{"id":"c3790083-940e-4551-8f91-969d1f73148e","arxiv_id":"2506.18720","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"TeNCA, a temporal neural cellular automata trained with adaptive losses at irregularly spaced time points, produces synthetic post-contrast breast MRI images that match ground truth more closely than U-Net and latent diffusion baselines.","lead":"This paper introduces TeNCA, a neural cellular automata model trained to generate synthetic contrast-enhanced breast MRI images from pre-contrast scans. It reports improved pixel-level agreement with real post-contrast images compared to a U-Net and a diffusion model, with far fewer parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Temporal conditioning is only well-defined if every acquisition time is an exact multiple of Δt=8s; the paper never verifies this, so the central temporal-consistency claim may rest on an unspecified rounding or skipped-loss rule.","rationale":"I read the central claim as TeNCA's superiority in image alignment plus temporal consistency. The most load-bearing point is not metric choice or error bars, but the mechanism that gives temporal meaning to NCA iterations: the paper's contribution is conditioning at physical times, and if the discrete steps cannot land on recorded times, the loss never fires for those phases. This is a concrete instantiation of the reader's weaker assumption about fixed-duration steps and physical time; my concern adds a specific, checkable condition about divisibility by Δt. I considered the PSNR bias toward the MSE-trained TeNCA and the absence of significance tests; these are real secondary issues, but LPIPS, SSIM, and MS-SSIM also favor TeNCA, so the aggregate image-quality claim does not collapse on that basis. The timing-alignment issue, by contrast, directly threatens the temporal-consistency component of the central claim. Because the concern is resolvable by inspecting data metadata and code, it warrants the same CONDITIONAL verdict the reader gave, not a rejection: if all acquisition times are indeed multiples of 8s, or if a clearly described rounding rule preserves per-phase conditioning, the stated mechanism stands. I therefore keep the reader's verdict unchanged while flagging the specific check that should be run before accepting the temporal claim.","tokens_in":8032,"tokens_out":6189,"duration_ms":70508,"concrete_test":"Audit the acquisition-time metadata for all 1,604 training pairs: extract every post-contrast acquisition time t_i and compute t_i mod 8. If any residual is nonzero, inspect the released TeNCA training code to determine the actual rule used for the Algorithm 1 equality check (exact, floor, nearest-step, or skipped loss). Then rerun the Table 1 and Figure 3 evaluation after replacing Δt=8 with the greatest common divisor of all observed t_i (or with the rounding rule actually implemented, if different), and compare the aggregate and per-phase metrics. If rankings or per-phase trends shift materially, the temporal-consistency claim is an artifact of an unspecified timing convention.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core novelty is that discrete NCA iterations correspond to physical time via Δt=8s and that intermediate states are conditioned at actual acquisition times. In Section 3, Equation (3) computes the loss at update step t_i/Δt, and Algorithm 1 increments t by Δt and applies the loss only when 't equals t_i'. For this mechanism to work as stated, every acquisition time t_i in the combined MAMA-MIA/Duke training set must be an exact integer multiple of 8 seconds. The paper emphasizes 'temporally sparse, non-uniformly sampled imaging data' and protocols with varying acquisition times, but it never states that times were rounded, that non-multiple cases were excluded, or that the equality check uses a tolerance. If any t_i is not divisible by 8, the corresponding ground-truth phase is silently ignored during training; the model is then conditioned on only a subset of phases, and the per-phase consistency shown in Figure 3 is not produced by the mechanism described. If the implementation instead rounds to the nearest step, the up-to-4s timing error is unreported and could be significant for early phases where contrast uptake changes rapidly. This is an internal-consistency concern, independent of whether the final image metrics are plausible: it targets the stated causal link between iterative updates and physical contrast kinetics, which is the basis for the temporal superiority claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TeNCA (Temporal Neural Cellular Automata), an NCA-based method for synthesizing post-contrast breast MRI phases from a pre-contrast image. The key idea is to reinterpret each NCA update step as a fixed physical time increment (Δt = 8 s) and to compute the training loss at intermediate states corresponding to observed acquisition times, thereby learning a continuous, temporally conditioned contrast-enhancement trajectory. The method is trained on a combined MAMA-MIA and Duke breast MRI dataset and compared with a U-Net baseline and the CC-Net latent diffusion model. The authors report that TeNCA achieves the best LPIPS, SSIM, MS-SSIM, and PSNR on the test set with about 13,000 parameters, while CC-Net achieves better FID and FRD. They further claim temporal stability and sequential consistency based on per-phase metric plots and qualitative videos.","tokens_in":8335,"tokens_out":5296,"duration_ms":57528,"significance":"If fully substantiated, the paper would make a useful contribution: a parameter-efficient architecture for synthetic contrast enhancement with explicit temporal conditioning, trained and evaluated on a multi-center public dataset, with code provided. The strengths include the use of external baselines, per-phase evaluation in Figure 3, the very small parameter count of TeNCA, and the transparent reporting that CC-Net attains better distribution-level metrics. These features make the work of interest to the medical image synthesis community. However, the current evidence falls short of the paper's strong 'surpassing current SOTA' and 'prove superiority' claims due to missing statistical validation, an under-specified temporal alignment mechanism, and a temporal-consistency evaluation that is only qualitative.","major_comments":[{"comment":"The loss is computed only at update step t_i / Δt (Eq. 3), and Algorithm 1 applies the loss only when 't equals t_i'. With Δt = 8 s, this mechanism is well-defined only if every acquisition time in the combined training set is an exact integer multiple of 8 s. The paper neither verifies this condition nor states a rounding rule or tolerance for the equality check. If some observed t_i are not multiples of 8 s, the corresponding ground-truth phase is silently excluded from training, and the per-phase results in Figure 3 are not produced by the claimed conditioning mechanism. If rounding is used, the resulting timing error (up to 4 s) is unreported and could be significant for early phases where contrast uptake is fast. This is a load-bearing internal-consistency issue for the paper's central temporal-modeling claim and must be clarified.","section":"Section 3, Eq. (3) and Algorithm 1"},{"comment":"The reported improvements over the U-Net baseline are small (LPIPS 0.12 vs. 0.13, SSIM 0.89 vs. 0.88, MS-SSIM 0.93 vs. 0.92, PSNR 32.26 dB vs. 31.93 dB), yet no error bars, confidence intervals, or significance tests are provided. The reader cannot tell whether these differences are consistent across the 300 test cases or driven by a few outliers. In addition, PSNR is a monotone transform of MSE, which is TeNCA's training loss, so the PSNR advantage is partly by construction; the paper should report metrics that are not direct transforms of the training objectives, or at least discuss this confound. Without statistical support, the claim of 'surpassing current SOTA' is not established.","section":"Table 1 and Figure 3"},{"comment":"The temporal consistency claim is supported only by per-phase mean metric plots (Figure 3) and qualitative example videos provided online. Per-phase fidelity averaged over the test set is not a temporal consistency metric: a model could have good per-phase metrics while producing non-physiological jumps between phases. To support 'prove TeNCA's superiority in learning temporal patterns', the paper needs a quantitative temporal measure, such as the smoothness of predicted intensity trajectories, the displacement or variance between consecutive predicted phases, or a comparison of kinetic curves against the ground-truth enhancement dynamics. Without such a metric, the central temporal-consistency claim is not evidenced.","section":"Section 4.4, 'Temporal stability' and 'Sequential consistency'"},{"comment":"CC-Net outperforms TeNCA on both distribution measures (FID 21.28 vs. 27.83 and FRD 20.00 vs. 48.68). The authors explain this as CC-Net 'hallucinating' realistic-looking images, which is plausible, but it also means that the overall 'superiority' claim is not uniformly true. The paper should temper the abstract and conclusion wording ('surpassing current SOTA performance', 'prove TeNCA's superiority') to reflect that TeNCA is better on pixel- and perception-level fidelity while CC-Net is better on distribution-level realism, and should discuss which of these properties is more clinically relevant for synthetic contrast-enhanced MRI.","section":"Table 1 and Section 4.4"}],"minor_comments":[{"comment":"There are grammatical errors in the first paragraph: 'To address this limitations' and 'Wide adoption ... is hindered by its high costs and lengthy acquisition times' should be corrected.","section":"Section 1"},{"comment":"The final hyperparameters for TeNCA (number of update steps N_steps, training epochs, learning rate, batch size) are not reported. The code link helps, but for reproducibility the paper should state the chosen values, especially N_steps, which directly determines the maximum simulated time horizon.","section":"Section 4.2"},{"comment":"The dataset description does not specify how many Duke cases were used, how they were selected, or whether the random validation split was stratified by center or protocol. Since the paper emphasizes diversity across protocols, this information is needed to assess the generality claims.","section":"Section 4.1"},{"comment":"The indentation of the loops in Algorithm 1 is ambiguous: it is unclear whether the inner loop over i is intended to run inside the update-step loop or only after it. The pseudocode should be restructured so that the scope of the loss accumulation is unambiguous.","section":"Algorithm 1"},{"comment":"There is a typo in the affiliation list ('Klinkum' should be 'Klinikum'), and the abstract uses 'prove' in 'prove TeNCA's superiority in learning temporal patterns'; given the current evidence, a more measured phrase such as 'provide evidence for' would be more appropriate.","section":"Author affiliations and abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper is on a relevant topic and the lightweight TeNCA architecture is appealing, but the central claims require stronger evidence than is currently provided. The most important issue is the unspecified handling of acquisition times that are not exact multiples of Δt = 8 s; if the implementation silently drops such phases, the temporal conditioning mechanism is not what the paper describes. I would like the authors to confirm the alignment rule, add statistical validation, and replace the qualitative temporal-consistency discussion with a quantitative metric before this paper is accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version: TeNCA is a genuine contribution to the NCA-for-image-synthesis line, and the paper is honest enough to report that the diffusion baseline wins on distribution metrics. But the paper's central temporal-consistency claim has a gap that the authors don't address, and the evaluation is thinner than the prose suggests.\n\nWhat's new: the training strategy that applies the NCA loss at irregularly spaced time points, with each update step mapped to Δt=8 seconds. Prior NCA temporal work used uniform or dense sampling; this adaptive-loss approach is a real extension. The model is also strikingly cheap (13k parameters) and the authors evaluate on a diverse multi-center dataset with code released. Credit where due: they report FID/FRD even though CC-Net beats them there, and they explain the pixel-level vs distribution tradeoff.\n\nThe soft spots, in order of importance.\n\nFirst, the Δt=8s assumption. Algorithm 1 conditions the loss only when the accumulated time 'equals t_i'. For that to work, every acquisition time in MAMA-MIA and Duke must be an exact multiple of 8 seconds. The paper never says that. If a t_i is not divisible by 8, that ground-truth phase is silently ignored; if the implementation rounds, the rounding error (up to 4s) is unreported and could matter in early phases where contrast uptake changes fast. This isn't a nitpick; it's the mechanism that grounds the iterative updates in physical time, and it's unverified.\n\nSecond, the evaluation. There are no error bars or significance tests on any metric. PSNR is a monotone transform of MSE, which is exactly TeNCA's training loss, so the PSNR comparison is stacked. Distribution metrics favor CC-Net, and the authors' explanation (hallucination) is plausible but not quantified. Temporal consistency is shown via videos and per-phase plots, but the per-phase plots inherit the divisibility problem.\n\nThird, minor: 'prove superiority' overclaims, and the U-Net baseline is a small variant that may not be well tuned.\n\nOverall: the core idea is sound and the paper is readable. The stress-test concern is real and needs to be addressed before the temporal claim is credible. But that's a fixable gap, not a fatal flaw.\n\nWho it's for: anyone working on synthetic contrast enhancement or NCA-based temporal modeling. It deserves a serious referee. I'd recommend sending it to peer review with a request that the authors clarify the time-step handling and add statistical validation.","headline":"TeNCA's irregular-time training trick is real, but the paper needs to verify its 8-second divisibility assumption and add error bars before the temporal claims hold.","tokens_in":8834,"tokens_out":2541,"would_cite":true,"duration_ms":24637,"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":"Eight-second neural-cell steps beat U-Net and diffusion on breast MRI","keywords":["neural cellular automata","temporal image synthesis","dynamic contrast-enhanced MRI","breast MRI","synthetic contrast enhancement","temporal consistency","lightweight models","multi-center medical imaging"],"falsifier":"Train or evaluate TeNCA on a cohort whose acquisition times differ from the training distribution, then compare a held-out middle time point (for example at 90 seconds) against the model's state after 90/8 steps; if per-phase image metrics at the held-out time are substantially worse than at the training times, or if protocol-stratified metrics show the advantage disappears for protocols with long or irregular spacing, the temporal-consistency claim is falsified.","tokens_in":7851,"feed_emoji":"🩻","tokens_out":5644,"duration_ms":53100,"temperature":0.7,"pith_summary":"This paper introduces TeNCA, a temporal neural cellular automaton that learns to evolve a pre-contrast breast MRI into post-contrast phases by treating each update step as eight seconds of physical time. The authors claim that, trained on a combined multi-center dataset of 1,604 cases, TeNCA matches ground-truth post-contrast sequences better than a U-Net and a latent-diffusion baseline under LPIPS, SSIM, MS-SSIM, and PSNR, while using only about 13,000 parameters. They further claim the model is temporally stable: it does not degrade in later phases the way the baselines do, and its intermediate states evolve continuously. This matters because synthetic contrast enhancement could shorten breast MRI protocols and remove contrast-agent injections, but only if generated sequences are trustworthy over time. The paper argues that conditioning the loss at every acquisition time forces the automaton to learn physiologically plausible contrast kinetics rather than a single static output.","feed_headline":"Eight-second neural-cell steps beat U-Net and diffusion on breast MRI","feed_subtitle":"A 13,000-parameter automaton tracks true post-contrast phases while larger models drift or hallucinate.","key_machinery":"The core object is the TeNCA transition function $F$, defined by $S_{t+1} = F(S_t) = S_t + (\\partial S/\\partial t)\\,\\Delta t$ with $\\Delta t = 8$ seconds, so the number of update steps to a phase is fixed by its acquisition time. Perception uses two learnable 3x3 convolutional kernels for neighbor communication plus an identity kernel for the cell's own state, all feeding a two-layer MLP; the state has 24 channels, one visible (the image) and the rest hidden. The mechanism that carries the argument is the loss in Eq. (3), computed at every time $t_i$ where ground truth exists, which forces the hidden state to encode a continuous transition rather than just the final contrast phase.","core_discovery":"The central claim is that a neural cellular automaton whose update rule is tied to a fixed time increment $\\Delta t = 8$ seconds can be trained on temporally sparse, non-uniformly spaced DCE-MRI acquisitions and produce intermediate states aligned with ground-truth post-contrast images. TeNCA initializes the visible state with the pre-contrast image, runs a shared transition function $F$, and computes the image loss at every step $t_i/\\Delta t$ for which a ground-truth acquisition exists. This adaptive sequential loss makes the temporal dimension part of the learning objective, not just a means to reach a final image. On the combined MAMA-MIA and Duke test data, TeNCA improves over both reference methods on image-level metrics while using far fewer parameters than the diffusion baseline; the paper attributes the baseline's poor pixel-level scores to hallucination, a known failure mode of diffusion models.","pith_inferences":["A testable extension is to withhold one intermediate acquisition time during training and compare TeNCA's state at that time against the held-out ground truth; success would directly validate the physical-time assumption.","If the fixed 8-second step truly encodes physical time, the same training scheme could apply to other dynamic imaging modalities with sparse temporal sampling, such as cine MRI, 4D CT, or perfusion imaging.","The paper does not verify the physiological meaning of TeNCA's hidden channels, so an interesting next step would be probing whether those channels encode contrast arrival or washout rates.","The comparison with diffusion suggests that local, continuity-enforcing inductive biases may matter more than raw generative capacity for medical temporal synthesis; this is an editorial reading, not a claim of the paper."],"forward_implications":["Generating all post-contrast phases from one pre-contrast scan requires no contrast injection and no extra acquisition time, so a reader could inspect several synthetic phases and recover some washout kinetics lost in abbreviated protocols.","Because the model has about 13,000 parameters and uses iterative local convolutions, it can run on modest hardware, making synthetic DCE-MRI feasible in resource-constrained screening settings.","Per-phase metrics show TeNCA maintains accuracy in later post-contrast phases where the U-Net baseline degrades, implying washout behavior is modeled rather than ignored.","A fixed 8-second time step standardizes heterogeneous protocols: the same learned transition function can be queried at any time point, enabling interpolation between acquired phases."],"supporting_citations":[{"why":"MAMA-MIA supplies the multi-center DCE-MRI benchmark and the train/test split used for evaluation.","marker":"[5]"},{"why":"Duke-Breast-Cancer-MRI augments the training cohort with additional T1-weighted fat-saturated cases.","marker":"[24]"},{"why":"Growing neural cellular automata provides the iterative cell-state evolution paradigm that TeNCA extends.","marker":"[16]"},{"why":"DyNCA supplies the NCA update formulation and the motivation for modeling temporal textures.","marker":"[20]"},{"why":"The latent diffusion model is the CC-Net baseline that TeNCA is compared against and whose conditioning approach is adapted.","marker":"[19]"},{"why":"The U-Net architecture is the other reference baseline and the source of the MCO-Net structure for the comparison.","marker":"[25]"},{"why":"LPIPS is used as the perceptual image-similarity metric for evaluating alignment with ground-truth post-contrast images.","marker":"[28]"},{"why":"MS-SSIM is used as a multi-scale structural similarity metric in the overall and per-phase evaluations.","marker":"[26]"},{"why":"Cellular automata as convolutional neural networks justifies representing the transition function with convolutional perception and an MLP update.","marker":"[6]"}],"fun_headline_variants":["13k-parameter automaton tracks breast MRI contrast over time","Temporal NCA beats U-Net and diffusion on breast MRI contrast","TeNCA: 8-second steps for synthetic breast MRI contrast","Lightweight neural automata model contrast enhancement in breast MRI","Neural cellular automata replace contrast agent in breast MRI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an 8-second fixed update step maps directly onto real contrast-agent kinetics and that one transition function learned from many centers can represent all protocols; if timing or dynamics vary by protocol in ways the model does not see, the intermediate states are not physiologically meaningful.","fun_headline_variants_meta":{"raw":{"variants":["13k-parameter automaton tracks breast MRI contrast over time","Temporal NCA beats U-Net and diffusion on breast MRI contrast","TeNCA: 8-second steps for synthetic breast MRI contrast","Lightweight neural automata model contrast enhancement in breast MRI","Neural cellular automata replace contrast agent in breast MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1417,"prompt_tokens":936,"completion_tokens":481,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":394}},"tokens_in":552,"tokens_out":481,"duration_ms":5281,"temperature":1.0,"reasoning_tokens":394,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:43:40.402351+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train or evaluate TeNCA on a cohort whose acquisition times differ from the training distribution, then compare a held-out middle time point (for example at 90 seconds) against the model's state after 90/8 steps; if per-phase image metrics at the held-out time are substantially worse than at the training times, or if protocol-stratified metrics show the advantage disappears for protocols with long or irregular spacing, the temporal-consistency claim is falsified.","supporting_citations":[{"cited_title":"British journal of cancer 119(4), 508–516 (2018)","cited_arxiv_id":null,"evidence_quote":"Duke-Breast-Cancer-MRI augments the training cohort with additional T1-weighted fat-saturated cases."},{"cited_title":"In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"DyNCA supplies the NCA update formulation and the motivation for modeling temporal textures."},{"cited_title":"In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"The latent diffusion model is the CC-Net baseline that TeNCA is compared against and whose conditioning approach is adapted."},{"cited_title":"In: Deep Breast Workshop on AI and Imaging for Diagnostic and Treatment Challenges in Breast Care","cited_arxiv_id":null,"evidence_quote":"The U-Net architecture is the other reference baseline and the source of the MCO-Net structure for the comparison."},{"cited_title":"In: The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003","cited_arxiv_id":null,"evidence_quote":"MS-SSIM is used as a multi-scale structural similarity metric in the overall and per-phase evaluations."},{"cited_title":"Physical Review E 100(3), 032402 (2019)","cited_arxiv_id":null,"evidence_quote":"Cellular automata as convolutional neural networks justifies representing the transition function with convolutional perception and an MLP update."}],"review_version":2}