{"id":"897b8e6b-7330-4674-8d39-c451e67afe54","arxiv_id":"2505.07159","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A skull-stripping model trained entirely on randomly generated ellipsoid images achieves comparable accuracy to established methods across human, primate, and rodent MRI.","lead":"This paper trains a 3D U-Net for skull stripping using only synthetic ellipsoid images, with no real brain scans or human-created labels. The model reportedly matches established skull-stripping tools on human, monkey, and rodent MRI data, suggesting that anatomical priors may be replaceable by abstract shape priors.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'no anatomical priors' claim is contradicted by the explicit ellipsoid shape prior in Section 3.1; the method's prior-free status is untested.","rationale":"The reader identified the ellipsoid shape prior as the load-bearing premise, which is correct. My stress-test sharpens this into an internal inconsistency: the paper's own stated assumptions include an anatomical prior, directly contradicting the abstract's 'without any ... anatomical priors.' This is a concrete, falsifiable concern rather than a general worry about pathology. The missing numeric results and training instability noted by the reader are real but secondary; the conceptual contradiction cuts at the novelty claim and is easily testable. I agree with the CONDITIONAL verdict because the method's practical value does not hinge on being literally prior-free, but the paper must either remove the overclaim or demonstrate the ellipsoid prior is unnecessary. The proposed control experiment with non-ellipsoid shapes would settle the question definitively.","tokens_in":10242,"tokens_out":7112,"duration_ms":71788,"concrete_test":"Retrain PUMBA with the identical intensity ranges, noise, random deformation field, and boundary-label scheme from Section 3.2, but replace the ellipsoid shapes with non-ellipsoid connected shapes (e.g., superquadrics with random roundness parameters or randomly deformed spheres from smooth noise fields), preserving the inner/outer shell structure. Evaluate the retrained model on the same TCGA, MINDS, and LPBA40 benchmarks using the same post-processing and report Dice, Jaccard, and Hausdorff distances. If the metrics are statistically unchanged, the ellipsoid prior is not load-bearing and the 'no anatomical priors' claim is supported; if they degrade, the method depends on an anatomical prior and the abstract's claim must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Section 4.1 claim the model works 'without any real images or anatomical priors.' However, Section 3.1 states the method relies on the prior that a mammalian brain 'resembles a somewhat ellipsoid structure' and that the axial axis is shorter than the other axes. Section 3.2 then builds every training image from ellipsoids, with a hollow outer ellipsoid for the head and a smaller inner ellipsoid for the brain, plus a dedicated boundary label. This is an explicit anatomical prior, albeit a coarse one. The central novelty of the paper is that prior-free training from random shapes generalizes; if the ellipsoid prior is actually necessary for the reported accuracy, then the claim is false and the contribution reduces to testing a weak shape prior rather than no prior. The paper provides no experiment that varies the shape distribution while holding the intensity and deformation model fixed, so the necessity of the ellipsoid assumption is untested. Additionally, the post-processing described in Section 3.2 ('selecting only the center most largest chunk') encodes another prior about the brain being the largest central connected component, which is not discussed in the 'no priors' claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PUMBA, a skull-stripping model trained exclusively on purely synthetic images consisting of nested ellipsoids with random intensities, deformations, and artifacts, with no real brain images or labels in the training set. The method is evaluated on human, marmoset, macaque, mouse, and rat MRI datasets, including tumor-bearing cases, and compared qualitatively and quantitatively with BET, Synthstrip, and Unet-studio. The central claim is that the model achieves comparable accuracy without real images or anatomical priors.","tokens_in":10505,"tokens_out":4556,"duration_ms":42089,"significance":"The idea of training a segmentation model on purely synthetic, non-anatomical primitives is original and, if rigorously supported, would be a meaningful step toward generalizable brain extraction. The paper provides publicly available code and evaluates on external public benchmarks, which are strengths. However, the headline claim that the method uses no anatomical priors is directly contradicted by the method description, and the results are presented qualitatively rather than through numeric metrics, making the central claims unverifiable as written. With corrected claims and a proper quantitative evaluation, the work could become a useful contribution to the field.","major_comments":[{"comment":"The repeated claim that PUMBA works \"without any real images or anatomical priors\" is contradicted by §3.1, which introduces the prior that a mammalian brain \"resembles a somewhat ellipsoid structure\" and that the axial axis is shorter, and by §3.2, which builds every training image from nested ellipsoids and applies a post-processing rule that selects \"the center most largest chunk.\" These are explicit anatomical/geometric priors. Please either rephrase the claim to say \"without real images or species-specific anatomical labels\" or, preferably, provide an ablation that removes the ellipsoid and central-chunk priors and shows that the model still performs adequately; otherwise the headline claim is not supported.","section":"Abstract and §4.1"},{"comment":"The quantitative evaluation is presented only as plots: no numeric Dice, Jaccard, or Hausdorff values, standard deviations, or sample sizes are given in the text or tables. The statements that the method shows \"comparable results\" and \"the highest accuracy in marmoset data\" are therefore not verifiable. Please include a results table with per-dataset means and standard deviations, and ideally paired statistical comparisons against BET, Synthstrip, and Unet-studio.","section":"§4.1 and Fig. 3"},{"comment":"A grid search over image-generation parameters was used to select the final model, and §5 states that the intensity range had the largest effect on performance. The manuscript does not specify the data on which the grid search was evaluated. If real images were used to select the synthetic-data parameters, the claim of training without real data is compromised. Please state the validation protocol explicitly, and if real data informed parameter selection, re-run the evaluation with parameters fixed before seeing the test data.","section":"§3 (Methods) and §5 (Discussion)"},{"comment":"The authors acknowledge \"instability when training the same pipeline with different images generated by the same rules,\" but only a single trained model is reported, with no mention of random seeds, number of repeated runs, or variance of the resulting metrics. Because the training set is a fixed set of 3,000 synthetic samples (§3.3), the reported performance could be seed-dependent. Please report results across multiple training runs with different synthetic dataset draws to support the generalization claim.","section":"§5 (Discussion)"},{"comment":"The post-processing step that selects the \"center most largest chunk\" is itself an anatomical prior about the brain being the central largest connected component. Its contribution to the reported performance should be quantified, for example by an ablation that removes the boundary label and the selection step, and it should be explicitly listed among the method's priors when describing the method's assumptions.","section":"§3.2"}],"minor_comments":[{"comment":"The phrase \"regarding of species\" should read \"regardless of species.\"","section":"§3.1"},{"comment":"The word \"inifinite\" is a typo and should be \"infinite.\"","section":"§3.1"},{"comment":"The word \"affect\" in \"have a noticeable affect\" should be \"effect.\"","section":"§5"},{"comment":"The tool name \"Freeurfer\" should be spelled \"FreeSurfer.\"","section":"§2.1"},{"comment":"The dataset name \"BORDEUX24\" should likely be \"BORDEAUX24\" to match the NHP dataset reference.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is overstated: the method explicitly relies on an ellipsoid shape prior and a central-connected-component prior, so the 'no anatomical priors' statement is not defensible as written. The absence of numeric results is a serious omission for a journal submission. However, the underlying idea is publishable if the claims are recalibrated, a thorough quantitative evaluation with statistical tests is added, and the training stability and grid-search validation questions are addressed. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper is worth reading and should go out for review, but the headline claim is oversold. PUMBA trains a 3D U-Net for skull stripping on purely synthetic images built from random ellipsoids—no real MRI, no labels, no atlas. The idea is genuinely new relative to Synthstrip/Synthseg, which still start from human label maps. The results, at least qualitatively, look strong across human, marmoset, macaque, mouse, rat, and tumor-bearing scans. That combination is enough to make me want to referee it seriously.\n\nWhat the paper does well: the method is clearly described, the authors give exact intensity parameters in a table and point to code, they compare against BET, Synthstrip, and a template-based method, and they are honest about training instability and the failure of fine-tuning. Those are not the moves of people hiding problems.\n\nThe main soft spot is the claim, repeated in the abstract and Section 4.1, that the model works \"without any anatomical priors.\" That is not accurate. Section 3.1 explicitly assumes the brain resembles an ellipsoid with the axial axis shorter, and Section 3.2 builds every training image from a hollow outer ellipsoid (head) and an inner ellipsoid (brain), plus a boundary label. The post-processing step that selects the largest central connected component is another prior. So the contribution is not \"no priors\"; it is \"a very weak, generic shape prior is enough,\" which is still interesting. The stress-test note is right that the paper never tests whether the ellipsoid prior is necessary—an ablation with different shape distributions would settle it.\n\nThe other soft spot is quantitative. The figures show Dice, Jaccard, and Hausdorff, but the text reports no actual numbers, no confidence intervals, and no significance tests. \"Comparable accuracy\" is hard to assess without numbers. That is fixable, but it is a real gap.\n\nMinor: one evaluation set is a privately acquired mouse dataset, though the authors disclose acquisition details. Acceptable but limits reproducibility.\n\nMy take: the central idea is solid, the execution is decent, and the overstated claim is easy to correct. I would send this to peer review.","headline":"PUMBA's synthetic-ellipsoid training is a genuinely new idea that seems to work across species, but the 'no anatomical priors' claim is overstated and the paper needs numeric metrics.","tokens_in":11022,"tokens_out":2486,"would_cite":true,"duration_ms":24058,"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":"PUMBA shows that a model trained exclusively on synthetic images of random ellipsoids can strip brains from real MRI scans of humans, marmosets, macaques, mice, and rats with accuracy comparable to established methods.","keywords":["skull stripping","brain extraction","purely synthetic training data","multi-species segmentation","multimodal MRI","ellipsoid shape prior","unsupervised segmentation","generalizable segmentation"],"falsifier":"Benchmark the same synthetic-only training on a collection of brain MRIs with large tumors, severe edema, or marked atrophy and compare Dice to manual labels; if accuracy falls substantially below the baselines on those cases, the ellipsoid shape prior is the limiting factor.","tokens_in":10061,"feed_emoji":"🧠","tokens_out":9491,"duration_ms":87618,"temperature":0.7,"pith_summary":"PUMBA asks how little information a skull-stripping model actually needs, and its answer is: none from real images. The authors train a standard 3D encoder-decoder segmentation network (a U-Net) on 3,000 procedurally generated volumes, each built from a filled ellipsoid inside a hollow ellipsoid shell, painted with random Gaussian intensity patches, studded with small ellipsoid artifacts, and warped by a smooth deformation field. On real MRI covering healthy and tumor-bearing human brains, marmosets, macaques, mice, and rats, this synthetic-only model delivers accuracy comparable to established skull-stripping tools. That matters because collecting and annotating real training data is the main obstacle to generalizable medical segmentation, and because it suggests a recipe that may transfer to other segmentation tasks.","feed_headline":"Random ellipsoid images alone train a brain extractor across species","feed_subtitle":"A model trained only on synthetic ellipsoid volumes matches established tools on human, monkey, and rodent MRI.","key_machinery":"The load-bearing mechanism is the synthetic image generator rather than the network. Each training volume starts with two nested ellipsoids of random axis lengths — an outer hollow ellipsoid representing the head shell and an inner ellipsoid representing the brain — plus small ellipsoids and holes that stand in for imaging artifacts or anomalies. Each ellipsoid is split into patches with independent Gaussian intensity statistics, so the model cannot memorize any tissue-specific intensity pattern; a smooth random deformation field then breaks the perfect ellipsoid geometry. The three-class label scheme, brain, brain boundary, and background, lets a post-processing step keep only the largest central mass, which is the brain in normal and most pathological cases. This generator carries the argument because it is the only place where anatomical assumptions enter the pipeline.","core_discovery":"The paper's central discovery is that the prior knowledge needed for brain extraction can be reduced to three facts: a mammalian brain is a connected, roughly ellipsoidal mass; it sits inside a surrounding shell with an identifiable intensity boundary; and its axial extent is shorter than its other axes. From these assumptions alone, the authors train a simple 3D U-Net on 3,000 synthetic volumes, each an inner ellipsoid inside a hollow outer ellipsoid with random Gaussian intensities and smooth deformations, and report Dice, Jaccard, and Hausdorff performance comparable to established methods on healthy adult human T1, multi-modal human, glioblastoma and low-grade glioma, marmoset, macaque, mouse, and rat scans. They emphasize that the metrics were achieved without any image priors, supervised or unsupervised, during training. A dedicated brain boundary label is used at inference to select the largest central connected component, which reliably identifies the brain among other regions.","pith_inferences":["The ellipsoid prior has an unquantified boundary: brains with large space-occupying lesions, severe midline shift, or atypical anatomy may fall outside the training distribution, and the paper does not test how far the shape model can bend.","The same generator could be adapted to other approximately convex organs by changing the outer shell and the contrast rule, but the largest-central-mass selection would need rethinking for organs that are not the dominant object in the field of view.","Because the paper reports training instability across fixed synthetic datasets, generating fresh synthetic volumes on every training step is a natural test of whether the approach's ceiling is higher than reported.","A quantitative sweep over ellipsoid axis ratios, deformation strength, and intensity contrast would reveal exactly which part of the synthetic distribution controls generalization to real scans."],"forward_implications":["Skull-stripping pipelines could be built for new species or scanners without any real training images or manual labels, as long as the target is roughly ellipsoidal and contrast-bounded.","A single synthetic-only model can replace separate human, non-human primate, and rodent brain-extraction tools, simplifying preprocessing pipelines.","Training with procedurally generated volumes sidesteps privacy and annotation cost entirely, since no patient data is used.","The boundary-label trick gives a general mechanism for selecting a central target object when the segmentation model lacks anatomical context.","If the shape prior holds for other organs, the same recipe may transfer to segmentation tasks beyond the brain."],"supporting_citations":[{"why":"Supplies the random-intensity and deformation recipe that PUMBA adapts for its ellipsoid training volumes.","marker":"[36]"},{"why":"Provides the classic intensity-thresholding skull-stripping baseline that PUMBA is compared against.","marker":"[1]"},{"why":"Provides the template-based one-shot learning baseline, showing what an atlas-free approach must beat.","marker":"[10]"},{"why":"Supplies healthy multi-modal human MRI used to test modality invariance.","marker":"[12]"},{"why":"Supplies marmoset brain MRI data used to test cross-species performance.","marker":"[13]"},{"why":"Supplies adult human T1-weighted scans with ground-truth masks for the standard case.","marker":"[16]"},{"why":"Supplies glioblastoma and low-grade glioma scans used to test pathology invariance.","marker":"[17, 18]"},{"why":"Supplies mouse and rat brain MRI used to test rodent performance.","marker":"[14, 15]"},{"why":"Supplies macaque brain MRI used to test non-human primate performance.","marker":"[43]"}],"fun_headline_variants":["Skull stripping trained on pure synthetic ellipsoids","No real brains: synthetic training matches MRI tools","PUMBA: synthetic-only skull stripping across species","Ellipsoid shapes replace real anatomy for brain extraction","Brain extraction without real images, still accurate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that every mammalian brain, healthy or pathological, is well approximated for learning purposes by a filled ellipsoid with a contrast boundary against surrounding tissue; if a real brain departs from that shape enough, the synthetic-only training has nothing to fall back on.","fun_headline_variants_meta":{"raw":{"variants":["Skull stripping trained on pure synthetic ellipsoids","No real brains: synthetic training matches MRI tools","PUMBA: synthetic-only skull stripping across species","Ellipsoid shapes replace real anatomy for brain extraction","Brain extraction without real images, still accurate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000558,"raw_usage":{"total_tokens":2589,"prompt_tokens":815,"completion_tokens":1774,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":431,"completion_tokens_details":{"reasoning_tokens":1701}},"tokens_in":431,"tokens_out":1774,"duration_ms":12862,"temperature":1.0,"reasoning_tokens":1701,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:23:00.409396+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Benchmark the same synthetic-only training on a collection of brain MRIs with large tumors, severe edema, or marked atrophy and compare Dice to manual labels; if accuracy falls substantially below the baselines on those cases, the ellipsoid shape prior is the limiting factor.","supporting_citations":[{"cited_title":"Hoopes, J","cited_arxiv_id":null,"evidence_quote":"Supplies the random-intensity and deformation recipe that PUMBA adapts for its ellipsoid training volumes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the classic intensity-thresholding skull-stripping baseline that PUMBA is compared against."},{"cited_title":"Yeh, Brain mri segmentation using template-based training and visual perception augmentation, ArXiv (2023) arXiv–2308","cited_arxiv_id":null,"evidence_quote":"Provides the template-based one-shot learning baseline, showing what an atlas-free approach must beat."},{"cited_title":"URL:https: //brain-development.org/ixi-dataset/, [Online]","cited_arxiv_id":null,"evidence_quote":"Supplies healthy multi-modal human MRI used to test modality invariance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies marmoset brain MRI data used to test cross-species performance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies adult human T1-weighted scans with ground-truth masks for the standard case."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies macaque brain MRI used to test non-human primate performance."}],"review_version":1}