{"id":"c369d408-fe60-4426-9752-4b2280095cd3","arxiv_id":"2508.21580","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Temporal Flow Matching applies flow matching to longitudinal medical imaging, modeling the difference between context and target images, and exceeds the last-context-image baseline on three datasets.","lead":"A new generative model called Temporal Flow Matching predicts future medical scans from a patient's past scans by learning to interpolate each past image toward the target image. It outperforms a simple copy-paste baseline and natural-image video prediction methods on three public datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significance testing: TFM's edge over LCI is within run-to-run variability on key metrics, undermining 'consistently surpasses'.","rationale":"The reader's verdict (CONDITIONAL) identifies generalization and fixed mask seed as the weakest assumption, but my concern is more specific: even on the reported test sets, the improvements over LCI may not be statistically significant. The paper lacks significance testing, and the effect sizes are small relative to run-to-run variance. This directly threatens the abstract's claim of 'consistently surpasses' and 'new state-of-the-art.' A concrete statistical analysis (e.g., paired bootstrap over test subjects) would settle this. The method itself is sensible and the comparison to natural-imaging baselines is clearly favorable, so the paper has merit, but the primary claim needs stronger statistical backing. Therefore, the verdict remains CONDITIONAL, contingent on adding significance tests and/or demonstrating that the improvements withstand multiple mask seeds.","tokens_in":12824,"tokens_out":6157,"duration_ms":69623,"concrete_test":"Compute a paired bootstrap or permutation test over the test subjects for each dataset and each metric (NRMSE, SSIM, PSNR), using the per-subject errors for TFM and LCI. Report 95% confidence intervals for the mean difference. If the CI includes zero for any dataset/metric, the claim 'consistently surpasses' is not supported. Additionally, rerun the Lumiere evaluation with 10 different mask seeds to assess sensitivity.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that TFM 'consistently surpasses' LCI is not supported by the reported statistics. Across the three datasets, the improvements over LCI are modest (NRMSE reductions ≈0.016–0.016, SSIM gains up to 2.0 points, PSNR gains ≈2–2.6 dB), while the standard deviations over the three runs are comparable to or larger than these differences (e.g., ACDC NRMSE 0.040±0.012 vs LCI 0.056; Lumiere SSIM 89.7±1.2 vs 89.3). The paper reports no confidence intervals, p-values, or per-subject error distributions. With test sets of only 14–50 subjects and three runs, a one-sample t-test on the run-level means shows that several comparisons, including ACDC NRMSE and Lumiere SSIM, would not reach significance; even a paired test on individual subjects is likely to yield broad CIs. Since the SOTA claim depends on a consistent, significant advantage over LCI, the current evidence is insufficient. The 'Difference Modeling' framing does not resolve this, as a residual U-Net with 1-step Euler is nearly as good (Table 4: 1 NFE SSIM 0.9566 vs 10 NFE 0.9599).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Temporal Flow Matching (TFM), a flow-matching framework for predicting a target 3D medical image from a sequence of irregularly sampled historical scans. The method pads all context images to the same temporal dimension, defines a linear interpolation between the sparsity-filled context sequence and the target repeated T times, and trains a U-Net to regress the constant velocity field (the residual to the target). Inference integrates this field with an ODE solver. Experiments on ACDC, ISLES, and Lumiere compare TFM with LCI, ConvLSTM, SimVP, and ViViT, reporting that TFM outperforms all baselines, with ablations on sparsity filling and number of integration steps.","tokens_in":13199,"tokens_out":6208,"duration_ms":72457,"significance":"The central formulation is mathematically sound and the experimental setup is described in enough detail to be reproducible. The paper draws attention to a genuine issue in longitudinal medical imaging: static anatomy dominates pixel-wise metrics, and a simple LCI baseline is very strong. TFM's difference modeling is a clean way to turn that observation into a training objective. If the empirical claims hold up, TFM would provide a useful and computationally efficient baseline for 4D medical image prediction. However, the paper currently overstates the evidence: the reported margins over LCI are often within run-to-run variability, and at least one table entry appears to contain a copy-paste error. The core idea is promising but needs stronger empirical validation.","major_comments":[{"comment":"The claim that TFM 'consistently surpasses' LCI is not supported by the reported statistics. On ACDC, NRMSE is 0.040±0.012 vs LCI 0.056; on Lumiere, SSIM is 89.7±1.2 vs 89.3. The standard deviations are comparable to or larger than the reported improvements, and no p-values, confidence intervals, or per-subject error distributions are given. With only three runs and test sets of 14-50 subjects, a paired significance test is needed to establish that the advantages are not due to chance. Additionally, the protocol fixes one mask seed per split, so all results are conditional on a single masking realization. Please add per-subject paired metrics with tests and CIs, or moderate the 'state-of-the-art' claim.","section":"Section 3.2 / Table 2"},{"comment":"The SimVP row for ISLES is identical to the SimVP row for ACDC (NRMSE 0.124, SSIM 52.8, PSNR 21.21), which is implausible for two different datasets. This appears to be a copy-paste error. If the ISLES SimVP numbers are not correct, the comparison of TFM against SimVP on ISLES is invalid and the benchmark conclusions are undermined. Please correct the table and re-run the affected analyses.","section":"Table 2, ISLES SimVP row"},{"comment":"The baselines from natural imaging (SimVP, ConvLSTM, ViViT) are trained with a simple L2 loss for 500 epochs without any reported hyperparameter search or convergence analysis. These methods are known to require substantial tuning and longer training schedules; their poor performance (e.g., SSIM 30-50 on ACDC) may reflect undertraining rather than an inherent limitation. To make the 'consistently surpasses spatio-temporal methods from natural imaging' claim credible, please report training/validation curves, show that baselines have converged, or provide evidence of a hyperparameter search. At minimum, discuss this limitation explicitly.","section":"Section 3.2 / Baseline training"}],"minor_comments":[{"comment":"Section 2.1 states that missing context images are set to 0, but Section 2.4 replaces them with sparsity filling. Clarify the order: are zero-filled inputs used at any stage before sparsity filling, or is sparsity filling applied before the model sees the data?","section":"Section 2.1 / Section 2.4"},{"comment":"The NFE ablation reports only SSIM. Since the paper's main comparisons use NRMSE and PSNR as well, reporting those metrics would help judge whether the choice of 10 integration steps is robust across metrics.","section":"Table 4"},{"comment":"The term 'Dimension Padding' is used without definition. Please define it explicitly and distinguish it from 'Temporal Pooling' with a formula or diagram.","section":"Section 2.3"},{"comment":"The text mentions 'Runge-Kutta integration' in the results discussion, while Algorithm 1 only illustrates Euler integration. Please state which solver was actually used for the main results and in Table 4.","section":"Section 4 / Algorithm 1"},{"comment":"The x-axis label says 'total number of masked frames' but the caption and text describe two masking protocols ('1 → T' and 'T → 1'). Clarify the axis and the meaning of the two curves.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising core but the empirical evidence is not yet at the level claimed. The identical SimVP rows in Table 2 suggest a data-handling error that must be corrected before any acceptance decision. I recommend requiring per-subject significance testing and a clarification of the baseline training protocol. If these issues are addressed, the paper could be a solid contribution to the field."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is sound. The authors take flow matching and adapt it to irregularly sampled 3D medical time series by padding the context dimension, filling missing frames with the nearest available image, and letting the target be replicated across the same dimension. That last move makes the velocity field literally the difference between context and target, which they call Difference Modeling. It is a simple formulation, but it is well matched to the problem: in these datasets temporal changes are small, and the LCI baseline is strong. The sparsity filling trick is straightforward and clearly helps in the ablations. The fixed validation mask protocol is a thoughtful response to a real pitfall with small validation sets. The LCI+FM experiment, where single-channel FM collapses, is genuinely informative and not something a lazy author would include.\n\nThe soft spots are mostly empirical. The headline claim that TFM \"consistently surpasses\" LCI is stronger than the numbers justify. On ACDC the NRMSE gap is 0.016 with a standard deviation of 0.012 across three runs; on Lumiere the SSIM gap is 0.4 points with a standard deviation of 1.2. No significance tests, no per-subject error distributions, no confidence intervals. With test sets of 14 to 50 subjects, the evidence for a real advantage is suggestive, not conclusive. The baselines are also weak: SimVP, ConvLSTM, and ViViT are natural-image methods, dropped into 3D without much adaptation, and they perform far below LCI. The paper lists medical baselines like NODER and ImageFlowNet in Table 1 but never compares against them, so \"first comprehensive benchmark\" is an overstatement.\n\nThere is also a likely data error. In Table 2, the SimVP row for ISLES is identical to the SimVP row for ACDC (NRMSE 0.124, SSIM 52.8, PSNR 21.21, same standard deviations). That looks like a copy-paste mistake, and it needs to be fixed before the paper is taken at face value.\n\nThat said, the method is clearly explained, the ablations are useful, and the limitations section is honest about data scarcity and resolution issues. The central idea is not novel enough to change the field, but it is a solid engineering contribution that could become a common baseline for medical trajectory prediction.\n\nThis paper deserves a serious referee. It needs a revision that adds significance testing or per-subject analysis, fixes the table, and ideally compares with at least one medical imaging baseline. I would read the revised version carefully, and I would probably cite it if I worked on longitudinal medical images.","headline":"A sensible, well-motivated adaptation of flow matching to longitudinal medical imaging, but the claimed consistent superiority over the Last Context Image baseline is not supported by the reported statistics, and there is a likely typo in the results table.","tokens_in":13602,"tokens_out":1829,"would_cite":true,"duration_ms":22151,"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":"Temporal Flow Matching predicts future medical images by learning a velocity field between context scans and a repeated target scan, modeling only the changes; it outperforms the last-context-image baseline on three longitudinal datasets.","keywords":["temporal flow matching","generative modeling","longitudinal medical imaging","4D image prediction","difference modeling","irregular time series","flow matching","spatio-temporal prediction"],"falsifier":"Re-run the three benchmarks under several different fixed validation-mask seeds and check whether TFM's margin over LCI persists in every rerun; if the advantage shrinks or reverses when the masks change, the claimed general superiority is a protocol artifact. A stronger test is an external longitudinal cohort acquired at a different site or scanner, since the velocity field must extrapolate to patients and acquisition conditions never seen in training.","tokens_in":12786,"feed_emoji":"🩻","tokens_out":7728,"duration_ms":75705,"temperature":0.7,"pith_summary":"Spatio-temporal prediction in medical imaging faces a strong trivial baseline: because anatomy changes slowly, simply returning the most recent scan often beats sophisticated video-prediction models. This paper introduces Temporal Flow Matching (TFM), a generative method that trains a network to predict the velocity field that transports a patient's past scans into a future scan, so the learned target is the difference between context and target rather than the full image. On three public longitudinal datasets—cardiac cine-MRI, stroke perfusion CT, and glioblastoma MRI—TFM reports better NRMSE/SSIM/PSNR than the last-context-image baseline and outperforms natural-image spatio-temporal baselines (ConvLSTM, SimVP, ViViT). The authors argue this establishes a new baseline for 4D medical image prediction and that difference modeling explains the gain, since pixel metrics are dominated by static anatomy.","feed_headline":"Modeling temporal differences tops last-scan baseline in 4D imaging","feed_subtitle":"A flow-matching model learns only what changes between scans, improving cardiac, stroke, and tumor forecasts.","key_machinery":"The central identity is the flow-matching velocity under linear interpolation: with X0=I′ (the filled context sequence) and X1=[Itarget,…,Itarget] (T copies), the interpolation Xτ=(1−τ)X0+τX1 has derivative dXτ/dτ=X1−X0=Itarget−I′. The network vθ(Xτ,τ) is trained to predict this velocity, which is exactly the spatio-temporal difference; the paper calls this Difference Modeling and notes it is a pure change of output space, imposing no architectural constraint. Inference integrates the learned field from τ=0 to 1, and the result is reduced by taking the last or mean time channel. Sparsity Filling—replacing missing context frames with the nearest available scan—keeps the velocity fields homoge","core_discovery":"The paper claims that a flow-matching model operating on the stacked sequence of context images, with the target image repeated T times, can learn a transport whose velocity is exactly the temporal difference Itarget−I′; training the network to predict this difference lets the model focus on changes rather than static content. This extends flow matching to irregularly sampled 3D time series via dimension padding and sparsity filling. Empirically the paper finds that TFM consistently surpasses both the last-context-image heuristic and spatio-temporal methods from natural imaging on ACDC, ISLES, and Lumiere, including on the hard small-cohort Lumiere dataset, and that the LCI+FM variant (same","pith_inferences":["Editorial inference: If difference modeling is the active ingredient, then change-focused evaluation (ROI metrics, perceptual metrics on temporal residuals) should reveal larger gains than pixel-level NRMSE/PSNR, which are dominated by unchanged anatomy.","Editorial inference: The formulation could be extended to continuous clinical time by replacing the abstract interpolation step τ with the actual elapsed time between scans; that would make the model directly usable for predicting a scan at any requested follow-up date.","Editorial inference: The fixed-mask validation protocol the authors introduce is itself a transferable practice for small medical datasets: masking realizations should be frozen before model selection, otherwise the trivial baseline's score fluctuates and 'best epoch' becomes arbitrary.","Editorial inference: Since TFM showed the largest absolute gains on the smallest cohort (Lumiere), a plausible testable hypothesis is that difference modeling degrades gracefully with training set size; testing on progressively smaller subsets of a large longitudinal dataset would quantify that."],"forward_implications":["A single 3D UNet trained end-to-end can predict full-resolution future volumes from multiple prior scans, without a separate temporal encoder or latent compression.","TFM inherits a cheap lower bound: if the context is uninformative, predicting the last available scan is a special case of the flow, so errors do not explode beyond the trivial baseline.","The method tolerates missing and irregularly spaced acquisitions: random masking during training plus sparsity filling keeps performance stable, and masking early context frames at inference barely affects quality.","Removing sparsity filling degrades NRMSE from 0.0261 to 0.0444 on the ACDC validation set, so the filling strategy is load-bearing for the reported results.","Aggregation by mean or by the last predicted time channel gives equivalent results, and a lightweight no-attention UNet performs nearly as well, suggesting the framework is not tied to a specific backbone."],"supporting_citations":[{"why":"Supplies the flow-matching training objective and ODE formulation that TFM extends to temporal sequences.","marker":"[9]"},{"why":"SimVP is a natural-image video prediction baseline and provides the dimension-padding idea; TFM is compared against it on all datasets.","marker":"[6]"},{"why":"ConvLSTM is a recurrent spatio-temporal baseline that TFM is benchmarked against.","marker":"[18]"},{"why":"Provides the ViViT variant used as the transformer-based baseline for fair comparison.","marker":"[22]"},{"why":"ACDC cardiac cine-MRI is one of the three benchmark datasets used to establish the claims.","marker":"[3]"},{"why":"ISLES perfusion-CT stroke data is one of the three benchmark datasets.","marker":"[17]"},{"why":"Lumiere longitudinal glioblastoma MRI is the small-cohort benchmark where baselines collapse and TFM still beats LCI.","marker":"[19]"},{"why":"Supplies the UNet backbone with cross-attention used for the flow network in all TFM experiments.","marker":"[21]"}],"fun_headline_variants":["Flow matching learns only what changes in 4D scans","Modeling temporal differences improves 4D medical forecasting","Learn the delta, not the scan: flow matching for 4D","Temporal flow model beats last-scan baseline in 4D imaging","TFM: predicting 4D trajectories beats last-context baseline"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that a velocity field trained on small, heterogeneous patient cohorts—as few as 48 training cases on Lumiere—generalizes to unseen patients, and that the edge over returning the last scan is not an artifact of the particular fixed masking seed and validation split.","fun_headline_variants_meta":{"raw":{"variants":["Flow matching learns only what changes in 4D scans","Modeling temporal differences improves 4D medical forecasting","Learn the delta, not the scan: flow matching for 4D","Temporal flow model beats last-scan baseline in 4D imaging","TFM: predicting 4D trajectories beats last-context baseline"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000828,"raw_usage":{"total_tokens":3439,"prompt_tokens":710,"completion_tokens":2729,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":2641}},"tokens_in":454,"tokens_out":2729,"duration_ms":20416,"temperature":1.0,"reasoning_tokens":2641,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:09:31.267769+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the three benchmarks under several different fixed validation-mask seeds and check whether TFM's margin over LCI persists in every rerun; if the advantage shrinks or reverses when the masks change, the claimed general superiority is a protocol artifact. A stronger test is an external longitudinal cohort acquired at a different site or scanner, since the velocity field must extrapolate to patients and acquisition conditions never seen in training.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the flow-matching training objective and ODE formulation that TFM extends to temporal sequences."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SimVP is a natural-image video prediction baseline and provides the dimension-padding idea; TFM is compared against it on all datasets."},{"cited_title":"Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting, Sept","cited_arxiv_id":null,"evidence_quote":"ConvLSTM is a recurrent spatio-temporal baseline that TFM is benchmarked against."},{"cited_title":"SADM: Sequence-Aware Diffusion Model for Longitudinal Medical Image Generation","cited_arxiv_id":null,"evidence_quote":"Provides the ViViT variant used as the transformer-based baseline for fair comparison."},{"cited_title":"Maier-Hein, Peter M","cited_arxiv_id":null,"evidence_quote":"ACDC cardiac cine-MRI is one of the three benchmark datasets used to establish the claims."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ISLES perfusion-CT stroke data is one of the three benchmark datasets."},{"cited_title":"The LUMIERE dataset: Lon- gitudinal Glioblastoma MRI with expert RANO evaluation","cited_arxiv_id":null,"evidence_quote":"Lumiere longitudinal glioblastoma MRI is the small-cohort benchmark where baselines collapse and TFM still beats LCI."},{"cited_title":"TorchCFM, Jan","cited_arxiv_id":null,"evidence_quote":"Supplies the UNet backbone with cross-attention used for the flow network in all TFM experiments."}],"review_version":1}