{"id":"0c5d68d4-2da0-4745-95af-8bc4f6282029","arxiv_id":"2608.00187","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A landmark-guided, band-limited spectral template warp, preceded by semi-supervised landmark detection, produces dense and anatomy-confined correspondences directly from imperfect 3D photogrammetry.","lead":"SCALP automatically builds statistical skull models from raw, cluttered 3D baby head scans by first detecting facial landmarks with a semi-supervised network, then warping a clean template using only low-frequency shape modes. If it works as reported, it could replace CT-based monitoring of infant skull shape with radiation-free photography.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'no folding by construction' guarantee is unsupported: band-limited displacement fields can still invert locally; the central robustness claim needs a proof or numerical check.","rationale":"The paper has genuine strengths: a clean two-stage design, a matched supervised ablation, an oracle-landmark ablation showing Stage I is not the bottleneck, an explicit spectral-vs-TPS ablation, and consistent large-margin improvements on the reported test set. These support the empirical outperformance claim for the tested cohort. However, the central advertised mechanism—that band-limiting in the LBO eigenbasis is a hard geometric constraint ruling out folding—is the load-bearing assumption behind the 'mathematical robustness guarantee.' It is not proven and is not implied by the parameterization: a low-frequency displacement can still be non-injective, and the regularizer only penalizes coefficient energy. This matters because the clinical motivation is precisely severe missing data and deformities. The test-set-based selection of r (Appendix A, Table 5) is an additional concern, but it is secondary: even at suboptimal r, the reported margins over baselines are large. The reader's weakest_assumption about template-basis expressiveness is adjacent but not identical; both concern Stage II reliability. The right remedy is to turn the asserted guarantee into a checkable condition: either prove a bound or verify numerically that optimized warps have positive Jacobians. Until then, CONDITIONAL remains the right verdict; my read does not move it.","tokens_in":15895,"tokens_out":7949,"duration_ms":95000,"concrete_test":"For every optimized output mesh, compute per-triangle signed volume (or vertex Jacobian determinant of V0 + ΦW) and count negative determinants / inverted triangles. Also run a synthetic stress test: fit the head template by low-rank (r=64, 120) deformations to a target with a deep concavity or large missing region, with W unrestricted except by Eq. 10. If any negative Jacobian appears, the 'by construction' claim is false. If none appear on this dataset, the claim should still be softened from a mathematical guarantee to an empirical property, and a sufficient condition (e.g., a bound on ||ΦW||_{C^1}) should be derived.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central theoretical claim (Sec. 1; Sec. 3.3) is that optimizing within the first r Laplace–Beltrami eigenfunctions eliminates high-frequency mesh folding 'by construction,' providing a mathematical robustness guarantee against missing regions and severe outlier clutter. This is asserted, not proved, and as stated it is false. The deformation is Vdef(W) = V0 + ΦW (Eq. 7). Occupying a low-dimensional subspace of displacement fields does not by itself prevent the map from ceasing to be locally injective: injectivity requires det(I + ∇(ΦW)) > 0 at every vertex, and a low-frequency field can have arbitrarily large gradient if the coefficients are large. The eigenvalue-weighted regularizer (Eq. 10) penalizes L2 coefficient energy but imposes no hard bound on the C1 norm of the displacement; the two-way Chamfer data term (Eq. 9) can drive coefficients large, especially under severe partiality or template mismatch. The Sec. 5 limitation statement concedes sensitivity to template deviation and severe posterior missingness, undercutting the universal guarantee. If the guarantee fails in the very missing-data/pathology regimes motivating the work, then the mechanism claimed for robustness is not established; Table 2 averages could hide localized fold/collapse failures.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"SCALP proposes a two-stage pipeline for constructing statistical shape models directly from unprocessed infant photogrammetry scans. Stage I localizes 23 craniofacial landmarks using a semi-supervised Point Transformer v2 with Hough voting, trained on 50 expert-annotated subjects plus a larger unlabeled cohort. Stage II deforms a normative template mesh by optimizing coefficients in a truncated Laplace–Beltrami eigenbasis, anchored by the predicted landmarks and matched to the target scan via two-way Chamfer distance. The authors evaluate on 13 held-out subjects against six unsupervised correspondence baselines, reporting substantially lower landmark error, surface-fit error, and better compactness/generalization/specificity. Ablations compare semi-supervised vs. supervised Stage I, spectral vs. TPS warping, oracle vs. predicted landmarks, and the truncation rank r.","tokens_in":16209,"tokens_out":5522,"duration_ms":67516,"significance":"The clinical problem is well-motivated: CT-free craniosynostosis assessment requires dense correspondence from imperfect photogrammetry, and existing pipelines require manual preprocessing or fail on peripheral clutter. If the reported results hold, SCALP would be a practically valuable step toward automated, radiation-free longitudinal monitoring. The paper's strengths include a systematic comparison against six published baselines, internally consistent ablations on the semi-supervised component and the spectral warp, and a clinically meaningful evaluation protocol. The main load-bearing theoretical claim—that optimizing in a band-limited eigenbasis eliminates folding 'by construction'—is asserted without proof and, as stated, is not correct. In addition, the spectral truncation rank r is selected using the held-out test cohort, which compromises the reported generalization numbers. These issues are fixable, but they affect the central claims and require a revision.","major_comments":[{"comment":"The claim that optimizing within the first r Laplace–Beltrami eigenfunctions 'eliminates high-frequency mesh folding by construction' is not established. Eq. (7), Vdef = V0 + ΦW, constrains the displacement field to a low-dimensional subspace, but local injectivity requires det(I + ∇(ΦW)) > 0 at every vertex. A band-limited displacement with large coefficients can have arbitrarily large gradient and can fold the surface. Eq. (10)'s eigenvalue-weighted penalty is a soft regularizer, not a hard bound on the C1 norm, and the two-way Chamfer term can drive coefficients large under severe partiality. The paper should either prove a precise injectivity guarantee, state a weaker formal property (e.g., bounded displacement energy), or add a numerical check such as the minimum Jacobian determinant on all test subjects. Without this, the 'mathematical robustness guarantee' in Sec. 1 and the relate","section":"Sec. 3.3, Eq. (7)–(10); Sec. 1 (contribution bullet)"},{"comment":"The spectral truncation rank r=120 is chosen by evaluating Table 5's metrics on the 13 held-out test subjects. The main text states that test subjects are 'strictly excluded from all training, model selection, and pseudo-labeling,' yet the r ablation uses test-cohort metrics to select a hyperparameter. This is test-set leakage and inflates the reported SSM generalization/specificity numbers. The rank should be selected on a validation split (or via nested CV) and the final test numbers re-reported with r frozen. Additionally, the table text claims that increasing r from 64 to 200 steadily reduces P2S and Warp S2S, but the r=120 row shows P2S=1.478 and Warp S2S=1.456, which are worse than at r=100; please reconcile or correct this non-monotonicity.","section":"Supplementary Material, Table 5; Sec. 4 (test split description)"},{"comment":"The Stage I contribution is described as combining 'confidence-filtered pseudo-labeling' with a Mean Teacher consistency objective, but no pseudo-label loss or confidence threshold appears in the training objective. The only unsupervised term is Lcons on bottleneck features (Eq. 5). The paper never specifies how pseudo-labels are generated, filtered, or weighted. Since this is a central claimed contribution and is used in the ablations (which say 'disabling ... pseudo-labeling'), the full objective and the confidence-filtering procedure must be provided; otherwise the semi-supervised method is not reproducible.","section":"Sec. 3.2, Eq. (4)–(6)"},{"comment":"The comparison against unsupervised baselines may be skewed by the evaluation protocol. Baseline particles are estimated on raw scans and then filtered by a template bounding box after ICP alignment, whereas SCALP's particles are generated on the template ROI by construction. If the experiment is meant to demonstrate robustness to peripheral clutter, the baselines should be scored on the same cropped region in an identical manner, or the post-hoc ROI filtering should be justified as not discarding meaningful baseline correspondences. Currently the large P2S/LLE gaps in Table 2 could partially reflect this protocol difference rather than only anatomical awareness.","section":"Sec. 4.3 (comparison protocol)"}],"minor_comments":[{"comment":"The notation for L_offset appears to omit the square on the 2-norm: it is written as ||...||_2, not ||...||_2^2, despite being described as an offset regression loss. Please standardize.","section":"Eq. (3)"},{"comment":"The phrase 'often by multiple orders of magnitude' is inaccurate for the reported numbers: P2S 1.478 vs. 13.348 is about a 9x improvement, which is less than an order of magnitude. Rephrase to 'substantially' or 'roughly an order of magnitude.'","section":"Sec. 4.4"},{"comment":"Minor grammar: 'Fig. 2 summarizes' should be 'Fig. 2 summarizes' (subject-verb agreement is fine; the issue is elsewhere: 'summarizes' vs 'summarize' in context). More importantly, please label units in Table 2's SSM metrics (all in mm where applicable).","section":"Fig. 2 caption / Sec. 4.4"},{"comment":"The sharpening temperature τ in Eq. (2) is introduced but its role is not discussed. If τ is small, the softmax becomes peaked; please state the chosen value and whether it is fixed or annealed.","section":"Sec. 3.2"},{"comment":"The limitations paragraph acknowledges sensitivity to template mismatch and severe posterior missingness. This is honest and useful, but it directly undercuts the universal 'by construction' guarantee in Sec. 1. Please revise the language to match the actual scope, e.g., 'reduces high-frequency deformation' rather than 'eliminates folding by construction.'","section":"Sec. 5 / Limitations"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is clearly written and the empirical direction is promising, but I would not accept in its current form. The main theoretical claim is not proven and as stated is false; the r selection on the test set is a methodological error that could change the headline numbers if corrected; and the pseudo-labeling component of Stage I is unspecified. These are all addressable within a revision: provide a proof or numerical fold-free check, select r on validation, and fill in the training objective. If the authors can also clarify the comparison protocol, the paper could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know about SCALP is that it is a real step forward in a practical direction—automating SSM construction from raw, cluttered photogrammetry—but the paper's central robustness claim is not what it purports to be, and there is a hyperparameter leak that needs addressing.\n\nWhat's new: the orchestration of a semi-supervised landmark detector (PTv2 + Hough voting + Mean Teacher) with a landmark-anchored Laplace–Beltrami spectral warp for subject-specific template fitting. Prior spectral work used LBO for descriptors or generative synthesis, not this kind of direct fitting on imperfect scans. The empirical results are strong: on 13 held-out infants, SCALP brings P2S down from >13 mm to 1.48 mm and Warp S2S from >4 mm to 1.46 mm, and the SSM metrics (compactness 0.99, generalization 1.11 mm, specificity 9.6 mm) are far better than the unsupervised baselines. The qualitative mode analysis (PC1 head size, PC2 cephalic index, PC3 midface) lines up with clinical expectations. The ablation showing semi-supervised beats supervised-only, and spectral beats TPS, is informative.\n\nNow the soft spots.\n\nFirst, the 'mathematical robustness guarantee' against folding is asserted, not proven, and as stated it is false. Band-limiting the displacement field to the first r eigenmodes does not prevent local inversion failure: Vdef = V0 + ΦW can have det(I + ∇(ΦW)) ≤ 0 for large coefficients, even in the low-frequency subspace. The eigenvalue-weighted regularizer is a soft penalty; the two-way Chamfer term can drive coefficients large, especially under missing regions and template mismatch. The paper's own Limitations section concedes sensitivity to template deviation and severe posterior missingness. So either a proof or a numerically verified fold check is needed, and the 'by construction' language should be toned down.\n\nSecond, the spectral rank r is selected on the held-out test set. In the supplementary Table 5, the authors evaluate r ∈ {64,100,120,150,200} and pick r=120 based on test-set generalization/specificity. That is a leak. Model selection should be done on a validation fold, and only then should the test set be touched.\n\nThird, no code or data are released, and the AUC metrics for compactness/generalization/specificity are reported without error bars, so it is hard to know how stable the improvements are.\n\nThis is for researchers in craniofacial imaging and anyone building SSMs from raw surface scans. The central idea is sound and the empirical contribution is real. The paper deserves a serious referee, but it needs a proof or softening of the folding claim, a clean hyperparameter split, and ideally code/data release. I would send it to peer review with major revisions expected, and I would bring it to the reading group—there is good discussion material here about what 'by construction' can and cannot buy you.","headline":"Useful integration with dramatic empirical gains, but the 'no folding by construction' guarantee is unsupported and the rank r is tuned on the test set.","tokens_in":16707,"tokens_out":5205,"would_cite":true,"duration_ms":49351,"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":"The paper claims that a two-stage pipeline—semi-supervised landmark localization plus a band-limited Laplace–Beltrami template warp—constructs statistical shape models directly from raw, unsegmented infant photogrammetry scans, outperformin","keywords":["statistical shape modeling","photogrammetry","craniosynostosis","Laplace–Beltrami operator","landmark localization","semi-supervised learning","spectral deformation","dense correspondence"],"falsifier":"A decisive test: on a cohort of severe metopic or sagittal craniosynostosis patients with CT ground truth, measure whether the reconstructed photogrammetry surface captures the pathological keel or scaphocephalic elongation. If errors cluster at high-curvature pathological features and persist as the band limit r is increased without a matching improvement in shape-model quality, the low-frequency basis is not expressive enough. Also, re-running the pipeline with a deliberately different normative template (e.g., an adult head) and observing mesh collapse or gross misalignment would falsify th","tokens_in":15807,"feed_emoji":"👶","tokens_out":3831,"duration_ms":43418,"temperature":0.7,"pith_summary":"SCALP is a two-stage framework for turning raw 3D photogrammetry of infant heads into dense anatomical correspondence. Stage I learns 23 craniofacial landmarks from 50 manually annotated scans plus a larger unlabeled cohort. Stage II deforms a normative head template within its own low-frequency Laplace–Beltrami eigenbasis, anchored by those landmarks. The paper's central claim is that restricting deformation to low-frequency modes eliminates mesh folding by construction and isolates the cranium from neck, shoulder, and noise clutter, yielding shape models with much lower generalization and specificity errors than unsupervised point-cloud baselines. If correct, this removes the manual preprocessing bottleneck that has kept radiation-free photogrammetry from replacing CT in longitudinal craniosynostosis monitoring.","feed_headline":"Spectral warp builds head shape models from messy 3D scans","feed_subtitle":"Landmark-guided spectral deformation isolates the cranium from noise and clutter without CT or manual segmentation.","key_machinery":"The load-bearing object is the Laplace–Beltrami eigenbasis of the template mesh: the eigenvectors of its cotangent Laplacian, ordered by eigenvalue, from global shapes to fine detail. By parameterizing the deformation as V0 + ΦW with only the first r modes, the optimization space itself excludes high-frequency folding, and the eigenvalue-weighted regularizer biases the fit toward coarse cranial shape. The predicted landmarks pull specific template vertices to anatomically meaningful targets, tying the smooth spectral fit to clinical anatomy.","core_discovery":"SCALP argues that dense, anatomically consistent statistical shape models can be produced from imperfect, unsegmented surface scans without manual cleanup, by combining sparse clinical landmark guidance with a spectral deformation space. The landmark detector supplies anatomical anchors using a small expert-annotated set, and the spectral warp—optimizing only the first r eigenfunctions of the template's Laplace–Beltrami operator—acts as a hard geometric prior. This makes the fit robust to missing regions and peripheral clutter, and the resulting statistical model captures clinically meaningful modes of variation (overall head size, cephalic index, mid-facial shape) rather than scanner noise.","pith_inferences":["If the hard-constraint claim holds, the band limit r becomes a principled, population-specific hyperparameter: one could estimate the required spectral content from a CT-derived atlas, then apply it to photogrammetry, yielding a testable transfer rule.","The two-stage recipe may generalize to other craniofacial conditions (e.g., plagiocephaly, cleft lip/palate) or other anatomies wherever a normative template and sparse landmarks exist, suggesting a general pathway for radiation-free longitudinal shape monitoring.","The current evidence is geometric rather than clinical; direct validation would compare SCALP-derived severity scores against CT-based scores on the same patients, which the paper lists as future work."],"forward_implications":["Raw photogrammetry scans, without segmentation or manual preprocessing, can be converted into cranial-only correspondence particles ready for downstream statistical shape analysis.","A small labeled set (50 subjects) plus unlabeled scans suffices; the semi-supervised Stage I outperforms a matched supervised version on landmark accuracy and downstream shape-model metrics.","The band-limited spectral warp prevents high-frequency folding in the presence of severe missing data and clutter, a failure mode that spatial methods like thin-plate splines exhibit.","The resulting PCA shape space has high compactness, generalization, and specificity, with principal modes corresponding to clinically relevant head shape variations."],"fun_headline_variants":["SCALP: shape models from messy scans without CT","Landmark-anchored spectral warp handles cluttered 3D scans","SCALP: robust shape modeling from imperfect scans","Spectral warp + landmarks: shape models from noisy scans","No CT, no cleanup: SCALP builds head shapes from raw scans"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The pipeline depends most on the assumption that the template's low-frequency eigenbasis is expressive enough to represent every target anatomy in the cohort—including severe deformities like metopic craniosynostosis—and that fitting by Chamfer distance alone does not land in a wrong deformation basin; the paper itself notes that the warp becomes sensitive if the template deviates too far from the cohort.","fun_headline_variants_meta":{"raw":{"variants":["SCALP: shape models from messy scans without CT","Landmark-anchored spectral warp handles cluttered 3D scans","SCALP: robust shape modeling from imperfect scans","Spectral warp + landmarks: shape models from noisy scans","No CT, no cleanup: SCALP builds head shapes from raw scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000763,"raw_usage":{"total_tokens":3214,"prompt_tokens":725,"completion_tokens":2489,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":2402}},"tokens_in":469,"tokens_out":2489,"duration_ms":18217,"temperature":1.0,"reasoning_tokens":2402,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T01:02:49.305811+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test: on a cohort of severe metopic or sagittal craniosynostosis patients with CT ground truth, measure whether the reconstructed photogrammetry surface captures the pathological keel or scaphocephalic elongation. If errors cluster at high-curvature pathological features and persist as the band limit r is increased without a matching improvement in shape-model quality, the low-frequency basis is not expressive enough. Also, re-running the pipeline with a deliberately different normative template (e.g., an adult head) and observing mesh collapse or gross misalignment would falsify th","supporting_citations":[],"review_version":1}