{"id":"8c6a9595-8b9f-4e13-b8d3-408e62035dc9","arxiv_id":"2508.03925","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A correspondence-preserving diffusion model generates ordered point clouds of hippocampi, keeping point-to-point anatomical correspondence across generated shapes.","lead":"This paper trains a diffusion model that generates point-based hippocampus shapes while keeping the same point index meaning the same anatomical location across all generated shapes. It shows the model beats point-cloud baselines on shape quality metrics and supports class-conditional and counterfactual generation for Alzheimer's studies.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that generated shapes preserve anatomical correspondences rests on ShapeWorks validity, which is never independently tested; L2 metrics and the colormap assume the correspondence.","rationale":"The reader identified the same load-bearing assumption: ShapeWorks-derived correspondences are anatomically valid, and preserving the training-data ordering preserves anatomical correspondence. I agree that this is the weakest link. The paper's own evidence for correspondence preservation is circular in part: L2-based metrics assume the correspondence rather than testing it, and the qualitative colormap is based on spatial location in the mean shape, not on independent anatomy. No internal inconsistency in the architecture or training objective was found; the model plausibly learns to preserve the index ordering it is trained on. The issue is external validity: if the training-data correspondence is wrong or unstable, the entire application to localized morphological analysis and counterfactual generation loses its meaning. The reader's other concerns (per-fold error bars, cosine schedule handling, no code release) are real but secondary. Since the reader already marked the verdict CONDITIONAL, and this concern does not by itself invalidate the method but does require additional validation, the verdict should remain CONDITIONAL. I therefore recommend UNCHANGED relative to the reader's verdict.","tokens_in":7609,"tokens_out":3706,"duration_ms":48184,"concrete_test":"Select a held-out set of real shapes and a matched set of generated shapes. Register an independent hippocampal subfield atlas (e.g., HippUnfold or a manual subfield segmentation protocol) to each subject and label each of the 512 points by subfield. For each point index i, compute the entropy of the subfield-label distribution across subjects, separately for real and generated shapes. If the mean per-index entropy is substantially higher for generated than for real shapes, or the majority-label agreement per index falls below (say) 80%, the claimed anatomical correspondence is not preserved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the model 'preserve[s] point correspondences that are present in the training data' (Abstract), and the downstream value (Figs. 5-6, Table 4) depends on those correspondences being anatomically meaningful. However, the correspondence itself is never validated. Section 3.1 states that ShapeWorks was used to build 512-point shape models, but no independent anatomical metric (e.g., hippocampal subfield overlap, registration-based correspondence, or manual landmark agreement) is reported. The quantitative evidence for correspondence preservation is Table 1's L2-based MMD/coverage/density, but those metrics are computed under the assumption that same-index points are corresponding; they cannot detect systematic correspondence errors. Figure 5's colormap is assigned from spatial location on the mean shape and then reused, so it only shows that indices are spatially consistent in the mean, not that they track the same anatomy across subjects. If ShapeWorks places particles on different subfields for some subjects, or if the model's generated orderings deviate from the training correspondence, the L2 metrics and the atrophy/counterfactual visualizations would be circular. This is load-bearing because the method's entire advantage over unordered point-cloud diffusion is anatomical interpretability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a correspondence-preserving diffusion model for point-based shape representations, targeting hippocampal shapes from OASIS-3. The model uses a PointNet-inspired U-Net with shared linear layers, learned correspondence embeddings that encode point index, and masked self-attention over a mean-shape kNN graph. Training minimizes the standard noise-prediction L2 loss on ordered point clouds. The authors compare against PCA, Luo et al. (2021), and Zeng et al. (2022) using MMD, coverage, and density under Chamfer, EMD, and L2 distances, report an ablation without correspondence embeddings, apply the method to amygdala data, and show conditional generation and counterfactual healthy-to-AD shape changes. The central claim is that generated shapes preserve the point correspondences present in the training data, enabling localized morphological analysis.","tokens_in":7895,"tokens_out":5628,"duration_ms":68700,"significance":"If the correspondence-preservation claim is validated, this is a useful contribution: existing deep point-cloud generators are permutation-invariant and cannot support point-index-based morphometry. The architecture is simple and parameter-efficient, and the ablation shows that the learned embeddings have a substantial effect. The appendix demonstrates transfer to another anatomical structure, and the downstream conditional-generation use case is timely. However, the strongest claims are currently supported mostly by metrics that assume the correspondence rather than test it, and the quantitative evaluation lacks uncertainty estimates. The paper is therefore promising but needs additional validation before the contribution is fully established.","major_comments":[{"comment":"The paper assumes that ShapeWorks-derived point correspondences are anatomically correct, but no independent validation is provided. All L2-based metrics and the qualitative colormap use same-index points as corresponding, so they cannot detect systematic correspondence errors. Please add an independent anatomical evaluation (e.g., overlap of point labels with hippocampal subfield parcellations, agreement with an alternative registration or manual landmark method, or ShapeWorks particle quality metrics) and show that generated samples' indices are consistent with this ground truth, not merely spatially consistent with the mean shape.","section":"Section 3.1, Section 3.4, Figure 5"},{"comment":"Results are reported as point estimates from a 5-fold cross-validation with no error bars, per-fold values, or significance tests. Given the variability likely present in these metrics, the claimed improvements over baselines and the hyperparameter choices may not be robust. Report mean ± std across folds and test whether differences are statistically significant, or at least provide per-fold tables.","section":"Table 1 / Section 3.1"},{"comment":"The ablation is interpreted as showing that correspondence embeddings are crucial, but no analysis is provided for why the ablation fails or whether the differences are statistically significant. Moreover, the L2 metrics used in the comparison are computed under the correspondence assumption; for the ablation, high L2 error simply means the model does not use index ordering, not necessarily that the generated shapes are anatomically wrong. Please provide a direct evaluation of generated correspondences (e.g., consistency of landmark indices across generated samples) and, if possible, a metric that does not presuppose correspondences.","section":"Section 3.3 / Table 1"},{"comment":"The hyperparameter discussion is internally inconsistent. The main results use the 'scaled linear beta schedule' (Table 1), but Appendix C concludes that the sigmoid schedule is better on most metrics (lower MMD under EMD/L2, higher coverage/density) and recommends it together with kNN=50. It is unclear why the reported main model uses the linear schedule. Also, hyperparameters appear to be selected using the same evaluation metrics on the test folds; please clarify the model selection procedure and avoid test-set selection bias.","section":"Appendix C / Table 3, Table 1"},{"comment":"The downstream conditional-generation claims are not backed by the reported numbers. Table 4 shows that data generated by the class-conditioned model yields lower AD accuracy (53.7% vs 63.4%) and lower F1 (0.67 vs 0.71) than real test data. The text says the generated data 'is able to capture the class differences,' which is overstated. Please provide details of the conditioning mechanism, the classifier and counterfactual guidance, and report statistical tests or more appropriate morphometric comparisons (e.g., overlap of atrophy maps between real and generated groups).","section":"Section 3.5 / Appendix D"}],"minor_comments":[{"comment":"The variance schedule β_t and the total number of timesteps T are not defined before Eq. (1); please define them explicitly.","section":"Section 2.1"},{"comment":"The colormap description is vague ('determined by its spatial location'). Specify how colors are assigned and note that this only visualizes spatial position, not anatomical label; the caption should not overclaim correspondence.","section":"Section 3.5 / Figure 5"},{"comment":"'epoches' should be 'epochs'.","section":"Table 2"},{"comment":"The default end beta value 0.0205 is described as 'very small'; please state the schedule range and whether this value was tuned on a validation set.","section":"Appendix C"},{"comment":"The citation for ShapeWorks appears first in Section 2.1 as 'Cates et al. (2017)', but Section 3.1 should also cite it directly when describing how the shape models were built. Also, the claim that 'no deep generative models have taken correspondences into account' should be checked against more recent literature.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The contribution is incremental but potentially useful for the shape-analysis community. The main risk is that the correspondence-preservation claim is asserted rather than demonstrated. I would encourage the editor to require the additional validation described in the major comments. The paper's fit with a medical-imaging or shape-analysis venue is appropriate, but the current evidence is not yet at the level expected for a full-length journal paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is the problem framing: point-based shape representations are ordered point sets, and the authors build a diffusion model that respects that ordering. That is a genuine gap in the generative shape-modeling literature. The architecture is sensible — PointNet-style shared weights, learned correspondence embeddings, masked attention over nearest neighbors on the mean shape — and the ablation makes a convincing case that the embeddings matter: without them, generation collapses. The quantitative comparison against Luo and Zeng shows clear wins on L2-based metrics, which is expected but still meaningful: the model actually preserves ordering, and that is the point. The appendix's amygdala transfer, even qualitative, suggests the method generalizes. The writing is clear and the method is simple enough to reproduce.\n\nThat said, there are real soft spots. First, no error bars or per-fold statistics anywhere. Five-fold cross-validation is performed, but only pooled numbers are reported. I cannot tell whether the MMD/coverage/density gaps between methods are within noise. That is a straightforward fix and should be required. Second, the paper overclaims when it says the model preserves \"anatomical correspondences.\" What the L2 distance actually verifies is that same-index points are spatially close in the data space. If ShapeWorks correspondences are systematically wrong for some subjects, the model will faithfully reproduce those errors. That is not a fatal flaw for the paper's stated goal — preserving training-data correspondences — but the authors should add a cautionary paragraph rather than leaning on the colormap visualization, which is not an independent anatomical check. Third, the hyperparameter appendix is internally inconsistent: Table 3 shows the scaled-linear schedule gives better MMD-CD and comparable density to the sigmoid, yet the text recommends sigmoid, and the main results (Table 1) actually use scaled-linear. This must be reconciled. Fourth, no code release, which is unfortunate for a methods paper of this simplicity.\n\nThe stress-test note worries about circularity because the model is trained and evaluated on the same ShapeWorks-derived data. I think that worry is overreaching if the claim is only about preserving the training representation; it would only bite if the paper claimed to validate ShapeWorks itself. The paper should be more careful with its language but does not need a new experiment. The reader's conditional verdict is fair: the central contribution is sound and useful, but the missing error bars and the overclaim on anatomical correspondence need attention.\n\nWho is this for? Researchers in medical image analysis who need generative models with point correspondences for morphometry, atrophy mapping, or counterfactual studies. It deserves a serious referee, not a desk reject, with the expectation of moderate revisions.","headline":"A solid, genuinely useful contribution to medical shape generation; the correspondence-preserving diffusion model works and the experiments support it, but the paper under-specifies uncertainty and overclaims on anatomical correspondence.","tokens_in":8347,"tokens_out":3103,"would_cite":true,"duration_ms":39390,"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":"This paper claims that a diffusion model which treats point order as anatomical information can generate realistic point-based shape representations that preserve the point correspondences present in the training data.","keywords":["point cloud generation","shape correspondence","diffusion model","point distribution model","hippocampus","Alzheimer's disease","conditional generation","counterfactual generation"],"falsifier":"Train the model on a synthetic population of shapes for which the true anatomical correspondence is known (for example, ellipsoids with labeled landmarks), and check whether each generated point index lands on the same true landmark; if the index-to-landmark mapping drifts, the preserved ordering is not anatomical correspondence.","tokens_in":7510,"feed_emoji":"🧠","tokens_out":9207,"duration_ms":102253,"temperature":0.7,"pith_summary":"Existing deep generative models treat point clouds as unordered sets, so they cannot generate shapes whose points correspond across subjects. This paper argues that for anatomical shapes such as the hippocampus, point order is information: an ordered list of 512 points can index the same anatomical location on every shape. The authors build a denoising diffusion model that generates ordered point sets, using learned per-index embeddings and attention over spatial neighbors to keep the indices aligned. On hippocampus data from an Alzheimer's disease cohort, the generated shapes are realistic and preserve the training correspondences, and conditional generation captures healthy-versus-disease shape differences. If correct, the method makes generative models usable for localized shape analysis, not just global object generation.","feed_headline":"Diffusion model outputs shapes whose points match across subjects","feed_subtitle":"Synthetic hippocampus point clouds keep each point's anatomical index, enabling localized Alzheimer's atrophy analysis.","key_machinery":"The central device is the correspondence-preserving diffusion model. It treats a shape as an ordered list of $N$ points, applies the standard Gaussian forward/reverse diffusion process from denoising diffusion probabilistic models, and uses a U-Net-like network in which convolutional layers are replaced by shared linear weights. Two components carry the correspondence argument: learned correspondence embeddings, one per point index, added to intermediate activations so the network can tell which index it is processing; and masked self-attention at the bottleneck, computed over the 50 nearest neighbors of each point in the mean shape, which lets spatial neighbors exchange information without","core_discovery":"On the paper's own terms, the discovery is that a denoising diffusion model operating on an ordered point list, rather than an unordered point cloud, can generate point-based shape representations that keep point correspondences across generated samples. The model's network is built from shared linear weights, with learned correspondence embeddings encoding each point's index and a masked self-attention layer over each point's nearest neighbors in the mean shape. With this architecture, the L2 objective at matching indices is enough to produce smooth, diverse hippocampus shapes, and the same point index consistently lands at the same anatomical location across samples. The paper demonstrates","pith_inferences":["If the preserved ordering really tracks anatomy, then every generated shape inherits a pointwise coordinate system, so one could compute vertex-level statistics over generated populations without re-registration; this is a direct but unstated consequence of the setup.","The method inherits the quality of the training correspondences; a natural stress test would be to train on shapes with a manually verified ground-truth correspondence and check whether learned index embeddings recover it, which the current experiments do not do.","The fixed k-nearest-neighbor graph from the mean shape could be extended to a per-sample or adaptive graph, which might improve generation for anatomies with large shape variation; this is a natural next step the paper does not explore.","The conditional and counterfactual results suggest a template for other longitudinal or disease studies: train on paired or labeled shape data and use index alignment to read off localized change vectors, rather than scalar volumes."],"forward_implications":["Generated hippocampus shapes can be used for localized morphological analysis: the mean healthy-to-AD difference computed from generated samples shows similar atrophy patterns to real data, so synthetic data can supplement small cohorts.","Conditional generation on class labels yields populations of healthy and AD subjects whose shapes are class-specific enough that a classifier trained on real data achieves comparable accuracy (F1 0.67 vs 0.71) on generated data.","Counterfactual generation maps a healthy hippocampus to its disease-progression counterpart while preserving global identity, providing subject-specific visualizations of atrophy.","Removing the correspondence embeddings sharply degrades quality, so the per-index embeddings, not just the ordered representation, carry the correspondence-preserving behavior.","The same architecture with fewer points (256) generates realistic amygdala shapes, indicating the method transfers to other anatomical structures beyond the hippocampus."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model framework (forward/reverse process, noise-prediction objective) that the method adapts to ordered point sets.","marker":"Ho et al. (2020)"},{"why":"Supplies the correspondence-based shape modeling approach used to construct the 512-point training shapes for each hippocampus.","marker":"Cates et al. (2017)"},{"why":"Provides the imaging cohort from which the hippocampus data are drawn.","marker":"LaMontagne et al. (2019)"},{"why":"Supplies the shared-linear-weights network design that the architecture modifies by adding correspondence embeddings and attention.","marker":"Qi et al. (2017)"},{"why":"Supplies the idea of positional embeddings in transformers, which the paper repurposes as learned correspondence embeddings for point indices.","marker":"Vaswani et al. (2017)"},{"why":"One of the two diffusion-based point cloud generation baselines; comparison shows its generated clouds lack correspondences.","marker":"Luo et al. (2021)"},{"why":"The other baseline (latent diffusion for point clouds); comparison shows its outputs do not preserve point correspondences.","marker":"Zeng et al. (2022)"},{"why":"Supplies the coverage and density metrics used to evaluate generation quality.","marker":"Naeem et al. (2020)"},{"why":"Motivates the downstream use of hippocampal shape correspondence for Alzheimer's disease and provides the shape representation construction approach.","marker":"Zhu et al. (2024)"}],"fun_headline_variants":["Diffusion model generates shapes with matching anatomical points","Correspondence-preserving diffusion for point-based shape generation","Hippocampus shapes from diffusion keep point indices aligned","Point correspondences preserved in generated shapes via diffusion","Diffusion model outputs shapes with consistent point labels"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The paper assumes that the 512-point correspondences on the training shapes mark the same anatomical location across subjects, so preserving the training data's point ordering is the same as preserving anatomical correspondence.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model generates shapes with matching anatomical points","Correspondence-preserving diffusion for point-based shape generation","Hippocampus shapes from diffusion keep point indices aligned","Point correspondences preserved in generated shapes via diffusion","Diffusion model outputs shapes with consistent point labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1018,"prompt_tokens":657,"completion_tokens":361,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":401,"completion_tokens_details":{"reasoning_tokens":286}},"tokens_in":401,"tokens_out":361,"duration_ms":4764,"temperature":1.0,"reasoning_tokens":286,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T01:00:54.494224+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the model on a synthetic population of shapes for which the true anatomical correspondence is known (for example, ellipsoids with labeled landmarks), and check whether each generated point index lands on the same true landmark; if the index-to-landmark mapping drifts, the preserved ordering is not anatomical correspondence.","supporting_citations":[{"cited_title":"ShapeWorks: Particle-Based Shape Correspondence and Visualization Software","cited_arxiv_id":null,"evidence_quote":"Supplies the correspondence-based shape modeling approach used to construct the 512-point training shapes for each hippocampus."},{"cited_title":"OASIS-3: Longitudinal Neuroimaging, Clinical, and Cognitive Dataset for Normal Aging and Alzheimer Disease","cited_arxiv_id":null,"evidence_quote":"Provides the imaging cohort from which the hippocampus data are drawn."},{"cited_title":"Attention is all you need","cited_arxiv_id":null,"evidence_quote":"Supplies the idea of positional embeddings in transformers, which the paper repurposes as learned correspondence embeddings for point indices."},{"cited_title":"Diffusion Probabilistic Models for 3D Point Cloud Generation","cited_arxiv_id":null,"evidence_quote":"One of the two diffusion-based point cloud generation baselines; comparison shows its generated clouds lack correspondences."},{"cited_title":"LION: Latent Point Diffusion Models for 3D Shape Generation","cited_arxiv_id":null,"evidence_quote":"The other baseline (latent diffusion for point clouds); comparison shows its outputs do not preserve point correspondences."},{"cited_title":"Reliable Fidelity and Diversity Metrics for Generative Models","cited_arxiv_id":null,"evidence_quote":"Supplies the coverage and density metrics used to evaluate generation quality."},{"cited_title":"Quantifying Hippocampal Shape Asymmetry in Alzheimer’s Disease Using Optimal Shape Correspondences","cited_arxiv_id":null,"evidence_quote":"Motivates the downstream use of hippocampal shape correspondence for Alzheimer's disease and provides the shape representation construction approach."}],"review_version":1}