{"id":"78bd823e-5eff-40f7-ae0b-581b6c1ee281","arxiv_id":"2501.13073","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CHaRM is an end-to-end deep learning method that conditions heatmap-based landmark localization on predicted tooth presence, achieving 0.56 mm error on standard dentitions and up to 14.8x faster inference than two-stage baselines.","lead":"A new deep learning pipeline detects dental landmarks directly on 3D tooth scans without first segmenting each tooth, using predicted tooth presence to guide the landmark search. It reports faster and, on complex cases, more accurate results than two-stage methods, and introduces a public dataset of 1,214 annotated dental models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The contribution of the CHaR conditioning (Eq. 6) is never isolated from the added classification head and BCE loss, so the reported gains cannot yet be attributed to the proposed module.","rationale":"The strongest claim of the paper is that CHaRM, and specifically the CHaR module, yields accurate and fast dental landmark localization. The experiments demonstrate that the full CHaR-based networks beat their vanilla backbones and the SOTA methods, but they do not demonstrate that the conditioning operation is what causes the improvement. Since the comparison in Section 5.4 changes three things at once—the classification head, the BCE loss, and the multiplication in Eq. 6—the central mechanism is confounded. This is more load-bearing than the TP-only metric issue: even if the evaluation protocol were perfect, the paper would still lack evidence for its stated contribution. The TP-only concern is real and worth reporting, but CHaRNet's higher F1 compared with TSMDL (Tables 8 and 9) means the TP-selection bias would tend to penalize CHaRNet, not inflate its advantage, so it does not undermine the comparative conclusion as directly. The missing ablation, by contrast, affects the attribution of every reported gain to the proposed module. A concrete three-arm ablation would settle the question: keeping the classification head and BCE loss while removing only the conditioning isolates the mechanism. Without that, the verdict should remain conditional, pending this evidence. I do not think the concern is fatal to the empirical system, but it prevents accepting the paper's central mechanistic claim as established.","tokens_in":19326,"tokens_out":9708,"duration_ms":105669,"concrete_test":"Run a three-arm ablation on the PointMLP backbone with identical data split, optimizer, epochs, and hyperparameters as Section 5.2: (A) heatmap MSE only (current 'Base'); (B) multi-task head and combined loss (Eq. 5) but without applying Eq. 6 at train or test—localize by argmax over the initial heatmaps and compute F1 from the classification head; (C) full CHaRM. Compare macro-averaged MEDE, MSR, and F1 on the test split, with a paired bootstrap or Wilcoxon test over dentition types. If (B) is within, say, 0.05 mm MEDE or 1 F1 point of (C), the conditioning itself is not the driver; if (C) is significantly better, the central claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.4 (Tables 5-7) compares each 'Base' backbone, trained only with the MSE heatmap loss (Eq. 3), against a 'CHaR' version that simultaneously adds (i) the 16-class presence classification head, (ii) the BCE loss (Eq. 4), and (iii) the multiplicative conditioning in Eq. 6. The observed gains—higher F1, lower MEDE, higher MSR—are therefore attributable to any or all of these three changes. The paper's central claim is that the conditioning operation itself, 'leveraging teeth-presence information,' drives the improvement, but no arm removes only the conditioning while keeping the classification head and BCE loss. Under the combined loss L = 0.001*L_MSE + L_BCE, the auxiliary task has a large weight, so it is entirely plausible that the gains come from multi-task regularization or from the extra capacity of the classification branch rather than from the CHaR multiplication. This is the load-bearing assumption: without this ablation, the 'novel module' is not empirically identified as the cause of CHaRNet's accuracy. The TP-only metric concern raised by the reader is valid and should be reported transparently, but it is secondary because CHaRNet also attains higher F1 than TSMDL, which would tend to make its TP-only MEDE harder, not easier.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"CHaRM is an end-to-end methodology for localizing five anatomical landmarks per tooth on 3D intraoral scans. The method uses a point-cloud encoder/decoder with a heatmap regression head, a teeth-presence classification head, and a CHaR module that multiplies each landmark's heatmap by the predicted presence probability of its tooth and weights an added null point by the absence probability. A new dataset, IOSLandmarks-1k (1,214 models), with a ten-class dentition taxonomy, is introduced. Experiments compare five backbones with and without CHaR and compare the best variant (CHaRNet, based on PointMLP) with TSMDL and ALIIOS. The paper reports a MEDE of 0.56 mm on standard dentitions and 1.12 mm macro-averaged across dentition types, MSR of 85.2% and 68.5%, and a 14.8x GPU speedup over TSMDL.","tokens_in":19615,"tokens_out":7613,"duration_ms":78094,"significance":"The dataset and the promised public release are valuable contributions to a field with scarce open data, and the efficiency gain from avoiding a segmentation stage is practically meaningful. The conditioning idea is simple and plausible, and evaluating five backbones gives evidence of general applicability. However, the central causal claim about the CHaR module is not yet empirically supported because the CHaR versions differ from the base models by more than the conditioning operation, and the evaluation metrics are true-positive-conditional with no uncertainty estimates. If the missing ablation and full-set metrics confirm the reported trends, CHaRNet would be a solid state-of-the-art result; as it stands, the strength of the conclusions exceeds the evidence.","major_comments":[{"comment":"The central claim that the CHaR module drives the improvements is not identified by the experimental design. In Tables 5–7, the 'Base' condition is trained only with the MSE loss (Eq. 3), while the 'CHaR' condition simultaneously adds the classification head, the BCE loss with λ_cls=1 and λ_reg=0.001 (Eq. 5), and the multiplicative weighting in Eq. (6). The gains could therefore be caused by multi-task regularization or by the extra capacity of the classification branch alone, rather than by the conditioning operation. An ablation that keeps the classification head and BCE loss but replaces Eq. (6) with an identity (or with ground-truth presence labels) is needed to attribute the improvement to the proposed module; without it, the paper's main novelty claim is unsupported.","section":"§4.2, §5.4, Eqs. (5)–(6)"},{"comment":"MEDE and MSR are computed only over true-positive landmarks, so landmarks that are missed (false negatives) never enter the localization error. Consequently, the reported MEDE of 0.56 mm for CHaRNet is a conditional quantity: it measures accuracy only on the subset of landmarks that the model detected. A model that detects fewer hard landmarks can look more accurate under this protocol. Reporting F1 alongside mitigates this only partially, and CHaRNet's higher F1 in the SOTA comparison makes its TP-conditioned errors relatively harder, but the headline numbers in the abstract are still not full-set localization errors. I recommend also reporting an error measure that penalizes false negatives (e.g., assigning FN a fixed distance penalty or computing MEDE over all landmarks of present teeth), and reporting confidence intervals or multiple-seed statistics, since all tables present single-split point estimates.","section":"§5.3, Eqs. (8)–(9)"},{"comment":"The Base-versus-CHaR localization comparison uses different rules for defining true positives. For base models, a tooth is considered absent if any of its five landmarks is predicted at the null point, which removes all five landmarks from MEDE/MSR; for CHaR versions, presence comes from the classification head and each landmark is evaluated independently. This asymmetry can distort Tables 6–7, because one misplaced landmark in a base model eliminates the other four (possibly accurate) landmarks from the localization statistics. The comparison should be repeated under a common TP rule, or with full-set metrics that do not depend on the heuristic presence criterion.","section":"§5.4, Tables 6–7"}],"minor_comments":[{"comment":"The text contains the typo 'CHarNet' in the introduction; it should be 'CHaRNet'.","section":"§1"},{"comment":"The text says CHaRNet reduces the micro-averaged MEDE to 0.80 mm, but Table 9 reports a micro-avg MEDE of 0.70 mm for CHaRNet; please reconcile the discrepancy.","section":"§5.5"},{"comment":"The discussion states that CHaR improves 'the encoding of local geometric information,' but CHaR reweights heatmap values rather than modifying the encoder; this sentence mischaracterizes the method and should be rephrased.","section":"§6"},{"comment":"The claim of being 'the first fully end-to-end deep learning approach' should be qualified, since c-SCN [7] is described in Section 2.2 as an end-to-end method for tooth segmentation and landmark localization, albeit on crowns only.","section":"§1 and §2.2"},{"comment":"The set-union notation in Eq. (6) is ambiguous for what should be an ordered heatmap vector; define the update by index (e.g., the final heatmap entries) so that the argmax in Eq. (2) is well-defined.","section":"§4.2, Eq. (6)"},{"comment":"The sentence describing training of CHaR-based networks with 'the MSE loss (Formula 3) and the combined loss (Formula 5), respectively' is ambiguous; clarify that base versions use Eq. (3) and CHaR versions use Eq. (5).","section":"§5.2"}],"recommendation":"major_revision","confidential_remarks":"The main technical weakness is the ablation confound in Section 5.4, which is fixable within the manuscript's scope. The dataset contribution is real and the efficiency comparison is useful. I would like the editor to ensure the authors provide the missing ablation and true-positive-inclusive metrics before acceptance; if those confirm the reported trends, the paper could be acceptable. The promise of public release is contingent on acceptance, which is normal but should be verified at production."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a serious empirical paper for dental landmark detection with a genuinely useful new dataset and a plausible conditioning mechanism. The CHaR module multiplies each landmark heatmap by the predicted probability that the corresponding tooth is present, and routes missing-tooth landmarks to a synthetic null point. That is a clean, segmentation-free way to handle partial dentitions, and the authors evaluate it across five point cloud backbones. They also introduce IOSLandmarks-1k (1,214 annotated models), which fills a real data gap in orthodontics. On that dataset, CHaRNet reports 0.56 mm MEDE on standard dentitions, 1.12 mm macro, and a 14.8x GPU speedup over the two-stage TSMDL. The writing is clear, the split is patient-level and stratified, and the experimental coverage is above average for the niche.\n\nThe main weakness is that the CHaR module is never isolated. The CHaR versions are trained with an added classification head plus a BCE loss weighted 1.0, while the base versions use only the MSE heatmap loss (weight 0.001). So the comparison in Tables 5-7 changes two things at once: the conditioning operation and the multi-task auxiliary loss. The gains attributed to CHaR could come from the auxiliary task or from extra model capacity. A proper ablation would keep the classification head and BCE loss and simply disable the multiplicative conditioning. Without that, the paper's central attribution is under-evidenced.\n\nA second issue is the evaluation protocol. MEDE and MSR are computed only over true-positive landmarks (Eqs. 8-9). The authors justify this because null-point distances would be meaningless, and they also report F1, so the omission is not hidden. Moreover, CHaRNet has higher F1 than the baselines, which makes its TP-only accuracy harder, not easier, to achieve. The protocol is defensible, but without confidence intervals or multiple seeds the numbers are point estimates. The single split also limits the strength of the claims.\n\nFinally, the framing oversells a bit: 'first fully end-to-end' is arguable given c-SCN, and 'several orders of magnitude faster' is true only against ALIIOS, not against TSMDL. Minor, but worth fixing.\n\nWho this is for: people working on dental/orthodontic image analysis or point cloud landmark detection. The dataset alone justifies a serious referee. I'd send it to peer review with a required ablation and a request for error bars or repeated seeds. Not a desk reject, and not an accept as is either.","headline":"Solid dental-landmarks paper with a real new dataset, but the headline CHaR module is never ablated from the multi-task head and loss, so the central attribution is under-evidenced.","tokens_in":20203,"tokens_out":4437,"would_cite":true,"duration_ms":43260,"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":"CHaRNet detects dental landmarks directly on 3D intraoral scan point clouds, reaching 0.56 mm mean error on standard models and avoiding tooth segmentation entirely.","keywords":["dental landmark detection","intraoral scans","point cloud deep learning","heatmap regression","conditioned heatmap regression","tooth presence classification","orthodontic treatment planning","IOSLandmarks-1k"],"falsifier":"Recompute MEDE and MSR over all existing landmarks, treating false negatives (landmarks predicted at the null point that actually exist) as errors at least as large as the largest observed distance, and re-rank CHaRNet against TSMDL and ALIIOS; if CHaRNet no longer leads, the reported accuracy gain is an artifact of the true-positive-only protocol.","tokens_in":19156,"feed_emoji":"🦷","tokens_out":3342,"duration_ms":36509,"temperature":0.7,"pith_summary":"The paper proposes CHaRM, a fully end-to-end deep learning methodology that localizes five dental landmarks per tooth directly on intraoral scan point clouds, eliminating the tooth-segmentation stage that previous methods require. The central idea is to condition the heatmap regression on predicted per-tooth presence probabilities, so landmarks of missing teeth are routed to a dedicated null point instead of hallucinating positions in the mesh. Evaluated on a new annotated dataset of 1,214 dental models, the best instantiation, CHaRNet, reports a mean Euclidean distance error of 0.56 mm on standard dentitions and 1.12 mm macro-averaged across all dentition types, while running up to 14.8x faster on GPU than the two-stage TSMDL baseline. A sympathetic reader would care because this removes a costly, error-prone preprocessing step and handles missing teeth in a single network.","feed_headline":"Dental landmark detection skips segmentation, hits 0.56 mm error","feed_subtitle":"Conditioning heatmaps on tooth-presence predictions lets one network handle missing teeth and run 14.8x faster.","key_machinery":"The load-bearing component is the Conditioned Heatmap Regression (CHaR) module, applied after the decoder and heatmap head. It takes the initial heatmap for landmark (t, g), the predicted presence probability p_t of tooth t, and a special null point placed outside the point cloud, then computes final scores as real-point likelihoods weighted by p_t and the null-point likelihood weighted by 1 - p_t. This switch-like weighting lets the network dynamically relocate landmarks of absent teeth to the null point while keeping landmarks of present teeth on the mesh, and it is trained jointly with a binary cross-entropy presence-classification head and an MSE heatmap regression loss.","core_discovery":"The paper claims that a simple multiplicative gating of heatmap scores by tooth-presence probabilities turns a generic point cloud encoder-decoder into a clinically useful dental landmark detector that works on incomplete dentitions. For each landmark associated with tooth t, the CHaR module multiplies the likelihoods of all real mesh points by the predicted presence probability p_t and multiplies the null-point likelihood by (1 - p_t), so the final heatmap maximum lands either on the tooth or on a fixed external placeholder. With PointMLP as the encoder, the resulting CHaRNet beats the segmentation-based TSMDL and ALIIOS pipelines on nearly every dentition subtype in localization error and success rate, and does so without segmenting teeth at inference time.","pith_inferences":["Because the reported MEDE and MSR are computed only over true-positive landmarks, a model that silently misses hard landmarks can appear more accurate than one that finds them with large errors; re-ranking the methods after counting false negatives as large errors would test how much of the advantage survives.","The same conditioning mechanism could transfer to other landmark-detection tasks on point clouds where individual parts may be absent, such as craniofacial or orthopedic keypoints, provided a null-point convention can be defined.","A direct ablation that feeds ground-truth presence labels instead of predicted probabilities into the CHaR module would isolate how much of the gain comes from the weighting rule versus the quality of the presence classifier."],"forward_implications":["Orthodontic workflows can obtain per-tooth landmarks without a separate tooth-segmentation stage, removing segmentation error propagation.","CHaRNet's GPU inference time of about 0.24 seconds per model makes real-time landmark placement feasible in treatment-planning software.","The CHaR mechanism improves five different point cloud backbones, so the conditioning idea transfers beyond a single architecture.","The public release of 1,214 annotated models gives the field a benchmark that includes missing teeth and third molars, which prior datasets largely lacked."],"supporting_citations":[{"why":"TSMDL is the two-stage segmentation-then-landmark baseline that CHaRNet must beat, and it supplies the main accuracy and inference-time comparisons.","marker":"[8]"},{"why":"ALIIOS is the other state-of-the-art landmark detection baseline, also requiring prior segmentation, and it provides the second accuracy comparison.","marker":"[9]"},{"why":"PointMLP is the point cloud backbone used in CHaRNet, the best-performing instantiation of CHaRM.","marker":"[22]"},{"why":"PointNet++ supplies the hierarchical set-abstraction encoder design that the point cloud backbones build on.","marker":"[20]"},{"why":"Heatmap-based landmark detection is the foundational approach that the paper adapts to 3D dental point clouds.","marker":"[51]"},{"why":"Teeth3DS+ is the only prior public dental landmark dataset, used to argue that IOSLandmarks-1k fills a data scarcity gap.","marker":"[61]"}],"fun_headline_variants":["End-to-end dental landmarks: no segmentation, 0.56 mm error","CHaRM: one network localizes teeth landmarks on 3D scans","Skip tooth segmentation: AI marks dental landmarks 14.8x faster","Conditioned heatmaps nail dental landmarks on missing teeth","Fully end-to-end tooth landmarking beats two-stage pipelines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation protocol measures localization accuracy only on landmarks the model correctly identified as present, so a model that misses difficult landmarks entirely can look more accurate than one that detects them with large errors.","fun_headline_variants_meta":{"raw":{"variants":["End-to-end dental landmarks: no segmentation, 0.56 mm error","CHaRM: one network localizes teeth landmarks on 3D scans","Skip tooth segmentation: AI marks dental landmarks 14.8x faster","Conditioned heatmaps nail dental landmarks on missing teeth","Fully end-to-end tooth landmarking beats two-stage pipelines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000635,"raw_usage":{"total_tokens":2975,"prompt_tokens":1036,"completion_tokens":1939,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":652,"completion_tokens_details":{"reasoning_tokens":1857}},"tokens_in":652,"tokens_out":1939,"duration_ms":12412,"temperature":1.0,"reasoning_tokens":1857,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:27:36.571534+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute MEDE and MSR over all existing landmarks, treating false negatives (landmarks predicted at the null point that actually exist) as errors at least as large as the largest observed distance, and re-rank CHaRNet against TSMDL and ALIIOS; if CHaRNet no longer leads, the reported accuracy gain is an artifact of the true-positive-only protocol.","supporting_citations":[{"cited_title":"Two-stage mesh deep learning for automated tooth segmentation and landmark localization on 3d intraoral scans","cited_arxiv_id":null,"evidence_quote":"TSMDL is the two-stage segmentation-then-landmark baseline that CHaRNet must beat, and it supplies the main accuracy and inference-time comparisons."},{"cited_title":"Automatic landmark identification on intraoralscans","cited_arxiv_id":null,"evidence_quote":"ALIIOS is the other state-of-the-art landmark detection baseline, also requiring prior segmentation, and it provides the second accuracy comparison."},{"cited_title":"Joint training of a convolutional network and a graphical model for human pose estimation","cited_arxiv_id":null,"evidence_quote":"Heatmap-based landmark detection is the foundational approach that the paper adapts to 3D dental point clouds."}],"review_version":1}