{"id":"4865290b-809a-4641-8963-df9b15b35258","arxiv_id":"2508.00155","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GEPAR3D combines a statistical-shape-model prior and 3D deep watershed energy maps to segment teeth and root apices in CBCT, reaching 95.0% average Dice across external test sets.","lead":"GEPAR3D is a deep learning system that segments every tooth, including narrow root tips, in 3D dental CT scans by combining a statistical shape model with watershed-based instance separation. It reports the best average Dice score of 95.0% across four test centers, which could support more reliable orthodontic root-resorption monitoring.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cui et al. appears in both training and external test sets with no documented split; until patient-level disjointness is shown, the +2.8 DSC margin cannot be attributed to generalization.","rationale":"Reader's CONDITIONAL verdict is appropriate. I looked for a stronger internal inconsistency and found none: the loss formulation, deep watershed adaptation, and ablation logic are coherent, and the qualitative figures support the root-apex emphasis. The load-bearing weakness is precisely the evaluation protocol. The text creates a direct ambiguity, with the same citation [7] in both training and test, and the comparison paragraph does not state that all baselines were retrained on the same split. This is not a disagreement with community consensus; it is a failure to document a condition necessary for the central numerical claim. A split audit is cheap because the paper promises released JSON and Zenodo annotations. If the audit shows overlap, the headline claim is invalid; if it shows disjoint IDs and matched baseline training, the CONDITIONAL can be upgraded. Hence I retain the reader's CONDITIONAL verdict and agree with the weakest-assumption analysis.","tokens_in":14587,"tokens_out":10561,"duration_ms":105222,"concrete_test":"Audit the released artifacts: download the test JSON referenced as the file IDs in the accompanying JSON and the Zenodo re-annotated training labels; enumerate scan identifiers for the 98 training scans and the 46 test scans, focusing on the Cui et al. subset. If any training ID appears in the test list, recompute Table 1's average DSC after removing the Cui et al. column and compare with SGANet's 92.2; if the margin disappears or reverses, the external-generalization claim fails. Independently of overlap, confirm that the training/validation split and each baseline's training data are stated in the paper; otherwise the evaluation protocol remains underspecified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 (Datasets and preprocessing) states that the method is trained and validated on a publicly available dataset of 98 CBCT scans [7] and immediately after that testing uses 46 CBCT scans from 4 centers, including the public datasets Cui et al. [7] and Tooth Fairy 2. No patient-level split, scan IDs, or validation partition is provided. The central claim, highest DSC across external test sets, requires the 46 test scans to be disjoint from all data used to train or validate GEPAR3D. As written, the manuscript does not establish this; if any of the 98 training scans appear in the Cui et al. test column of Table 1, that column is in-domain and the 95.0% average DSC, including the +2.8% margin over SGANet, reflects memorization rather than external generalization. The same evaluation paragraph says tooth-specific baselines are trained per their original protocols, so training data are not controlled across methods; a baseline trained on a smaller or different dataset could underperform for that reason alone. Together these facts make the reported superiority unverified, even though the method and ablation logic in Table 2 are internally coherent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GEPAR3D, a single-step 3D CBCT tooth segmentation method that combines a statistical shape model (SSM) prior encoded in a Wasserstein Dice loss with a 3D deep watershed instance regression branch. The model is trained on 98 publicly available CBCT scans from Cui et al., reannotated into 32 classes, and evaluated on 46 external scans from two in-house and two public centers. The authors report an average DSC of 95.0% (+2.8% over SGANet), a recall of 95.2%, and improved root apex delineation, with ablations attributing the gains to the geometric prior, energy map, and direction losses.","tokens_in":14786,"tokens_out":6175,"duration_ms":57283,"significance":"If the evaluation is validated, GEPAR3D appears to advance the state of the art for root-sensitive tooth segmentation, with a clear architectural contribution and a welcome emphasis on root apices. The planned release of code and reannotated labels is a strength, and the ablation study is logically structured. However, the current manuscript does not establish the external validity or the controlled comparison needed to support the headline claim.","major_comments":[{"comment":"The manuscript trains and validates on 98 CBCT scans from Cui et al. [7] and simultaneously lists Cui et al. [7] among the external test datasets, without providing scan IDs, a patient-level split, or a validation partition; this leaves the independence of the Cui et al. test column in Table 1 unverifiable, and if any of the 46 test scans overlap with the 98 training/validation scans, the reported 95.0% average DSC and the +2.8% margin over SGANet reflect in-domain memorization rather than external generalization. Please provide the exact scan IDs and a documented disjoint split, and report the external metrics separately for each center.","section":"Sec. 3 (Datasets and preprocessing), Appendix A, Table 1"},{"comment":"The evaluation is not controlled across methods because tooth-specific baselines are 'trained per their original protocols' while GEPAR3D and the general baselines are trained on the reannotated 32-class Cui dataset; a baseline trained on a smaller or differently labeled dataset can underperform for reasons unrelated to GEPAR3D's components, so the comparison does not establish that the proposed method is superior. Please retrain all baselines on identical training folds with identical preprocessing and annotation schema, or restrict the superiority claim to the controlled general baselines and report the training data size and protocol for every method.","section":"Sec. 3 (Evaluation details), Table 1"},{"comment":"The loss weights Lambda1=10, Lambda2=0.1, Lambda3=1e-6, the background penalty b=2, the quadrant penalty modifiers in Eq. (1), and the watershed threshold beta=0.5 are all hand-set, and no sensitivity analysis is reported; because the ablation gains are small (e.g., rows #6 and #7 differ by 0.05 DSC), the robustness of the central claim to these choices is not established. Please add a sensitivity analysis for the loss weights and penalty values, or justify them with a tuning procedure.","section":"Sec. 3 (Implementation details), Eq. (1)-(2), Table 2"},{"comment":"The ablation table does not state the dataset or number of test scans used, nor does it report the statistical test, sample size, or number of runs for the daggered p-values; without this information the small per-component differences cannot be evaluated. Please specify the ablation evaluation set and the statistical protocol.","section":"Sec. 3 (Table 2)"}],"minor_comments":[{"comment":"The phrase '+2.8% over the second-best method' is inconsistent with Table 1, where the second-best overall method (V-Net) is 93.5% DSC; the +2.8% margin is over SGANet, not the second-best overall. Please correct the headline or state that SGANet is the best tooth-specific baseline.","section":"Abstract and Sec. 1"},{"comment":"The captions contain 'T able 1' and 'T able 2'; these should be 'Table 1' and 'Table 2'.","section":"Table captions"},{"comment":"The statement 'file IDs in accompanying JSON' does not correspond to any JSON file in the manuscript; please include the file IDs or a link to them.","section":"Sec. 3 (Datasets and preprocessing)"},{"comment":"The URL 'zenodo.org/records/GEPAR3D' appears to be a placeholder; please provide the actual DOI or permanent link for the released annotations.","section":"Appendix A"},{"comment":"The notation p_i, \\hat{p}_i, and the exact form of W^M are introduced only in prose; please define all symbols explicitly at first use and clarify that W^M is computed for each voxel i over the class dimension.","section":"Sec. 2 (Eq. (2) and loss definitions)"},{"comment":"The expression O = 1/4 (G11 + G21 + G31 + G41) mixes coordinate pairs with scalar weights; please specify that the sum is vector-valued or clarify the indexing of the four central incisors.","section":"Sec. 2 (Geometric prior)"},{"comment":"The paper claims improvements at root apices but evaluates only whole-tooth metrics; please consider reporting a root-apex-specific metric (e.g., DSC restricted to the apical third) to substantiate the qualitative claims.","section":"Sec. 3 (Evaluation details)"}],"recommendation":"major_revision","confidential_remarks":"The train/test overlap is the decisive issue for this paper. If the authors cannot demonstrate patient-level disjointness between the 98 training/validation scans and the 46 test scans, the central claim of external generalization fails and rejection would be warranted. The baseline training protocol also needs to be controlled or the claims restricted accordingly. The methodological idea and ablation logic are otherwise plausible and worth a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: the method is worth reading, but the evaluation as written does not support the headline. GEPAR3D introduces a genuinely new configuration—an SSM-derived Wasserstein penalty matrix combined with a 3D deep watershed energy and direction regression—and the ablation in Table 2 shows each component contributes a modest, additive gain. The writing is clear, the method section is reproducible in principle, and the authors commit to releasing code and annotated labels. That is real credit.\n\nThe soft spot is load-bearing and exactly where the reader flagged it. The paper trains on Cui et al. [7] and also lists Cui et al. [7] as one of the external test sets, with no patient-level split or scan IDs provided. The appendix repeats this without clarifying disjointness. If any of the 98 training scans are in the test column, that column is in-domain and the average DSC of 95.0% is not external generalization. The comparison with tooth-specific baselines is also unfair: they are 'trained per their original protocols,' so SGANet's lower DSC could reflect training data rather than the method. The general baselines are trained in-house, but the tooth-specific ones are not, which muddles the +2.8% margin.\n\nA smaller but real concern: the Tooth Fairy 2 references [3,6] describe inferior alveolar canal segmentation, not tooth segmentation. The paper should clarify what ground truth is used for teeth in that set.\n\nThese issues are fixable. The authors can retrain all baselines on the same data, document the train/test split with scan-level IDs, and clarify the TF2 label derivation. The ablation logic suggests the method itself is sound; the external claim is just unverified. This deserves a serious referee, not a desk rejection, but the revision request should be substantive.\n\nFor a colleague: if you work in dental CBCT, this is a useful read because the method idea is clever and the evaluation pitfalls are instructive. I would not cite the +2.8% claim until the split and baselines are sorted out, but I would engage with the method itself.","headline":"A promising method with a clean ablation, but the external comparison is confounded by train/test overlap and protocol mismatch, so the headline DSC claim needs verification.","tokens_in":15409,"tokens_out":3290,"would_cite":true,"duration_ms":32026,"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 single-stage CBCT tooth segmenter that injects a statistical shape model of dentition and models each tooth as a 3D energy basin reaches an average Dice of 95.0%, with the largest gains at root apices.","keywords":["tooth segmentation","cone-beam computed tomography","statistical shape model","geometric prior","deep watershed","instance segmentation","root resorption","medical image segmentation"],"falsifier":"Retrain all five baselines on the exact same 98 scans with the same 32-class re-annotations and the same splits as GEPAR3D, then evaluate on a held-out set whose images are verified by pairwise registration to share no patient with training; if GEPAR3D's +2.8% average Dice margin over the second-best method does not persist, the claim is not supported.","tokens_in":14370,"feed_emoji":"🦷","tokens_out":7584,"duration_ms":65722,"temperature":0.7,"pith_summary":"GEPAR3D is a method for segmenting individual teeth in cone-beam CT (CBCT) scans, aimed at the fine root apices that matter for diagnosing orthodontic root resorption. The paper's central claim is that adding two ingredients to a single-stage segmentation network, a statistical shape model of normal dentition used to build a geometry-aware loss and a deep watershed branch that represents each tooth as a continuous 3D energy basin, improves both detection and segmentation. Across external test sets from four centers, GEPAR3D reports an average Dice of 95.0% (+2.8% over the second-best method) and a binary recall of 95.2% (+9.5%), with qualitative gains concentrated at root apices. If this holds, automated CBCT analysis can support reliable baseline segmentation for longitudinal root-resorption assessment.","feed_headline":"Geometry prior lifts tooth segmentation to 95% Dice","feed_subtitle":"Adding a dentition shape model and watershed energy maps raises recall by 9.5 percent.","key_machinery":"The load-bearing machinery is the pairing of a Geometric Wasserstein Dice loss with a deep-watershed instance-regression branch. The GeoWDL is built from an SSM-derived matrix of statistical inter-tooth distances, augmented by quadrant penalty modifiers that make within-arch confusions cheap, across-arch confusions expensive, and tooth-to-background errors most expensive; this matrix replaces the empirical dissimilarity penalties of the Generalized Wasserstein Dice Loss. The instance branch regresses two quantities per voxel: an energy value equal to the Euclidean distance to the nearest tooth boundary, so each tooth is a continuous 3D basin, and a unit vector pointing along the energy descent derived from a 3D Sobel operator. Watershed seeds are extracted from basin halves, the watershed transform separates instances, and per-instance class labels come from majority voting over the semantic segmentation. Together these components let the model resolve the narrow, tapering root apices where intensity contrast is low.","core_discovery":"On its own terms, the paper establishes that tooth segmentation in CBCT improves when the network is guided by a population-level anatomical prior and when instance separation is posed as an energy-basin regression rather than a separate detection stage. The geometric prior is a Statistical Shape Model of normal dentition from which pairwise inter-tooth distances are extracted and converted into a penalty matrix; the segmentation loss, called GeoWDL, then charges higher cost for confusing morphologically distant or across-arch teeth than for confusing adjacent or symmetric ones. The instance branch regresses a continuous energy map that encodes each voxel's distance to the nearest tooth boundary, plus a per-voxel energy descent direction, and the watershed algorithm on these maps produces separated tooth instances whose labels are assigned by majority voting from the semantic branch. GEPAR3D reports the best average multi-class Dice (95.0%), binary recall (95.2%), and instance detection accuracy (99.2%) against five baselines on external test data from four centers, with the clearest improvements on root apex boundaries.","pith_inferences":["Because the SSM prior is derived solely from normal adult dentition, GEPAR3D's margin may narrow on pediatric or severely maloccluded arches; a testable extension is rebuilding the prior from growth-stage or pathological samples.","The energy-direction loss is the component most directly aimed at apex boundaries, so an ablation that removes only the direction term while keeping energy regression and GeoWDL would isolate whether the binary-recall gain comes from watershed instance separation or from the geometric prior.","A static population-based penalty matrix cannot adapt to patient-specific anomalies; a learned or instance-adaptive geometric prior could push the same machinery further.","The reported binary recall gain suggests fewer false negatives in tooth tissue, which is what a baseline root-length measurement would need; quantifying root-length error on the released masks could translate the Dice gain into a clinical number."],"forward_implications":["Root apex segmentation becomes reliable enough that serial CBCT comparison can detect true root shortening rather than under-segmentation, supporting orthodontic root-resorption assessment.","Single-stage instance detection plus multi-class segmentation removes the error accumulation of coarse-to-fine pipelines, raising recall on thin structures without sacrificing precision.","The SSM-derived penalty matrix transfers across scanners: a model trained on one public center maintains high Dice on four external centers.","The released 32-class re-annotations and code provide a reproducible training and evaluation setup for future tooth-segmentation work.","Class-aware penalties from the geometric prior naturally handle third-molar class imbalance better than uniform cross-entropy, since confusions among similar molars are inexpensive."],"supporting_citations":[{"why":"Supplies the public CBCT training dataset (98 scans) and one of the external test sets, re-annotated into 32 classes.","marker":"[7]"},{"why":"Provides the statistical shape model of normal dentition used to derive inter-tooth distances for the geometric prior.","marker":"[16]"},{"why":"Source of the deep watershed transform, adapted here to 3D energy-basin regression for tooth instances.","marker":"[2]"},{"why":"Defines the Generalized Wasserstein Dice Loss whose dissimilarity penalties are replaced by the SSM-based geometric matrix.","marker":"[10]"},{"why":"Baseline competitor enforcing shape consistency via tooth skeletons; compared in the main evaluation.","marker":"[9]"},{"why":"Second-best baseline in the comparison; its graph-based adjacency constraints are contrasted with the unrestricted geometric prior.","marker":"[20]"},{"why":"Baseline whose learned dynamic adjacency is noted to be vulnerable to training-data bias; the reference for the binary-recall gain.","marker":"[22]"},{"why":"Supplies the Tooth Fairy 2 external test dataset used for cross-center generalization.","marker":"[3]"}],"fun_headline_variants":["Shape prior lifts tooth segmentation to 95% Dice","Geometry-guided watershed sharpens root segmentation","Statistical model boosts recall to 95.2%","Single-step learning beats two-stage tooth segmentation","GEPAR3D: shape prior achieves top segmentation accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported margins assume the comparison is fair, meaning every baseline is trained under conditions equivalent to GEPAR3D's and the public Cui et al. test set is genuinely external to the training scans, so that the +2.8% Dice difference reflects the method rather than a data or protocol advantage.","fun_headline_variants_meta":{"raw":{"variants":["Shape prior lifts tooth segmentation to 95% Dice","Geometry-guided watershed sharpens root segmentation","Statistical model boosts recall to 95.2%","Single-step learning beats two-stage tooth segmentation","GEPAR3D: shape prior achieves top segmentation accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000661,"raw_usage":{"total_tokens":3051,"prompt_tokens":1002,"completion_tokens":2049,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":1976}},"tokens_in":618,"tokens_out":2049,"duration_ms":19029,"temperature":1.0,"reasoning_tokens":1976,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:20:03.031216+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain all five baselines on the exact same 98 scans with the same 32-class re-annotations and the same splits as GEPAR3D, then evaluate on a held-out set whose images are verified by pairwise registration to share no patient with training; if GEPAR3D's +2.8% average Dice margin over the second-best method does not persist, the claim is not supported.","supporting_citations":[{"cited_title":"Nature Communications 13(1), 2096 (2022) 10 T.Szczepański et al","cited_arxiv_id":null,"evidence_quote":"Supplies the public CBCT training dataset (98 scans) and one of the external test sets, re-annotated into 32 classes."},{"cited_title":"Clinical Oral Investigations pp","cited_arxiv_id":null,"evidence_quote":"Provides the statistical shape model of normal dentition used to derive inter-tooth distances for the geometric prior."},{"cited_title":"In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)","cited_arxiv_id":null,"evidence_quote":"Source of the deep watershed transform, adapted here to 3D energy-basin regression for tooth instances."},{"cited_title":"In: MICCAI BrainLes Workshop 2017","cited_arxiv_id":null,"evidence_quote":"Defines the Generalized Wasserstein Dice Loss whose dissimilarity penalties are replaced by the SSM-based geometric matrix."},{"cited_title":"In: Information Processing in Medical Imaging","cited_arxiv_id":null,"evidence_quote":"Baseline competitor enforcing shape consistency via tooth skeletons; compared in the main evaluation."},{"cited_title":"IEEE Transactions on Medical Imaging41(11), 3116–3127 (2022)","cited_arxiv_id":null,"evidence_quote":"Second-best baseline in the comparison; its graph-based adjacency constraints are contrasted with the unrestricted geometric prior."},{"cited_title":"IEEE Transactions on Medical Imaging (2024)","cited_arxiv_id":null,"evidence_quote":"Baseline whose learned dynamic adjacency is noted to be vulnerable to training-data bias; the reference for the binary-recall gain."},{"cited_title":"IEEE Transactions on Medical Imaging (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the Tooth Fairy 2 external test dataset used for cross-center generalization."}],"review_version":1}