{"id":"4d215f43-edf5-45c8-b8c1-029e25c45d30","arxiv_id":"2504.19402","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A weight-space diffusion model trained on implicit neural representations of 452 real livers can generate realistic synthetic 3D livers, but the claimed downstream benefit to reconstruction is not tested.","lead":"This paper trains a diffusion model on the neural network weights that encode 3D liver shapes, then uses it to synthesize new liver shapes to supplement a small real dataset. The authors also find that only about 48% of automatically segmented liver objects in a public dataset are complete and clean enough to use.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central augmentation claim is never tested: no downstream reconstruction model is trained on real vs. real+synthetic livers, so 'improving accuracy' is asserted, not demonstrated.","rationale":"The reader's weakest assumption precisely identifies the gap I consider load-bearing: the paper's stated goal is to address data scarcity by supplementing real data with synthetic livers, but no downstream task is evaluated. The generation fidelity metrics and expert survey, even if fully valid, only show that the synthetic shapes resemble the training distribution. They do not establish that adding these shapes to a training set improves a downstream reconstruction model. This is not a mere presentation issue; it is the difference between the paper's title claim ('Boosting 3D Liver Shape Datasets') and the actual experiments (generating synthetic examples).\n\nI also note the unresolved inconsistency in the expert survey (Real/Fake/Not sure counts versus the statement that 'all were classified as real'), which further weakens the realism evidence. The reader did not overlook this; it was flagged as an 'unresolved inconsistency.' I agree that it is secondary to the missing downstream evaluation but worth fixing.\n\nThe verdict remains CONDITIONAL because the missing experiment is both necessary and easily addable: train a reconstruction model with and without synthetic augmentation on a fixed real training set and measure held-out accuracy. If the result is positive, the central claim is supported; if negative, the paper should be revised to a more modest fidelity-only claim. The reader's recommendation to condition acceptance on such an experiment is appropriate, so no change to the verdict is needed.","tokens_in":8020,"tokens_out":3779,"duration_ms":40193,"concrete_test":"Train a standard 3D reconstruction model (e.g., Occupancy Networks or a mesh regression network) on (i) the 452 usable real livers only, and (ii) the same 452 real livers plus a fixed set of N synthetic livers generated by the trained 3D Liver HyperDiffusion (e.g., N=100, 200, 400). Evaluate both on a held-out real liver test set that was not used in MLP or diffusion training (e.g., a separate hospital cohort or a held-out subset of TotalSegmentator). If reconstruction accuracy (IoU, Chamfer-L1, or similar) does not improve with synthetic augmentation across at least three random seeds, the abstract's claim that the method \"improve[s] the accuracy and reliability of 3D liver reconstruction\" fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and introduction promise that the synthetic liver dataset will \"improve the accuracy and reliability of 3D liver reconstruction\" and \"address the problem of data scarcity.\" Section 4, however, evaluates only generation fidelity (MMD, COV, 1-NNA, FPD) and an expert survey. No 3D reconstruction or segmentation model is ever trained on real-only versus real-plus-synthetic data. The synthetic shapes are shown to resemble the training distribution, but resemblance to training data does not imply usefulness for downstream tasks—especially when the real dataset itself is small and biased, and the generated shapes may simply reproduce or interpolate the same biases. The expert survey is also internally inconsistent: the text reports Real (139), Fake (4), Not sure (7), yet claims \"all were classified as real.\" This ambiguity weakens the only human-validated evidence of realism. The central claim that the generated shapes \"supplement the limited real dataset\" is therefore an untested assumption, and the paper's stated practical motivation remains unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper addresses data scarcity in 3D medical shape datasets, specifically liver shapes. The authors analyze the TotalSegmentator-derived liver objects (939 subjects) and, with expert surgical review, classify only 48.14% as directly usable, citing incomplete shapes and segmentation-to-3D conversion artifacts. To supplement the limited usable data (452 objects), they adopt the HyperDiffusion framework: first training an MLP as an implicit neural representation (INR) for each liver, then training a Transformer-based diffusion model over the flattened MLP weights to generate new weight vectors, which are decoded into novel liver meshes via Marching Cubes. They evaluate MLP reconstruction quality (VIoU 0.9747, Chamfer-L1 0.0028, NC 0.9780, F-Score 1.0) and compare the Transformer-based weight-space diffusion against a 1D UNet baseline on MMD, COV, 1-NNA, and FPD, reporting that the Transformer performs better. They also report an expert survey in which clinicians classified 150 real/synthetic livers, with outcomes Real (139), Fake (4), Not sure (7). The paper claims the synthetic livers are realistic and can improve 3D liver reconstruction accuracy and reliability, addressing data scarcity.","tokens_in":8224,"tokens_out":3947,"duration_ms":40704,"significance":"If the central claim were fully validated, this work would offer a practical recipe for expanding small, curated medical shape datasets using implicit neural representations and weight-space diffusion, which could benefit downstream reconstruction and segmentation tasks. The paper has concrete strengths: it performs a careful, expert-annotated dataset curation (452/939 usable with explicit failure categories), reproduces an established framework rather than introducing unvalidated machinery, and reports standard generation metrics on a held-out test set (MMD, COV, 1-NNA, FPD). The MLP reconstruction results in Table 1 indicate that the INR representation is of high fidelity. However, the paper's stated contribution—that the generated shapes 'improve the accuracy and reliability of 3D liver reconstruction' and 'address the problem of data scarcity'—is not actually tested: no downstream reconstruction or segmentation model is trained on real-only versus real-plus-synthetic data. In addition, the expert survey, the only human-judgment evidence of realism, contains a clear internal inconsistency. These gaps currently prevent the paper from supporting its main claims.","major_comments":[{"comment":"The abstract and introduction promise that the synthetic liver dataset will 'improve the accuracy and reliability of 3D liver reconstruction' and 'address the problem of data scarcity,' but Section 4 evaluates only generation fidelity (MMD, COV, 1-NNA, FPD) and an expert survey. No downstream reconstruction or segmentation model is trained on real-only versus real-plus-synthetic data, so the central augmentation claim is never demonstrated. The authors should add an experiment that trains a concrete downstream model (e.g., an occupancy-network-style reconstruction or a segmentation model) on the real dataset and on the real+synthetic dataset, and compare accuracy; alternatively, the paper should be re-scoped to a generation-only claim and the abstract revised accordingly.","section":"Abstract and Section 4 (Experiments)"},{"comment":"The expert survey results are internally inconsistent: the text reports Real (139), Fake (4), and 'Not sure' (7) out of 150 objects (75 real and 75 synthetic), and then states that 'all were classified as real.' Since 139+4+7=150, the claim that all were classified as real cannot be correct unless 'all' refers only to the real objects. The paper should report the full confusion matrix (e.g., how many of the 75 synthetic objects were classified as real, fake, or not sure), state how many experts participated and whether each object was rated by one or multiple reviewers, and use this corrected analysis to support or qualify the realism claim.","section":"Section 4, Expert Evaluation of 3D Liver Objects"},{"comment":"The generation results in Table 2 report single-point estimates with no variance or significance testing. The MMD difference between Transformer (0.24) and 1D UNet (0.25) is small, and the 1-NNA value of 53.68% is close to the 50% target, so it is unclear whether the Transformer's superiority is meaningful. The authors should report mean and standard deviation over multiple random seeds, and ideally a permutation test or confidence interval for the differences, before claiming that the Transformer-based model outperforms the 1D UNet across all metrics.","section":"Table 2 and Evaluation Metrics"}],"minor_comments":[{"comment":"In Eq. (1), the parameter vector is written as θ_i ∈ R, which should be a vector space such as R^d; this is likely a typo but should be corrected.","section":"Equation (1)"},{"comment":"The description of tokenization and the projection/merge operations for the 8 weight tensors (sizes [3456, 128, 16384, 128, 16384, 128, 128, 1]) is terse; a brief clarification of how the per-tensor linear projections are defined and why a sequence length of 8 is sufficient would improve reproducibility.","section":"Section 3.2"},{"comment":"The dataset analysis focuses entirely on TotalSegmentator-derived livers, but the introduction and Fig. 1 claim conclusions about 'current 3D medical shape datasets' such as MedShapeNet and SARAMIS. Please either provide quantitative evidence for the other datasets or restrict the claims to TotalSegmentator.","section":"Section 2"},{"comment":"The caption notes that MMD is scaled by 10^2, which is easy to overlook; consider using a unified notation, such as 'MMD (×10^2)' in the table header, to avoid reader confusion.","section":"Table 2 caption"},{"comment":"The paper does not report how many experts completed the survey, whether each object was reviewed by a single expert or multiple, or whether the experts were blinded to the real/fake ratio. These details are needed to interpret the survey as a validity check.","section":"Section 4"},{"comment":"The final sentence suggests diffusion models can be applied to other downstream tasks in 3D medical imaging, but this is speculative and not supported by any experiment in the paper; consider citing related work or removing the suggestion.","section":"Conclusions"}],"recommendation":"major_revision","confidential_remarks":"The paper's practical motivation is an augmentation claim, but the experimental section stops at generation fidelity. A downstream reconstruction or segmentation experiment is the natural and feasible fix. The expert survey inconsistency is also fixable but must be addressed carefully, as it is the only human-judgment evidence. If the authors instead choose to re-scope the paper to a generation-only contribution, the title and abstract would need to be revised and the downstream claim dropped. The paper does not mention code or dataset release, which would be useful for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: this paper applies HyperDiffusion to liver shapes and contributes a genuinely useful manual analysis of the TotalSegmentator liver data. The catch: the paper's stated motivation—that synthetic livers will improve 3D reconstruction—is never tested.\n\nWhat's new and good. The dataset analysis is the most solid part. They had a surgical team manually review 939 liver segmentations and found only 48% usable, with clear categories for the rest. That kind of ground-truth curation is valuable for the community. The MLP reconstruction scores are strong (VIoU ~0.97, Chamfer-L1 ~0.003), suggesting the INR stage is working. Comparing a Transformer-based weight-space diffuser against a 1D UNet on held-out test shapes is a reasonable baseline, and the Transformer does look better on the standard generation metrics.\n\nWhere it falls short. The abstract says the method will \"improve the accuracy and reliability of 3D liver reconstruction and generation,\" but no reconstruction model is ever trained or evaluated on real vs. real-plus-synthetic data. The experiments stop at generation fidelity. Showing that synthetic livers resemble the training distribution is not the same as showing they help a downstream task—especially when the real dataset is small and biased. This is a load-bearing gap.\n\nThe expert survey is also internally inconsistent: they report Real (139), Fake (4), Not sure (7) out of 150, but the text claims \"all were classified as real.\" That can't be right. Either the numbers or the sentence is wrong, and the paper needs to correct it. There are also no error bars or significance tests on the generation metrics, and no code or data released, which makes the comparison hard to trust.\n\nOverall. This is a modest application paper with a useful data-quality analysis and a plausible but incomplete evaluation. The central claim is unsupported, but it's the kind of gap that can be fixed by adding a downstream experiment—train a reconstruction or segmentation model on real-only versus real-plus-synthetic data and report the accuracy difference.\n\nRecommendation. I'd accept it for peer review, but the reviewers should insist on that downstream experiment and a corrected survey section before publication.","headline":"A useful liver dataset analysis and a clean HyperDiffusion baseline, but the paper's central claim—that synthetic livers improve downstream reconstruction—is never tested.","tokens_in":8746,"tokens_out":4091,"would_cite":false,"duration_ms":36668,"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":"A weight-space diffusion model over implicit neural fields can generate realistic synthetic 3D livers and expand a scarce medical shape dataset.","keywords":["3D liver generation","diffusion models","implicit neural representations","weight-space diffusion","dataset augmentation","occupancy networks","medical shape analysis"],"falsifier":"Train a 3D liver reconstruction or segmentation model on the 452 real livers alone and on the same set augmented with synthetic livers, then measure accuracy on held-out real CT volumes; if the augmented training set does not improve accuracy, the paper's motivating promise of boosting liver reconstruction is refuted.","tokens_in":7857,"feed_emoji":"🩻","tokens_out":8516,"duration_ms":83212,"temperature":0.7,"pith_summary":"Public 3D liver shape datasets are mostly not usable as-is: the paper finds that in a TotalSegmentator-derived collection only 48.14% of 939 liver objects are complete, artifact-free, and anatomically plausible. To work around this scarcity, it proposes 3D Liver HyperDiffusion, which encodes each usable liver as an implicit neural representation (a small network that predicts whether a 3D point is inside the liver) and then trains a diffusion model directly on the flattened network weights. New synthetic livers are produced by denoising random weights and extracting surfaces with Marching Cubes. The paper's central claim is that these synthetic livers are realistic and diverse enough to expand the real dataset, with expert reviewers overwhelmingly classifying them as real. The experiments verify generation quality; the promised downstream gain in 3D reconstruction accuracy is a stated motivation rather than a measured result.","feed_headline":"Diffusion on MLP weights yields realistic synthetic livers","feed_subtitle":"A weight-space diffusion model expands 452 curated liver shapes into a larger synthetic dataset for medical 3D imaging.","key_machinery":"The central object is a denoising diffusion model over flattened MLP weights. Each liver is first encoded as an implicit neural representation: a small fully connected network (three hidden layers of 128 ReLU neurons, positional-encoded 3D input) that predicts occupancy inside or outside the liver surface. The trained weight vectors are flattened into eight tensors, projected into a common token embedding, concatenated with a time-step token, and denoised by a Transformer; new weights are sampled with DDIM and decoded into meshes by Marching Cubes. The work this does is to turn 3D shape generation into a distribution-learning problem over weight space, avoiding explicit mesh or point-cloud generation.","core_discovery":"The central claim is that weight-space diffusion on implicit neural fields, the \"3D Liver HyperDiffusion\" model, can generate novel 3D liver shapes whose geometry is statistically close to real livers and that surgeons classify as real. The paper demonstrates this on 452 quality-controlled real livers: each liver is overfitted by a small occupancy MLP with positional encoding, the flattened weights are treated as diffusion data, and a Transformer-based denoiser learns to sample new weight vectors that decode into liver-shaped occupancy fields. Quantitative metrics (volumetric IoU near 0.975, Chamfer-L1 near 0.003, F-score 1.0 for the INR encodings; MMD, COV, and 1-NNA on the generated set) are reported, and an expert survey overwhelmingly classified the synthetic objects as real. The paper positions this as a scalable data-augmentation strategy: combine synthetic livers with the usable real ones to boost reconstruction and generation, though the downstream improvement is proposed rather than measured.","pith_inferences":["A direct, untested consequence is that the method's practical value depends on downstream reconstruction: adding synthetic livers to training data could help, hurt, or leave accuracy unchanged, and the paper does not measure which.","Because synthetic shapes come from denoising weight vectors, rare or extreme liver anatomies in the 452-shape training set will constrain the generated distribution; the method is unlikely to invent anatomies it never saw.","The same INR-plus-weight-diffusion recipe could transfer to other organs, but its success would likely depend on how many clean seed shapes exist for that organ and on how much shape variation the MLP can capture."],"forward_implications":["A hybrid dataset of the 452 curated real livers plus generated synthetic livers becomes available for training downstream 3D reconstruction and generation models.","The method offers a way to rescue public medical shape datasets whose raw segmentation exports are largely unusable: quality filtering to the usable fraction, then generative expansion.","Weight-space generation with occupancy fields naturally yields closed, volumetric shapes, avoiding some mesh-artifact issues of segmentation-to-mesh conversion.","Because the generative model is unconditional, a natural next step is conditioning on text or images, which the paper names as future work."],"supporting_citations":[{"why":"Computes the generalized winding numbers used to label sampled points as inside or outside the liver.","marker":"[1]"},{"why":"Supplies the HyperDiffusion weight-space diffusion framework that the paper adapts to liver MLPs.","marker":"[4]"},{"why":"Provides the denoising diffusion probabilistic model training objective underlying HyperDiffusion.","marker":"[8]"},{"why":"Marching Cubes is used to extract the liver surface from predicted occupancy values.","marker":"[14]"},{"why":"Introduces occupancy networks, the INR formulation used to encode each liver shape as an MLP.","marker":"[17]"},{"why":"DDIM is the sampler used to turn random noise into novel MLP weights.","marker":"[26]"},{"why":"The TotalSegmentator dataset is the source of the liver segmentations analyzed and curated.","marker":"[29]"}],"fun_headline_variants":["Weight-space diffusion yields realistic synthetic 3D livers","Diffusion on neural weights generates lifelike liver shapes","Synthetic livers via diffusion on implicit neural representations","Diffusion models expand 3D liver datasets with realistic shapes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that synthetic livers judged realistic by geometry metrics and expert review will also improve downstream 3D liver reconstruction accuracy when added to real training data, a benefit the paper states but does not test.","fun_headline_variants_meta":{"raw":{"variants":["Weight-space diffusion yields realistic synthetic 3D livers","Diffusion on neural weights generates lifelike liver shapes","Synthetic livers via diffusion on implicit neural representations","Diffusion models expand 3D liver datasets with realistic shapes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000404,"raw_usage":{"total_tokens":2088,"prompt_tokens":917,"completion_tokens":1171,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":1105}},"tokens_in":533,"tokens_out":1171,"duration_ms":10206,"temperature":1.0,"reasoning_tokens":1105,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:53:08.312320+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a 3D liver reconstruction or segmentation model on the 452 real livers alone and on the same set augmented with synthetic livers, then measure accuracy on held-out real CT volumes; if the augmented training set does not improve accuracy, the paper's motivating promise of boosting liver reconstruction is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Computes the generalized winding numbers used to label sampled points as inside or outside the liver."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the HyperDiffusion weight-space diffusion framework that the paper adapts to liver MLPs."},{"cited_title":"In: Proceedings IEEE Conf","cited_arxiv_id":null,"evidence_quote":"Introduces occupancy networks, the INR formulation used to encode each liver shape as an MLP."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DDIM is the sampler used to turn random noise into novel MLP weights."}],"review_version":1}