive SSIM and regional MAE against BrLP, CounterSynth, TADM-3D, and SADM.","tokens_in":13993,"tokens_out":4913,"duration_ms":46320,"significance":"If the results hold, the paper makes a useful contribution to longitudinal brain MRI prediction by demonstrating that direct voxel-space diffusion modeling can be made effective through a hierarchical coarse-to-fine architecture, avoiding the information loss of latent autoencoders. The flow-matching formulation is standard and the ablations (Table 2) support the importance of the coarse pathway and coarse-to-fine cross-attention. The evaluation across three cohorts and two settings is thorough and includes clear preprocessing and implementation details. However, the central state-of-the-art claim is currently fragile: the reported advantage over baselines is small (0.20–0.28 dB PSNR), no significance tests are provided, and the method's performance depends strongly on an unreported, manually chosen regional weighting in the training loss. These issues must be resolved before the claim can be accepted.","major_comments":[{"comment":"The regional weights λ_r in the weighted flow-matching objective are never reported. The ablation in Table 2 shows that removing them degrades PSNR by 1.64 dB (pairwise) and 3.45 dB (trajectory) on ADNI, while the claimed margins over the strongest baselines are only 0.20–0.28 dB. This implies that the state-of-the-art result is not an inherent property of the hierarchical architecture but is contingent on a hand-chosen weighting scheme whose values, and tuning procedure, are undisclosed. The authors must report the λ_r values, provide code or a sensitivity analysis, and present results with uniform weights. Additionally, because the same anatomical regions emphasized in the loss (hippocampus, ventricles, etc.) are used to compute R-MAE, the regional metric is partially aligned with the training objective; this potential circularity should be explicitly discussed.","section":"§3.7, Eq. (21) and §4.3, Table 2"},{"comment":"No statistical significance testing is reported. The PSNR differences between ProgFormer-S and the strongest baseline are within one standard deviation on all three cohorts (e.g., 26.39±1.68 vs. 26.19±2.64 on ADNI). Without paired significance tests (e.g., Wilcoxon signed-rank or bootstrap confidence intervals over test subjects) for PSNR, SSIM, and R-MAE, the claim that ProgFormer achieves state-of-the-art performance is not supported. The authors should add such tests for all pairwise and trajectory comparisons.","section":"§4.1, Table 1"}],"minor_comments":[{"comment":"The condition vector in Equation (22) is written as c_{T+1}, but the covariate vector is defined as c_T in Eq. (1) and the text says c_T remains fixed during inference. This is likely a typo and should be corrected.","section":"Eq. (22)"},{"comment":"The ablation 'w/o Coarse Pathway' is difficult to interpret because the fine pathway in the full model receives coarse representations via cross-attention. With the coarse pathway removed, the fine pathway has no coarse context; the resulting collapse (PSNR 7.23) is expected and not informative. Please clarify what variant is used in this ablation.","section":"§4.3, Table 2"},{"comment":"The row label 'T SADM' is not defined in the table caption or text. It is later explained that SADM is evaluated in single-input (I=1) and multiple-input (I=T) configurations. Please label this row clearly (e.g., 'SADM (I=T)') and, for completeness, report the single-input SADM trajectory result as well.","section":"Table 1"},{"comment":"The sentence 'Compared with the strongest I=1 result for each metric, using the observed history improves PSNR by 1.28 on ADNI, 0.17 on AIBL, and 0.79 on OASIS' is ambiguous about the comparison basis. From Table 1, the difference between ProgFormer-S pairwise and trajectory on ADNI is 1.51 dB, not 1.28. Please reconcile these numbers and specify exactly which baseline and protocol (pairwise vs I=1 trajectory) are being compared.","section":"§4.2, Trajectory prediction"},{"comment":"The colorbar for the absolute error maps is not shown, making it difficult to gauge the magnitude of the differences. Consider adding a color scale or normalizing the maps consistently.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea and experiments are plausible, and the architecture is well motivated. However, the missing λ_r values and the lack of significance testing are load-bearing issues for the state-of-the-art claim. These are fixable within the manuscript's scope, so major revision is appropriate rather than rejection. I would also recommend the editor ensure that the authors provide code or detailed supplementary material for the loss weights and ablations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look: ProgFormer is a well-built hierarchical voxel-space diffusion transformer for longitudinal brain MRI. The architecture is new — coarse patch-level spatio-temporal attention, fine voxel-level refinement with coarse-to-fine cross-attention, all in voxel space with flow matching. The ablations are clean and support the design. But the headline SOTA result is weaker than it looks: the margins over baselines are sub-dB, there are no significance tests, and the loss depends on undisclosed regional weights λ_r whose removal drops PSNR far more than those margins.\n\nThe paper does several things well. The two-pathway design is clearly motivated and the ablations (Table 2) show that removing either pathway, the cross-attention, or the region weights degrades performance substantially. The visualizations in Figure 6 demonstrate that the fine pathway concentrates its changes on hippocampus, ventricles, and cortex, which is the intended behavior. The evaluation on three cohorts with both pairwise and trajectory settings is thorough.\n\nThe soft spots are real. The pairwise PSNR gains over CounterSynth are 0.20–0.28 dB, while the reported standard deviations are 1.5–2.6 dB. Without paired significance tests, the SOTA claim is not established. More importantly, the training loss in Eq. (21) uses SynthSeg-derived region weights λ_r that are never reported. Removing those weights in the ablation drops ADNI PSNR by 1.64 dB (pairwise) and 3.45 dB (trajectory) — far exceeding the margin over baselines. So the advantage is contingent on an unreported, hand-set weighting scheme. There is also some circularity in the evaluation: R-MAE is computed over the same regions (hippocampus, amygdala, ventricles, etc.) that the loss weights emphasize, which could inflate the apparent regional benefit. No code or data are released, so the results are not reproducible as-is.\n\nNone of this makes the method invalid. The hierarchy is sensible, and the ablations support its contribution. But the paper needs to disclose the λ_r values, add significance tests or confidence intervals, and ideally release code before the SOTA claim can be taken at face value.\n\nThis paper is for researchers in longitudinal medical imaging and voxel-space generative models. It deserves a serious referee — the architecture is worth exploring — but it needs revision. I would bring it to a reading group to discuss the design, though I wouldn't cite it in my own work until the weights are available.","headline":"Novel voxel-space hierarchy, but the SOTA claim is undercut by undisclosed loss weights and missing significance tests.","tokens_in":14427,"tokens_out":3691,"would_cite":false,"duration_ms":39312,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Predicting future brain MRI directly in voxel space with a coarse-to-fine transformer outperforms latent-space and single-pathway baselines, the paper claims.","keywords":["longitudinal brain MRI","diffusion transformer","flow matching","voxel-space generation","Alzheimer's disease","medical image synthesis","coarse-to-fine","trajectory prediction"],"falsifier":"Run ProgFormer-S on the ADNI pairwise split with all regional weights λ_r set to 1 and report PSNR/SSIM; if the resulting drop is substantially smaller than the 1.64 dB and 5.96 SSIM points reported in Table 2, the strong dependence on the manual weighting—and hence the claim that the hierarchical architecture alone delivers the gain—is falsified.","tokens_in":13540,"feed_emoji":"🧠","tokens_out":6695,"duration_ms":62344,"temperature":0.7,"pith_summary":"The paper attempts to establish that longitudinal brain MRI prediction is better done directly in voxel space, avoiding the information loss of latent autoencoders, but with a two-pathway design that separates global structural modelling from local refinement. It introduces ProgFormer, where a coarse pathway reasons over 3D patch tokens with spatial and causal temporal attention to set the whole-brain structure, and a fine pathway refines individual voxels within each patch using the coarse representation as spatio-temporal grounding. The two pathways jointly predict a velocity field via conditional flow matching, and the future scan is generated by Euler integration from Gaussian noise. On ADNI, AIBL, and OASIS, the paper reports the best pairwise PSNR on all three cohorts and the best regional accuracy for the interleaved variant, along with trajectory improvements from observed history. A sympathetic reader would care because the method addresses a real trade-off: preserving stable anatomy while capturing subtle, localized disease-driven changes.","feed_headline":"Voxel-space transformer beats latent rivals for brain MRI forecast","feed_subtitle":"It separates global structure from local refinement and reports the top PSNR on ADNI, AIBL, and OASIS.","key_machinery":"The central mechanism is the hierarchical velocity-field decomposition: a single network predicts a voxel-space velocity νθ = νcoarse + νfine through conditional flow matching, where the coarse pathway produces a patch-level velocity from spatio-temporal token attention and the fine pathway produces a voxel-level velocity guided by cross-attention to the coarse features. This avoids a separately learned autoencoder and lets global structure and local change be modelled by separate pathways within one end-to-end network. The weighting of the loss by SynthSeg regional labels is a second load-bearing component that directs training toward progression-relevant tissue.","core_discovery":"The central claim is that hierarchical voxel-space diffusion modelling can outperform both latent-space and unified voxel-space predictors for longitudinal brain MRI. ProgFormer's coarse pathway operates on non-overlapping p×p×p patch tokens, applying spatial attention within each scan and causal temporal attention across observed scans to encode whole-brain structure and longitudinal context. The fine pathway treats each voxel in the flow state as a token, applies voxel-level self-attention within each patch, and uses coarse-to-fine cross-attention to condition voxel refinement on the patch-level representation. The predicted velocity is the sum of the coarse and fine velocities, and traini","pith_inferences":["The unreported values of the regional loss weights λ_r are a reproducibility gap: if they were tuned on one cohort and then applied to others, part of the reported advantage could be dataset-specific; a uniform-weight ablation is the natural check.","The causal temporal attention receives scans in chronological order with zero-padding to a common length; a natural next experiment is varying the number of observed scans to see whether prediction error saturates or the model learns to down-weight old scans.","The same two-pathway recipe could transfer to other longitudinal volumetric imaging tasks—such as cardiac MRI or tumour progression—where the majority of the volume is static and changes are localized; the paper does not test this, but the mechanism is architecture-agnostic.","Because the fine pathway operates on individual patches with coarse-to-fine attention but no cross-patch voxel attention, the model may be limited in refining structures that span patch boundaries; increasing patch size or adding fine-level cross-patch attention is a testable modification."],"forward_implications":["The reported pairwise PSNR gains (0.20–0.28 dB over the strongest baseline) on ADNI, AIBL, and OASIS would establish ProgFormer-S as the best single-scan predictor among the compared methods.","Using the full observed history in the trajectory setting improves PSNR by up to 1.28 dB on ADNI over single-scan input, showing that the causal temporal attention is extracting useful longitudinal signal.","The ablation study attributes the bulk of the performance to the coarse pathway: removing it collapses PSNR from roughly 26–28 dB to about 7 dB, confirming that patch-level spatio-temporal reasoning is the backbone of the method.","The fine pathway and coarse-to-fine cross-attention contribute 1.4–1.8 dB and 2–8 SSIM points, and their contribution grows under trajectory input, indicating that coarse grounding matters more when history is available.","The interleaved variant achieves the lowest regional MAE on all three cohorts, suggesting a trade-off between image-level fidelity (sequential) and regional structure (interleaved)."],"fun_headline_variants":["Voxel-space ProgFormer beats latent methods for brain MRI","Hierarchical voxel diffusion sharpens longitudinal MRI prediction","Coarse-to-fine voxel transformer outdoes latent rivals in MRI","ProgFormer splits brain prediction into global and local refinements"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method's advantage relies on hand-assigned per-region loss weights λ_r, derived from SynthSeg labels, whose specific values are not reported; if these weights are not reproducible or transferable to other cohorts, the reported performance gains may not hold.","fun_headline_variants_meta":{"raw":{"variants":["Voxel-space ProgFormer beats latent methods for brain MRI","Hierarchical voxel diffusion sharpens longitudinal MRI prediction","Coarse-to-fine voxel transformer outdoes latent rivals in MRI","ProgFormer splits brain prediction into global and local refinements"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00021,"raw_usage":{"total_tokens":1263,"prompt_tokens":777,"completion_tokens":486,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":415}},"tokens_in":521,"tokens_out":486,"duration_ms":5527,"temperature":1.0,"reasoning_tokens":415,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T06:02:02.045357+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ProgFormer-S on the ADNI pairwise split with all regional weights λ_r set to 1 and report PSNR/SSIM; if the resulting drop is substantially smaller than the 1.64 dB and 5.96 SSIM points reported in Table 2, the strong dependence on the manual weighting—and hence the claim that the hierarchical architecture alone delivers the gain—is falsified.","supporting_citations":[],"review_version":1}