{"id":"3d211206-ae48-4915-8fe1-35537f508172","arxiv_id":"2508.09757","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"NEUBORN combines hierarchical diffeomorphic registration with a Neo-Hookean biomechanical loss to model individual neonatal brain growth between two MRI scans.","lead":"Researchers describe NEUBORN, a deep learning method that maps how a baby's brain changes shape between two MRI scans, guided by a simple biomechanical model of brain tissue. The method matches existing registration accuracy while producing anatomically smoother brain deformations, but its strongest evidence about growth is weakened because it uses the follow-up scans' volume measurements to guide the growth and then tests against those same measurements.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported growth-trajectory accuracy is enforced by target-derived growth maps, making the ASPVC and trend claims circular rather than predictive.","rationale":"The reader correctly identifies the target-informed growth map as load-bearing and flags the resulting circularity. My analysis goes further: the circularity is not only a risk from spatial heterogeneity; it is a direct consequence of the loss formulation. The term (J_e−1)^2 in Eq. 7 explicitly penalizes deviations of the local volume change from the prescribed target ratio, making the low ASPVC essentially an enforcement of the growth map rather than a measured prediction. The reader's weakest_assumption also emphasizes the homogeneity of the growth prior; while that is a secondary concern, the primary issue is that the evaluation metric is tied to the same information used to construct the prior. The registration-quality aspects (Dice, negative Jacobians) are independent and credible, so the appropriate verdict remains CONDITIONAL: the growth-trajectory claim requires an out-of-sample test before it can be accepted. Since the reader already reached CONDITIONAL, no verdict change is needed, but the concrete test above should be run.","tokens_in":7642,"tokens_out":5193,"duration_ms":61022,"concrete_test":"Re-evaluate the 19 test subjects with growth maps g estimated without access to each test subject's follow-up segmentation: for example, fit a per-structure volume-age model on the 64 training subjects only, predict each test subject's g from their baseline age and volume, and use these out-of-sample g values in the same trained NEUBORN network. Then recompute ASPVC_cortex and plot the predicted volumes against a population trend fitted without the 19 test subjects. If ASPVC remains near 3.82 and the volumes still follow the trend, the subject-specific growth claim survives. If ASPVC degrades toward the no-biomechanical-loss value (~16.8), the reported accuracy is an artifact of target-informed growth maps.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 prescribes F_g=(g^{-1/3})I with g=V_{t1,seg}/V_{t2,seg} measured from the same follow-up segmentations later used as ground truth. Because det(F_g)=g^{-1}, the Neo-Hookean loss (Eq. 7) contains (J_e−1)^2 with J_e=det(F_e)=g·det(F). Minimizing this term drives the determinant of the learned deformation toward 1/g=V_{t2}/V_{t1} for every voxel in each structure. The reported ASPVC_cortex=3.82 and the visual alignment of predicted volumes with the population trend in Fig. 4 are therefore consequences of injecting the target volume ratios into the loss, not evidence that the model predicts subject-specific growth. The trend itself is computed from all 92 subjects, including the 19 test subjects, further weakening the 'trend adherence' comparison. The ablation result (ASPVC 16.8 without the biomechanical loss) is fully consistent with this: removing the loss removes the enforced target volumes. The registration-quality findings (Dice comparable to VoxelMorph, dramatically fewer negative Jacobians) remain credible, but the central growth-trajectory claim is not supported by the current evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents NEUBORN, a deep-learning framework for longitudinal diffeomorphic registration of neonatal brain MRI. It couples a hierarchical two-resolution U-Net stationary-velocity-field architecture with a Neo-Hookean hyperelastic biomechanical loss intended to enforce biologically plausible growth. The model is trained and evaluated on 92 dHCP preterm infants (64 train, 9 validation, 19 test). The authors report that NEUBORN matches VoxelMorph in Dice overlap, produces drastically fewer negative Jacobians than VoxelMorph, and yields lower ASPVC for cortical volume, which they interpret as better adherence to subject-specific and population-level growth trajectories. The abstract and conclusions claim that the framework 'generates growth trajectories that better follow population-level trends' and 'accurately preserves complex subject-specific cortical growth trajectories.'","tokens_in":7977,"tokens_out":5116,"duration_ms":60382,"significance":"If the growth-trajectory claims held, this would be a valuable contribution: it would combine state-of-the-art alignment accuracy with biomechanically interpretable deformations, potentially improving normative modeling of early brain development. The registration-quality findings—comparable Dice to a strong baseline and a four-order-of-magnitude reduction in negative Jacobians—are credible and of practical interest. However, the central growth-accuracy claim is not independently established because the biomechanical growth map is prescribed from the same follow-up volumes used as ground truth for evaluation. The paper deserves attention for its registration methodology, but the predictive and biological-validity claims require substantial reframing or re-evaluation.","major_comments":[{"comment":"The growth map is prescribed as F_g = (g^{-1/3})I with g = V_{t1,seg}/V_{t2,seg}, using the follow-up segmentation volumes that later serve as ground truth for ASPVC. Since det(F_g) = g^{-1}, the biomechanical loss (J_e − 1)^2 with J_e = det(F_e) = g·det(F) drives the determinant of the total deformation toward 1/g = V_{t2,seg}/V_{t1,seg} in every voxel of each structure. Thus the reported ASPVC_cortex = 3.82 (Table 1) is a direct consequence of injecting the target volume ratios into the objective, not evidence of prediction. The ablation in Table 2 (ASPVC rises to 16.8 without the biomechanical loss) is fully consistent with this interpretation. To support the growth-trajectory claim, the evaluation must use growth maps derived without access to the follow-up volumes (e.g., predicted from timepoint 1 or from a population prior) rather than the measured V_{t2,seg}.","section":"§2.2, Eq. (7); §3, Eq. (8), Table 1"},{"comment":"The definition of g requires V_{t2,seg}, the follow-up segmentation, at test time. The manuscript does not specify whether this volume is available during inference or whether g is computed from the reference image in a standard registration setting. As written, the method cannot 'generate growth trajectories' or 'enable forward projections of future brain changes' (Conclusions) because it needs the future volumes to define the growth map. Please clarify the exact inference protocol: if g is computed from the reference scan (which is legitimate for registration), the predictive claims must be removed; if g is predicted from timepoint 1 alone, the evaluation must reflect that setting.","section":"§2.2 'Subject-specific growth maps'; §4 Conclusions"},{"comment":"The population-level age-volume trends in Figure 4 are computed from 'all 92 available subjects,' which includes the 19 test subjects whose simulated follow-up volumes are compared against that same trend. This makes the claim that NEUBORN 'better follow[s] population-level growth trajectories' non-independent. The trend should be estimated from the training/validation subjects only, or via leave-one-out (or split-half) schemes. Additionally, the Wilcoxon test for NEUBORN cortical volumes against ground truth yields p=0.05, and all methods show significant white-matter volume differences; the abstract's statement that the model 'accurately preserves complex subject-specific cortical growth trajectories' overstates the evidence.","section":"§3, Figure 4; §3 'Evaluation methods'"}],"minor_comments":[{"comment":"The title uses 'BiOmechanical' while the section heading on page 3 reads 'Using BiOmechnical RemodelliNg' (missing 'a'). Please unify the spelling.","section":"Title and §3 heading"},{"comment":"The composition formula ϕ = ϕ1 ◦ ϕ2 + ϕ2 is unusual. If ϕ1 and ϕ2 are velocity fields that are integrated separately, the addition is not standard; please clarify the notation and confirm the intended composition.","section":"Eq. (3)"},{"comment":"The column 'Jϕ ≤ 0 (%)' is ambiguous: is this the percentage of negative-Jacobian voxels per image, or per total brain tissue? Please define clearly in the caption. Also, the Dice standard deviations (e.g., 0.0016) appear overly precise; consider reporting to 1-2 significant digits.","section":"Table 1"},{"comment":"The statement 'with fewer negative Jacobians' relative to state-of-the-art baselines is misleading when Elastic Syn has exactly zero negative Jacobians. Please qualify this claim as relative to VoxelMorph specifically.","section":"§3, Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper does not mention code or data availability, which is expected for a conference-style submission but should be addressed if the authors intend to claim reproducibility. The novelty relative to prior work [7] is incremental: the same biomechanical loss is adapted from atrophy to growth by prescribing per-structure target volumes. The main contribution is the demonstration that such a loss can be integrated into a hierarchical registration network without sacrificing Dice; the 'growth trajectory prediction' framing currently overreaches. Given that the load-bearing evaluation issue is identifiable and fixable by reframing or re-evaluating with a true predictive growth map, major revision is appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has two parts. The first, a biomechanically constrained diffeomorphic registration network for longitudinal neonatal MRI, is worthwhile and mostly sound. The second, the claim that it predicts subject-specific cortical growth trajectories, is compromised by a circular evaluation design. The stress-test note is correct on this.\n\nWhat's new: the specific integration of a Neo-Hookean hyperelastic loss, with per-structure growth scalars, into a two-level VoxelMorph-style architecture applied to dHCP longitudinal scans. The registration findings are credible. Dice is on par with VoxelMorph, and the negative-Jacobian reduction is dramatic. Those results are independent of the growth-model evaluation and support the authors' claim about smoother, more anatomically plausible warps.\n\nThe problem is in Section 2.2. The growth map uses g = V_t1,seg / V_t2,seg measured from the same follow-up segmentations that later serve as ground truth for ASPVC and the trend plots. Minimizing the Neo-Hookean term with det(F_g) = 1/g drives the Jacobian of the learned deformation toward 1/g, so the follow-up volume is essentially fed into the loss and then recovered. The ASPVC of 3.82 is not an independent prediction. The ablation result (ASPVC 16.8 without the biomechanical loss) is consistent with that reading. The trend comparison also includes the 19 test subjects in the trend computation, which further weakens the 'adherence' claim.\n\nThis is a fixable flaw. The authors could estimate growth maps from a subset of subjects, or from an intermediate timepoint, or evaluate on a held-out pair using a growth map derived from the first timepoint only. Until then, the growth-trajectory claim should be presented as a modeling choice rather than validated predictive accuracy.\n\nThe paper is a solid registration contribution with an overreaching evaluation. I'd send it to peer review, but with the expectation of a major revision focused on the circularity. Code and data release would also make a real difference.","headline":"The registration contribution is real, but the headline growth-trajectory claim is not yet supported: the growth prior is computed from the same follow-up volumes used for evaluation, so the accuracy is partly enforced rather than predicted.","tokens_in":8436,"tokens_out":2109,"would_cite":false,"duration_ms":23559,"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":"NEUBORN claims that injecting a tissue-level isotropic growth prior into a diffeomorphic deep registration network produces longitudinal warps that match state-of-the-art alignment accuracy while being orders of magnitude more anatomically","keywords":["biomechanical registration","longitudinal registration","neonatal brain MRI","cortical growth","diffeomorphic deformation","deep learning","Neo-Hookean hyperelasticity","growth decomposition"],"falsifier":"Compare NEUBORN's Jacobian determinant fields against regionally sub-parcellated volume changes computed from an independent longitudinal surface correspondence: if local Jacobian expansion does not track local tissue growth even though global volumes match, the subject-specific growth claim fails. Alternatively, run NEUBORN on test-retest scans with no growth; if it still predicts substantial volume expansion, the target-informed prior is overfitting.","tokens_in":7553,"feed_emoji":"🧠","tokens_out":7833,"duration_ms":78013,"temperature":0.7,"pith_summary":"This paper tries to show that longitudinal brain registration of neonatal MRI can be made both alignment-accurate and anatomically plausible by feeding a deep learning registration network a subject-specific biomechanical growth prior. The prior prescribes uniform expansion within each segmented brain tissue, derived from the volume change between the two time points, and a Neo-Hookean hyperelastic loss enforces mechanical plausibility. In tests on 19 unseen subjects, the method matches the alignment accuracy of a leading learning-based baseline while producing orders of magnitude fewer folded (negative Jacobian) voxels and simulated follow-up cortical volumes much closer to true individual volumes. If right, it makes individual growth-trajectory modelling from routine longitudinal scans both sharper and more biologically interpretable.","feed_headline":"Deep learning grows infant brains along true trajectories","feed_subtitle":"Matches top alignment accuracy while cutting anatomically impossible folds by four orders of magnitude.","key_machinery":"The load-bearing mechanism is the multiplicative decomposition of the deformation gradient $F = F_e F_g$ (elastic times growth), with $F_g$ prescribed isotropically per segmented structure as $g^{-1/3} I$ where $g$ is the linear-aligned volume ratio between time points, and $F_e$ computed through a differentiable Neo-Hookean strain-energy loss. This converts a biomechanical growth prior into a differentiable regulariser that steers the network's deformation away from implausible folds while letting the image-similarity loss handle alignment.","core_discovery":"The core claim is that prescribing growth at the tissue level inside a diffeomorphic, learning-based registration network yields deformations that are simultaneously as accurate, more diffeomorphic, and more biologically faithful than current baselines. The deformation is built from two hierarchical stationary velocity fields integrated by scaling and squaring; the biomechanical loss penalises strain energy after the deformation is factored as elastic response times prescribed growth, with growth set isotropically inside each segmented structure from the tissue's volume ratio across time. On held-out subjects, the method attains Dice overlap statistically indistinguishable from the learning-","pith_inferences":["The reported volume accuracy may be partly inherited from the target-informed growth map: the growth factor $g$ is computed from the follow-up volumes the network is asked to reproduce. A stricter test would predict growth from the baseline scan alone, without using the follow-up volume ratio, and compare ASPVC.","The paper itself notes only relative growth is assessed because linear registration removes absolute volume changes; this means the claim of 'growth trajectories' concerns relative expansion, not absolute brain size change.","A natural stress test is to replace the homogeneous isotropic growth with a spatially varying map and see whether the low ASPVC persists; this would separate the biomechanical regulariser's contribution from the information already contained in the volume-ratio prior."],"forward_implications":["By interpolating the learned velocity field, the framework can produce intermediate brain states between the two scan ages, giving denser growth trajectories from sparse follow-ups (flagged as future work in the paper).","The same per-subject deformation can be projected forward to simulate later brain structure at the patient level, supporting early identification of deviant cortical development.","Smoother, less folded warps make the resulting anatomical correspondences more reliable for downstream morphometric analyses.","The general structure — a differentiable biomechanical loss plus a prescribed growth field — can be adapted to other longitudinal brain-change scenarios, such as atrophy, by changing the growth prescription."],"supporting_citations":[{"why":"The learning-based diffeomorphic registration baseline whose Dice accuracy NEUBORN matches.","marker":"[3]"},{"why":"Provides the differentiable Neo-Hookean biomechanical loss that NEUBORN adapts for growth.","marker":"[7]"},{"why":"Supplies the scaling-and-squaring diffeomorphic integration used for the velocity fields.","marker":"[8]"},{"why":"The linear registration tool used to align volumes before tissue volume ratios $g$ are computed.","marker":"[12]"},{"why":"The coarse-to-fine multi-resolution registration strategy adapted for the twin U-Nets.","marker":"[18]"},{"why":"Introduces the stress-dependent finite-growth decomposition $F = F_e F_g$ that motivates the growth prescription.","marker":"[20]"},{"why":"The probabilistic registration with spatially adaptive regularisation from which the paper builds.","marker":"[24]"},{"why":"Biomechanical model of cortical folding that motivates linking growth patterns to mechanical tissue properties.","marker":"[26]"}],"fun_headline_variants":["Biomechanically constrained warps yield realistic infant brain growth","Infant brain growth modeled with biomechanical constraints","Biomechanical growth reduces anatomically impossible folds in infant brains","Diffeomorphic warps with tissue growth keep infant brain folds realistic","Fewer negative Jacobians: biomechanical growth improves infant brain warps"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"Everything hinges on the assumption that within each segmented brain structure, growth is a single uniform expansion whose rate is taken from the volume ratio after linear alignment; if true growth is uneven inside a structure, the reported volume accuracy is a fitting artifact and not evidence of biological fidelity.","fun_headline_variants_meta":{"raw":{"variants":["Biomechanically constrained warps yield realistic infant brain growth","Infant brain growth modeled with biomechanical constraints","Biomechanical growth reduces anatomically impossible folds in infant brains","Diffeomorphic warps with tissue growth keep infant brain folds realistic","Fewer negative Jacobians: biomechanical growth improves infant brain warps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000399,"raw_usage":{"total_tokens":1867,"prompt_tokens":636,"completion_tokens":1231,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":380,"completion_tokens_details":{"reasoning_tokens":1145}},"tokens_in":380,"tokens_out":1231,"duration_ms":8625,"temperature":1.0,"reasoning_tokens":1145,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:49:37.619587+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare NEUBORN's Jacobian determinant fields against regionally sub-parcellated volume changes computed from an independent longitudinal surface correspondence: if local Jacobian expansion does not track local tissue growth even though global volumes match, the subject-specific growth claim fails. Alternatively, run NEUBORN on test-retest scans with no growth; if it still predicts substantial volume expansion, the target-informed prior is overfitting.","supporting_citations":[{"cited_title":"IEEE Transactions on Medical Imaging 38(8), 1788–1800 (2019)","cited_arxiv_id":null,"evidence_quote":"The learning-based diffeomorphic registration baseline whose Dice accuracy NEUBORN matches."},{"cited_title":"Medical Image Analysis5(2), 143–156 (2001)","cited_arxiv_id":null,"evidence_quote":"The linear registration tool used to align volumes before tissue volume ratios $g$ are computed."},{"cited_title":"Large Deformation Diffeomorphic Image Registration with Laplacian Pyramid Networks","cited_arxiv_id":"2006.16148","evidence_quote":"The coarse-to-fine multi-resolution registration strategy adapted for the twin U-Nets."},{"cited_title":"Journal of Biomechanics 27(4), 455–467 (1994)","cited_arxiv_id":null,"evidence_quote":"Introduces the stress-dependent finite-growth decomposition $F = F_e F_g$ that motivates the growth prescription."},{"cited_title":"Medical Image Analysis 26(1), 203–216 (2015)","cited_arxiv_id":null,"evidence_quote":"The probabilistic registration with spatially adaptive regularisation from which the paper builds."}],"review_version":1}