{"id":"041c8c01-bc9c-4434-8ca8-960ae2677024","arxiv_id":"2506.09668","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A conditional implicit neural atlas trains on MRI and labels to generate high-resolution fetal and neonatal brain atlases in minutes, with conditioning on age, birth age, ventricle volume, and corpus callosum presence.","lead":"CINeMA builds fetal and neonatal brain atlases with a neural network that maps 3D coordinates and per-subject codes to MRI tissue maps, skipping image registration. It claims minutes-scale training, support for rare conditions like enlarged ventricles and missing corpus callosum, and better segmentation and age prediction than existing atlases.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The conditional pathology claim rests on a disentanglement that the paper admits is unenforced; the only quantitative checks are model outputs, not causal tests of the conditioning variable.","rationale":"The reader identified the same load-bearing assumption: the explicit conditioning dimensions xi are not strictly disentangled from latent anatomy, and the paper itself states in Section VI-A that the network could ignore explicit conditions. I agree that this is the critical point for the paper's central claim. A conditional atlas that silently ignores its conditioning label is not a conditional atlas; it is a population-average atlas with a decorative input. The current evidence has three gaps: (1) For VM, the only quantitative evidence is a correlation computed on the model's own outputs, which cannot distinguish causal control from confounded latent structure. (2) For ACC, there is no quantitative outcome measure at all, only visual inspection. (3) The segmentation experiments, where the method is quantitatively strongest, explicitly do not use the explicit conditioning, so Table I does not support the conditioning contribution. The permutation test I propose is decisive because it directly tests whether the model uses xi or whether the apparent conditioning is an artifact of correlations with other encoded factors. If the permuted control reproduces the condition-dependent anatomy, the paper should be revised to either enforce disentanglement (e.g., with an auxiliary prediction loss or a penalty that makes xi necessary for reconstructing the conditioned anatomy) or to weaken the claims to non-causal 'style' manipulation. If the permuted control does not reproduce the effect, the conditioning claim is substantially supported and the conditional acceptance is warranted. I therefore keep the reader's CONDITIONAL verdict unchanged: the framework is plausible and the empirical work is mostly careful, but the central conditional-generation claim should not be accepted at face value until this causal check is performed.","tokens_in":18600,"tokens_out":3203,"duration_ms":40003,"concrete_test":"Run a permutation control on the MarsFet VM cohort: train CINeMA identically, but with the LV-volume labels xi randomly permuted across training subjects, breaking the correspondence between xi and the subject's actual ventricular volume while preserving the marginal distribution and all other training details. Then, for several fixed ages, generate atlases with the same regressed latent code z_t and vary xi from -1 to +1; segment LV volume and compute the GI-LV correlation on the generated atlases. If the permuted-xi model reproduces the same monotonic LV enlargement and GI-LV trend, the explicit condition is not the causal variable and the conditioning claim collapses. If the permuted model yields flat or random LV behavior, the conditioning is real.","verdict_should_be":"UNCHANGED","load_bearing_attack":"CINeMA's central contribution beyond its predecessor CINA is flexible conditioning on pathologies such as ventriculomegaly and agenesis of the corpus callosum. The mechanism for this is the concatenation of the explicit condition xi to the spatial latent code in Eq. 2 (Section III-A.3) and the claim in Section III-D that this 'enables disentangled conditioning.' However, Section VI-A explicitly concedes that 'this separation is not strictly enforced, theoretically allowing the network to ignore explicit conditions.' The evidence that the network does not ignore them is thin. For VM, the principal quantitative support is Fig. 3b, a correlation (r = -0.97) between conditioned LV volume and the gyrification index of the generated atlas. That is an output of the model, not evidence that xi is causally controlling anatomy; the same trend could be produced by age- or subject-correlated structure in the latent code z_t even if xi is unused. For ACC, Section V-B.3 provides only qualitative visual inspection of generated atlases, with no metric such as corpus callosum volume or overlap. Moreover, Section V-A reports that explicit conditioning gave no benefit in the segmentation experiments, and Table I was generated without explicit conditioning. The headline accuracy claim therefore does not certify the conditioning claim. If the network ignores xi in the VM and ACC experiments, then the 'conditional atlas for pathologies' contribution is not established, and the generated 'VM' and 'ACC' atlases are confounded reconstructions of the training population rather than controlled anatomical manipulations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CINeMA, an auto-decoder implicit neural representation framework for building spatio-temporal, multimodal atlases of the perinatal brain. Subject-specific variability is encoded in spatial latent codes and learned rigid alignments, while explicit conditioning variables (gestational age, birth age, ventricle volume, corpus callosum status) are concatenated to the latent code. Atlas generation at a target time point is performed by a Gaussian-weighted average of training latent codes followed by a single forward pass. The authors report improved segmentation, scan-age prediction, and modality translation compared to registration-based and GAN-based baselines on dHCP and MarsFet data, and they demonstrate qualitative conditional atlases for ventriculomegaly and agenesis of the corpus callosum. The central claims are that CINeMA surpasses state-of-the-art accuracy, operates in low-data regimes, and enables flexible, disentangled conditioning on pathologies.","tokens_in":18922,"tokens_out":6772,"duration_ms":75847,"significance":"If established, CINeMA would be a practically valuable alternative to registration-based atlas pipelines: it requires no deformable registration, builds an atlas in minutes, supports multiple modalities, and can be adapted to individual subjects in seconds. The paper provides a public code release, evaluates on multiple datasets including pathological cases, and includes a thorough ablation study of architectural choices (multimodal input, rigid alignment, spatial latent codes, explicit conditioning). The runtime advantages and the breadth of downstream tasks are genuine strengths. However, the most novel component—conditional generation of pathological atlases—is not quantitatively validated, and the paper itself concedes that the explicit condition is not enforced to be disentangled. The global accuracy claim is also stronger than the evidence in Table I, which shows a PSNR deficit on term neonates relative to Atlas-GAN. These issues are central to the paper's contribution and need to be addressed before the claims can be accepted.","major_comments":[{"comment":"The abstract states that CINeMA surpasses state-of-the-art methods 'in accuracy,' and the conclusion claims 'higher anatomical fidelity and more accurate representations than established baselines.' Table I does not fully support this. On dHCP term neonates, Atlas-GAN achieves PSNR 23.75±0.88 versus CINeMA's 23.43±0.55, and only SSIM and DSC are marked as significantly better for CINeMA. The accuracy advantage is thus metric- and cohort-dependent. The authors should either qualify the claim or explain the PSNR trade-off, e.g., by reporting a combined evaluation or discussing why higher structural similarity does not translate to pixel-wise fidelity.","section":"Abstract, Section VII, Table I"},{"comment":"The conditional pathology atlas is the central novel contribution, but the manuscript does not demonstrate that the explicit condition ξ actually controls the generated anatomy. For ventriculomegaly, the only quantitative evidence is Fig. 3b, which reports a correlation (r=-0.97) between the conditioned LV volume and the gyrification index of the generated atlas. This is a model output, not a causal test: the same trend could be produced by age- or subject-correlated structure in the latent code z even if ξ were ignored. For ACC, Section V-B.3 provides only qualitative visual inspection, with no metric such as corpus callosum volume, overlap, or a classification accuracy on generated images. Section VI-A explicitly concedes that 'this separation is not strictly enforced, theoretically allowing the network to ignore explicit conditions.' To support the conditioning claim, the authors should provide direct quantitative tests, such as measuring LV volume in generated atlases as a function of ξ and comparing to target values, decoding the condition from generated images, or ablating ξ during training (e.g., training with shuffled condition labels and showing that conditioning performance degrades).","section":"Section III-D, Section V-B.2, Section V-B.3, Section VI-A"},{"comment":"The text in Section V-A states that CINeMA demonstrates 'more accurate tissue segmentation throughout all datasets,' but for the severe ventriculomegaly cohort the DSC metric is reported only for the lateral ventricles (dagger footnote), not for the full tissue classes used in the other cohorts. This overstates the pathological segmentation result. The authors should either report multi-class DSC for this cohort or explicitly state that the evaluation is limited to the lateral ventricles in the main text, not only in a table footnote.","section":"Section V-A, Table I (MarsFet severe ventriculomegaly row)"}],"minor_comments":[{"comment":"The Gaussian kernel standard deviation is given as σ_b=0.5 weeks, but the training ages t_i are said to be normalized to [-1,1]. Please clarify whether t in Eq. (3) is expressed in weeks or in the normalized coordinate system, and report a sensitivity analysis for the choice of σ_b, especially given the admitted limitations for unbalanced data in Section VI-A.","section":"Section III-B, Eq. (3)"},{"comment":"The notation 'z 256×3^3' and the CNN kernel-size entries are difficult to parse, and the 'shaded row' indicating the final configuration may not be visible in black-and-white printing. Please make the configuration indicator explicit, e.g., with a row label or a footnote.","section":"Table II"},{"comment":"The text says the LV volume 'ξ' is added as a 'disjoint, i.e., static, dimension,' but Section III-C describes ξ as learnable during test-time adaptation. Please clarify what 'static' means during training versus inference.","section":"Section V-B.2"},{"comment":"There is a typo: 'primising' should be 'promising' in the discussion of spatial transformer networks.","section":"Section VI-A"},{"comment":"The note that CINeMA was not explicitly conditioned on LV volume in the segmentation experiments and that explicit conditioning gave no benefit is important context; it should be moved to or repeated in the limitations discussion so that readers do not conflate the segmentation results with the conditional generation results.","section":"Section V-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically solid in its core reconstruction and segmentation experiments, and the code release is a strength. The main risk is that the conditional pathology modeling—the headline novelty relative to CINA—relies on an unenforced disentanglement and is validated only qualitatively or indirectly. The accuracy claims also need to be aligned with Table I. These issues are fixable with additional experiments and revised claims, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a credible extension of CINA, with useful empirical work, but the headline claim about conditional pathology atlases rests on thinner evidence than the paper suggests. The paper does several things well: it integrates multimodal input, spatial latent codes, and learnable rigid alignment into the auto-decoder atlas framework; trains in about 12 minutes; and evaluates on neonatal and fetal datasets including pathological cases. The ablation study is honest and informative. The main tables show CINeMA generally beats baselines in segmentation and age prediction, especially in low-data and severe VM conditions.\n\nThe soft spots are real but not fatal. First, the abstract says 'surpassing state-of-the-art methods in accuracy,' but Table I shows Atlas-GAN has higher PSNR on term neonates (23.75 vs 23.43); CINeMA wins on SSIM and DSC. That overstatement should be fixed. Second, the conditional pathology claim—the main novelty over CINA—depends on a disentanglement the authors admit is not enforced (Section VI-A). The evidence that the network actually uses the explicit condition is thin: for VM, a single correlation between LV volume and gyrification index of generated atlases (r=-0.97), which is a model output, not a causal test; for ACC, only qualitative visual inspection. And in segmentation, explicit conditioning gave no benefit, and Table I was generated without it. So the conditional atlases for VM/ACC are plausible but not convincingly validated. The severe VM cohort is just 5 subjects, so those numbers are suggestive, not definitive. The code is released but not pinned to a commit, which is a minor practical issue.\n\nOverall, this is a solid paper that deserves peer review. The core contribution—a fast, latent-space, multimodal atlas with spatial latent codes—appears to hold up. The conditioning claim needs either stronger causal or predictive evidence (e.g., showing the condition variable is actually used, or quantitative evaluation of generated ACC anatomy). I'd tell the authors to soften the abstract and strengthen the conditioning validation.","headline":"Useful extension of CINA with honest ablations, but the conditional pathology claim is not yet backed by strong evidence.","tokens_in":19466,"tokens_out":2028,"would_cite":true,"duration_ms":20705,"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":"A single conditional implicit neural network replaces registration-based atlas pipelines, producing high-resolution spatio-temporal atlases of healthy and pathological fetal and neonatal brains in minutes from small cohorts.","keywords":["conditional atlas","perinatal brain imaging","implicit neural representation","spatio-temporal atlas","magnetic resonance imaging","tissue segmentation","ventriculomegaly","agenesis of the corpus callosum"],"falsifier":"Take a trained CINeMA model, fix a subject's latent code, and regenerate the atlas while sweeping the explicit conditioning value from its minimum to its maximum; if the lateral-ventricle volume present in the generated segmentation maps does not change along the sweep, or if toggling the ACC label does not make the corpus callosum appear or disappear, then the explicit conditions are not controlling the generated anatomy and the pathological atlases are not condition-driven.","tokens_in":18428,"feed_emoji":"🧠","tokens_out":9903,"duration_ms":97082,"temperature":0.7,"pith_summary":"Brain atlases are standard tools for studying how the fetal and neonatal brain grows, but conventional atlas pipelines require many age-balanced scans and days of deformable registration, which breaks down for rare pathologies where data are scarce. CINeMA, a conditional implicit neural multimodal atlas, claims to remove that bottleneck: one neural network is trained across a cohort, enrolling each subject's anatomy in a personal latent code, and any desired atlas—at any gestational age, resolution, or modality, and optionally conditioned on ventricle size or corpus-callosum status—is a weighted averaging of latent codes followed by a single forward pass. Training takes roughly 12 minutes and atlas inference about 3 seconds, and the same model performs tissue segmentation, scan-age and birth-age prediction, and cross-modal translation. The paper reports higher anatomical fidelity and segmentation accuracy than registration-based and GAN-based atlases, especially for severe ventriculomegaly, where the non-pathological baselines largely fail. If the central claim is correct, CINeMA offers a practical way to model atypical neurodevelopment from small cohorts and to generate synthetic but anatomically informed data.","feed_headline":"One neural network builds perinatal brain atlases in minutes, not days","feed_subtitle":"Conditioning on ventricle size and callosal agenesis lets it model pathological brains from small cohorts.","key_machinery":"The load-bearing object is the auto-decoded, conditionally modulated implicit neural representation: per-subject spatial latent codes $z_i \\in \\mathbb{R}^{256 \\times 3 \\times 3 \\times 3}$, queried by trilinear interpolation at each 3D coordinate, are concatenated with explicit condition variables $\\xi$ and converted by a linear layer into scale-and-shift parameters that modulate sinusoidal (SIREN) MLP layers, while per-subject rigid rotations and translations $R_i$ align all brains into a single reference space during training. The shared weights $\\theta$ of the MLP encode the population-level anatomy; the latent codes absorb subject-specific anatomy. To build an atlas of age $t$, the training latent codes are combined through a Gaussian kernel with width 0.5 weeks: $\\bar{z}_t = \\sum_i w(t,t_i) z_i$, and one forward pass through the modulated MLP outputs intensities and tissue probabilities in any requested modality and resolution. This architecture is what removes deformable registration, makes atlas construction a single forward pass, and enables conditioning and latent-space interpolation.","core_discovery":"CINeMA is an extension of the conditional implicit neural atlas idea. A single multilayer perceptron with sinusoidal activations is trained across a cohort of fetal or neonatal brain scans; the network's shared weights learn the anatomy common to the population, while each subject's anatomical individuality is pushed into a subject-specific spatial latent code, and a per-subject learnable rigid transformation handles gross misalignment. Because variability lives in low-dimensional codes rather than in deformation fields, atlas construction requires no image registration. At inference, a temporal atlas for any target postmenstrual age is produced by taking a Gaussian-weighted average of training latent codes around that age and running one forward pass through the network; explicit extra dimensions concatenated to the code allow the atlas to be conditioned on continuous or discrete anatomical characteristics such as lateral ventricle volume or the complete absence of the corpus callosum. The paper reports that this procedure yields higher anatomical fidelity, sharper tissue segmentations, and more accurate age prediction than registration-based atlases and a GAN-based conditional atlas, and that it does so with far less data and in minutes rather than days, including for pathological ventriculomegaly and agenesis of the corpus callosum populations.","pith_inferences":["Editorial: The same conditional mechanism could be used as an in-silico assay: generate atlases across a ventricle-volume sweep and compare the implied gyrification index or cortical-thickness trajectories with real longitudinal cohorts, testing the atlas as a causal model rather than a descriptive one.","Editorial: Because the network is trained without any hard constraint tying the condition dimensions to anatomy, a cheap diagnostic—randomly permuting condition labels across subjects and retraining—would reveal whether the reported pathological atlases are a genuine effect of conditioning or just population averages.","Editorial: If latent codes prove stable and compact, they could serve as a nuisance-removed phenotype for group studies, letting researchers compare preterm versus term populations directly in latent space instead of in image space.","Editorial: The framework should extend to other rare malformations by stacking additional binary or discrete condition dimensions, but the paper only demonstrates two pathologies in isolation; combined multi-pathology conditioning, which is untested, would show whether the conditioning space scales."],"forward_implications":["Atlas construction for a new population would take about 12 minutes of training on a single GPU plus a roughly 3-second forward pass, instead of the days required by registration-based pipelines.","A user could generate a continuous atlas at any gestational age and spatial resolution, and condition it on lateral ventricle volume or corpus-callosum status, including values interpolated between, or extrapolated beyond, the training distribution.","For a new subject, adapting the atlas to the scan's T2w image alone yields tissue probability maps, scan-age and birth-age estimates, and translation to T1w (and vice versa) without external registration.","Tissue-volume growth trajectories and gyrification trends derived from CINeMA atlases match established normative curves and reproduce the known inverse relation between ventricle enlargement and cortical folding (reported as $r = -0.97$ on generated atlases).","On the reported cohorts, CINeMA reports higher PSNR/SSIM, higher Dice scores in atlas-based segmentation, and lower scan-age prediction error than all three baselines; the largest margin appears for severe ventriculomegaly fetuses."],"supporting_citations":[{"why":"Predecessor framework CINA whose auto-decoder conditioning design this work extends.","marker":"[14]"},{"why":"Introduces sinusoidal activation functions (SIREN) used as the INR backbone for high-frequency detail.","marker":"[29]"},{"why":"Auto-decoder formulation central to encoding per-subject variability in latent codes without registration.","marker":"[30]"},{"why":"Supplies the linear scale-and-shift modulation layers used to condition the INR on latent codes.","marker":"[32]"},{"why":"Origin of spatial latent codes with interpolation, replacing the 1D latent codes of CINA.","marker":"[37]"},{"why":"Fetal BD-Atlas baseline against which CINeMA's segmentation and age-prediction accuracy are compared.","marker":"[6]"},{"why":"Neonatal BD-Atlas baseline providing the main traditional comparison for tissue segmentation.","marker":"[21]"},{"why":"Atlas-GAN baseline, the main conditional generative competitor, especially for pathological anatomy.","marker":"[11]"},{"why":"Registration library used to build a groupwise registration baseline from the same training data.","marker":"[45]"},{"why":"Quality-controlled tissue segmentation pipeline that provides ground truth labels for training and evaluation.","marker":"[40]"}],"fun_headline_variants":["Perinatal brain atlas generated in minutes with implicit neural net","CINeMA: implicit neural atlas for rare brain conditions","Fast spatio-temporal brain atlas from few scans, no registration","Multimodal fetal brain atlas built with a single MLP","CINeMA: minutes to atlas, handles pathologies with small cohorts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire conditional-atlas claim rests on the assumption that the network actually uses the explicit condition variables rather than ignoring them, a separation the authors themselves note is 'not strictly enforced.'","fun_headline_variants_meta":{"raw":{"variants":["Perinatal brain atlas generated in minutes with implicit neural net","CINeMA: implicit neural atlas for rare brain conditions","Fast spatio-temporal brain atlas from few scans, no registration","Multimodal fetal brain atlas built with a single MLP","CINeMA: minutes to atlas, handles pathologies with small cohorts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00021,"raw_usage":{"total_tokens":1462,"prompt_tokens":1048,"completion_tokens":414,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":664,"completion_tokens_details":{"reasoning_tokens":327}},"tokens_in":664,"tokens_out":414,"duration_ms":5044,"temperature":1.0,"reasoning_tokens":327,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:43:35.575678+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained CINeMA model, fix a subject's latent code, and regenerate the atlas while sweeping the explicit conditioning value from its minimum to its maximum; if the lateral-ventricle volume present in the generated segmentation maps does not change along the sweep, or if toggling the ACC label does not make the corpus callosum appear or disappear, then the explicit conditions are not controlling the generated anatomy and the pathological atlases are not condition-driven.","supporting_citations":[{"cited_title":"Cina: Conditional implicit neural atlas for spatio-temporal representation of fetal brains,","cited_arxiv_id":null,"evidence_quote":"Predecessor framework CINA whose auto-decoder conditioning design this work extends."},{"cited_title":"Deepsdf: Learning continuous signed distance functions for shape rep- resentation,","cited_arxiv_id":null,"evidence_quote":"Auto-decoder formulation central to encoding per-subject variability in latent codes without registration."},{"cited_title":"From data to functa: Your data point is a function and you can treat it like one,","cited_arxiv_id":null,"evidence_quote":"Supplies the linear scale-and-shift modulation layers used to condition the INR on latent codes."},{"cited_title":"Construction of a consistent high-definition spatio- temporal atlas of the developing brain using adaptive kernel regression,","cited_arxiv_id":null,"evidence_quote":"Fetal BD-Atlas baseline against which CINeMA's segmentation and age-prediction accuracy are compared."},{"cited_title":"Regional growth and atlasing of the developing human brain,","cited_arxiv_id":null,"evidence_quote":"Neonatal BD-Atlas baseline providing the main traditional comparison for tissue segmentation."},{"cited_title":"Generative adversarial reg- istration for improved conditional deformable templates,","cited_arxiv_id":null,"evidence_quote":"Atlas-GAN baseline, the main conditional generative competitor, especially for pathological anatomy."},{"cited_title":"deepali: Image, point set, and surface registration in pytorch (0.6.3),","cited_arxiv_id":null,"evidence_quote":"Registration library used to build a groupwise registration baseline from the same training data."},{"cited_title":"The developing human connectome project: A minimal processing pipeline for neonatal cortical surface reconstruc- tion,","cited_arxiv_id":null,"evidence_quote":"Quality-controlled tissue segmentation pipeline that provides ground truth labels for training and evaluation."}],"review_version":1}