{"id":"b2b86258-7e97-4e80-8d8b-b281b36852da","arxiv_id":"2608.09993","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"APCReg combines ordered projection alignment, overlap-gated learned matching, arch-structured hypothesis selection, and a test-time retention rule to improve CBCT-to-IOS dental registration, reporting 0.87 mm mean Chamfer distance on 60 jaw pairs.","lead":"This paper proposes APCReg, a coarse-to-fine system that registers dental cone-beam CT scans to intraoral surface scans by aligning multi-view projections and then refining with learned overlap-aware matching. The authors report a mean Chamfer distance of 0.87 mm on a 60-jaw held-out test set and claim first place across six metrics against open-source baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 91.5% inlier ratio may be largely explained by the 0.3 overlap threshold rather than by registration accuracy, since 89.9% of Predator's points already pass this same threshold.","rationale":"The reader's REJECT verdict is based on a concrete, suspicious metric inconsistency: RTE = 13.10 mm alongside CD = 0.87 mm and RMSE = 1.00 mm, with RTE defined as the norm of the translation component. I agree that this inconsistency is real and deserves scrutiny. However, I interpret it slightly differently: the large RTE likely arises from a subset of hard pairs with large translation errors, while the median pair has small CD; the paper itself reports a hard subset with RTE >= 20 mm or RRE >= 10 deg, so the distribution is bimodal. The reader's concern about mutual consistency is still the most load-bearing, but the more precise formulation is that the reported RTE is dominated by a few outliers while the headline CD/IR are dominated by the rest, and the paper does not report per-pair RTE distributions. My own additional concern is the IR metric's dependence on the overlap mask and the 2 mm threshold; since IR gives APCReg only a 1.6-point gain over Predator (91.5 vs 89.9), the first-place claim across all six metrics rests on a metric that is not fully specified. If the overlap masking and threshold are applied equivalently to all methods, the ranking may still hold, but the paper does not provide enough detail to verify this. I therefore agree with the reader's REJECT verdict (the evidence is insufficiently reproducible and internally inconsistent as reported), and I would not adjust the verdict, though I would add the IR-overlap dependence as an explicit audit step when the authors provide code or data.","tokens_in":11882,"tokens_out":1701,"duration_ms":16025,"concrete_test":"Re-run the evaluation code on the STSR 2025 test split using only publicly available metadata and transformations: compute the IR for each method both with the overlap mask derived from hand-labeled or challenge-provided GT overlap and with the trivial full-surface mask, and report the per-pair IR distribution. If Predator's IR drops below 70% when the overlap mask is replaced by the full IOS surface, or if APCReg's IR does not exceed Predator's by more than a few points under the original mask, then the IR-based first-place claim is not a robust margin and the headline should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The inlier ratio (IR) is defined as the fraction of GT-overlapping IOS points whose predicted point lies within 2 mm of the GT IOS surface. In this CBCT-IOS setting, after the coarse stage the IOS is roughly crown-aligned, so a substantial fraction of points may already be within 2 mm even for imperfect poses. Table 1 shows Predator at 89.9% IR, yet its RTE is 29.18 mm and CD is 1.22 mm. A translation error of 29 mm would move every point by 29 mm, so any definition under which Predator reaches 89.9% IR is not a pointwise criterion on the transformed IOS alone; it must effectively be a surface-distance criterion after dropping or weighting non-overlapping points. The contribution section reports only '91.5% inlier ratio' and the gain over Predator is a mere 1.6 percentage points, which is small relative to the CD and RRE improvements. The load-bearing metric for the headline claim is therefore IR, yet the text does not specify how GT overlap is computed, how the 2 mm threshold is applied, or how the surface is sampled for the denominator; without this, the 91.5% value cannot be reproduced or compared, and the claim that APCReg 'ranks first across the six reported metrics' is not falsifiable from the paper's evaluation protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes APCReg, a coarse-to-fine CBCT-to-intraoral-scan registration pipeline. A multi-view anatomical coarse registration (MACR) stage sequentially aligns buccal, proximal, and occlusal projections; an overlap-aware residual registration (OARR) stage refines the pose with KPConv features, an arch-length cue, and Sinkhorn matching; a dental-arch-structured hypothesis selection (DAHS) stage ranks candidate poses on held-out correspondences; and a coarse-retention guard (CRG) chooses at test time between the residual update and the identity. The method is evaluated on 60 held-out jaw pairs from the MICCAI STSR 2025 benchmark and is reported to rank first among the evaluated baselines on all six metrics, with CD 0.87 mm, HD 2.92 mm, RMSE 1.00 mm, RTE 13.10 mm, RRE 3.88°, and IR 91.5%.","tokens_in":12212,"tokens_out":7922,"duration_ms":96364,"significance":"If the reported results are correct, APCReg would be a practically valuable contribution: it combines classical anatomical projection constraints with learned residual matching, and the test-time retention guard is an interesting idea for avoiding harmful refinements. The paper has genuine strengths: comparison against a wide panel of open-source baselines on a shared benchmark protocol, clustered-bootstrap CI for the headline CD comparison, stagewise and order-stress ablations for MACR, a hard-case subset analysis, and a ground-truth audit of the CRG decisions. These are the right kinds of evidence for a registration paper. However, the central quantitative claim rests on the consistency and reproducibility of the reported metrics, and that is currently the weakest point of the manuscript.","major_comments":[{"comment":"Table 1 is internally inconsistent under the metric definitions stated in the 'Metrics' paragraph. RTE is defined as the norm of the translation component of the relative transform. For FPFH+RANSAC, RTE=205.62 mm while CD=6.03 mm; for OBB+PCA, RTE=352.31 mm while CD=10.76 mm. For a rigid transform with a translation component of hundreds of millimeters, every transformed IOS point is displaced by a comparable amount, so directed and bidirectional surface distances cannot be on the order of a few millimeters unless the ground-truth surface contains a matching copy of the IOS at that offset, which is not plausible for jaw anatomy. Similarly, Predator has RTE=29.18 mm and IR=89.9% under a 2 mm inlier threshold, and APCReg has RTE=13.10 mm with RMSE=1.00 mm. These values cannot all hold under the stated definitions. If some centering, normalization, or overlap masking is applied before computing the surface metrics, it must be stated precisely, because it changes the meaning of RTE and RMSE. Since every headline claim is a rank over these six metrics, the authors must correct the definitions, correct the numbers, or provide the exact evaluation code and per-pair transforms. This issue is load-bearing and currently blocks assessment of the main claim.","section":"Metrics paragraph and Table 1"},{"comment":"The inlier ratio is not reproducible as defined. The text says IR is 'the fraction of GT-overlapping IOS points whose predictions lie within 2 mm of the ground-truth IOS surface,' but it does not specify how GT overlap is computed, what threshold defines overlap, which points form the denominator, whether the 2 mm check is a one-sided nearest-neighbor distance over the full GT surface or only over the overlap region, or how the two jaws are pooled. Because Predator reaches 89.9% IR with RTE=29.18 mm, the definition must effectively exclude or heavily downweight non-overlap points; otherwise the reported number is impossible. The exact formula must be given, and the evaluation code should be released. This matters especially because APCReg's gain over Predator on IR is only 1.6 percentage points, so the 'ranks first across all six metrics' claim depends on this underspecified quantity.","section":"Metrics paragraph, IR definition"},{"comment":"Only the CD pairwise comparison receives a confidence interval and a pairwise win count. The other five metrics, including RTE, RMSE, HD, and IR, are reported as point estimates. With n=60 pairs and a hard subset defined after inspecting MACR's coarse errors, the claim that APCReg ranks first across all six metrics needs per-metric uncertainty quantification, pairwise win/loss counts for each metric, and sensitivity analysis to the hard-case selection rule. Otherwise the reported rank could be driven by a small number of difficult pairs or by threshold choices that are not visible in the table.","section":"Experiments, 'Comparison with Registration Baselines'"},{"comment":"The manuscript excludes 48 of the 179 pose-labeled cases because their ground-truth labels have det(R)=-1. This is a large fraction of the labeled data, and the paper does not report whether the excluded cases differ systematically from the retained ones, nor whether the official challenge uses the same exclusion rule. If the excluded labels follow a different coordinate convention, the exclusion may be justified, but the authors should show that the retained 131 cases form a representative split and that the test partition is not cherry-picked. This is relevant because the reported 60-pair test set is the entire basis for the headline comparison.","section":"Experiments, 'Dataset and split'"}],"minor_comments":[{"comment":"The 'Full APCReg' row includes CRG, while the cumulative ablation rows do not, so the drop from 1.29 mm (A+B+C+D+E) to 0.87 mm confounds the learned-scorer and CRG contributions. The text discusses this, but the table should make the ablation coupling explicit, for example by adding a 'Full without CRG' row.","section":"Table 3"},{"comment":"The row label 'Disrupted P–O–B+ICP' is not defined. The paper should state what disruption is applied and why this variant is a meaningful control.","section":"Table 2"},{"comment":"No code, checkpoints, or per-pair result files are provided. Given the metric inconsistency in Table 1, releasing the evaluation code and raw per-pair errors is essential for verification.","section":"Data and code availability"},{"comment":"Equation (7) ends with a comma in the displayed formula; the equation should end with a full stop or no punctuation.","section":"Eq. (7)"},{"comment":"The baseline algorithm 'OBB+PCA' in Table 1 has no citation or implementation description in the text, which makes the baseline hard to reproduce.","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"The central problem is the internal inconsistency of Table 1 under the stated metric definitions. I would ask the authors to provide the exact evaluation code and per-pair transforms before any further assessment. If the reported values cannot be reproduced with consistent definitions, the main ranking claim should be considered unsupported. No concerns about novelty or citation practices beyond the need for code release."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a better paper than the rejection you passed me. The reader's central contradiction is not actually one. RTE is the translation component of the relative transform at the CBCT origin. With a 3.9-degree rotation about an axis near the dental arch and an origin a couple hundred millimeters away, a 13.1 mm RTE is exactly what you'd expect while tooth surfaces stay sub-millimeter. Same for Predator: 29 mm RTE with 6.9-degree rotation and CD 1.22 mm is consistent with a distant origin. So the metric set is internally coherent, but it does mean RTE is origin-dependent and the paper never centers the data or discusses this. That is a real weakness in interpretation and comparability, just not a fatal inconsistency. The stress-test note makes the same error when it says a 29 mm translation would move every point by 29 mm; it ignores the rotation component.\n\nWhat is genuinely new: the ordered buccal-proximal-occlusal projection schedule, the arch-structured hypothesis selection, and the test-time coarse-retention guard. None is revolutionary on its own, but the integration is clean, and the ablations are genuinely informative: the order-stress test, the CRG audit separating harmful from useful refinement, and the hard-case analysis all show where the method works and where it does not. That is honest engineering validation.\n\nThe soft spots are reproducibility and the strength of the claim. The IR definition—how GT overlap is computed and how the 2 mm threshold is applied—is underspecified. The benchmark is controlled-access, and no APCReg code is released. The gain over Predator is modest: 0.35 mm in CD and 1.6 points in IR, though the CI for the CD gain excludes zero and the RRE improvement is large. The baselines are standard but not current; there is no comparison to other STSR challenge entrants beyond PointNetLK.\n\nNet: this deserves a serious referee. If I were the editor, I would send it out, expecting the authors to clarify the coordinate-frame issue, specify the IR protocol, and ideally release code. The method is plausible and the evaluation careful enough to spend referee time on. I disagree with the REJECT verdict.","headline":"A competent engineering paper whose apparent metric inconsistency dissolves once you account for the coordinate frame; the real issues are underspecified IR protocol and missing code, not a contradiction.","tokens_in":12740,"tokens_out":10241,"would_cite":false,"duration_ms":110352,"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":"APCReg, a coarse-to-fine CBCT–IOS registration pipeline, decomposes the alignment search into ordered anatomical projections, then conditionally accepts or rejects a residual refinement, reporting 0.87 mm mean Chamfer distance and first…","keywords":["CBCT–IOS registration","anatomical prior","coarse-to-fine","multi-view projection","point cloud registration","reliability control","dental surgical planning","MICCAI STSR 2025"],"falsifier":"Recompute all six metrics from the authors' released predicted transforms and ground-truth transforms using the paper's metric definitions. If the translation component of the relative transform is indeed 13.10 mm on average, a simple geometric calculation implies the mean Chamfer distance cannot be 0.87 mm; a discrepancy would settle that the reported metrics are not mutually consistent.","tokens_in":11701,"feed_emoji":"🦷","tokens_out":6547,"duration_ms":61654,"temperature":0.7,"pith_summary":"This paper introduces APCReg, a fully automatic pipeline for aligning cone-beam computed tomography (CBCT) scans with intraoral scans (IOS) of the jaw, a step needed for surgical planning. The central claim is that the alignment problem should be split into a coarse stage that uses anatomical structure—specifically three ordered 2D projections (buccal, proximal, occlusal)—to narrow the six-degree-of-freedom search, and a fine stage that refines with learned correspondences and then decides, without ground truth, whether to keep the refinement or fall back to the coarse pose. On 60 held-out jaw pairs from the MICCAI STSR 2025 benchmark, APCReg reports 0.87 mm mean Chamfer distance, 2.92 mm Hausdorff distance, and 91.5% inlier ratio, ranking first across all six reported metrics among the open-source baselines tested. If these numbers hold, the approach would offer clinicians a reliable automatic alternative to manual alignment.","feed_headline":"Anatomy-guided dental scan alignment hits 0.87 mm","feed_subtitle":"Fully automatic pipeline beats tested baselines on all six metrics of the MICCAI STSR 2025 jaw benchmark.","key_machinery":"The load-bearing mechanism is the staged decomposition of the rigid registration problem. MACR rasterizes the two surfaces into three orthogonal silhouettes and sequentially aligns them (Y-Z, then X-Z, then X-Y), each stage fixing a subset of the translation and rotation parameters, then refining with ICP. The residual stage, OARR, builds a soft transport matrix from KPConv features, overlap predictions, and a folded arch-length cue, and DAHS scores candidate poses on disjoint correspondences; CRG then chooses between the residual update and identity by comparing one-sided IOS-to-CBCT distances. The central compositional identity is \\(\\hat{T} = T_a T_c\\) with \\(T_a \\in \\{T_r, I\\}\\), letting the system discard a harmful refinement.","core_discovery":"On its own terms, the paper claims that CBCT–IOS registration is best solved by respecting dental-arch anatomy at every stage: MACR aligns the coarse pose through ordered orthogonal projections that each constrain a subset of the 6-DOF pose, OARR matches the remaining residual using overlap-gated cross-attention with a PCA-axis-reversal-invariant arch coordinate, DAHS selects among structured pose hypotheses using held-out correspondences, and CRG compares the refined pose with the coarse pose using a fixed geometric rule to avoid harmful updates. The final transform is the composition \\(\\hat{T} = T_a T_c\\). The evaluation on 60 jaw pairs shows first place on all six metrics versus the tested open-source baselines, with a 28.4% improvement in mean Chamfer distance over the strongest baseline, Predator.","pith_inferences":["A transferable design principle is to let a registration system abstain from refinement: the CRG rule could be appended to any point-cloud registration network as a zero-cost safety gate, since it only computes one-sided distances.","The multi-view projection decomposition is not limited to teeth; any anisotropic structure with a clear principal axis (e.g., long bones) could use the same ordered-silhouette strategy to reduce the search dimension.","The evaluation protocol's definition of RTE should be checked: if the translation component of the relative transform is really 13.10 mm while the Chamfer distance is 0.87 mm, then either the definition or the reported numbers need clarification before the ranking is interpreted."],"forward_implications":["If the reported results replicate, APCReg provides a fully automatic CBCT–IOS registration path that outperforms the tested classical and learned baselines, potentially removing a manual step in digital dental surgery planning.","The order of the projection stages is load-bearing: reordering buccal–proximal–occlusal increases mean Chamfer distance by about 0.75 mm, so the anatomical schedule itself is part of the method's correctness.","The coarse-retention guard (CRG) is a ground-truth-free safety mechanism: it improves mean Chamfer distance from 1.157 mm to 0.869 mm on the held-out set by rejecting refinements that degrade geometric consistency.","On the 17 hard cases where MACR initialization is poor, APCReg recovers 11 of them, indicating the residual stage can correct large initial errors rather than only polishing near solutions."],"supporting_citations":[{"why":"Supplies the MICCAI STSR 2025 benchmark dataset, train/validation/test splits, and the comparison protocol.","marker":"[Wang et al., 2025]"},{"why":"Predator, the strongest baseline, whose 1.22 mm CD and 89.9% IR are the direct comparison points.","marker":"[Huang et al., 2021]"},{"why":"PointNetLK, the differentiable SVD baseline whose reported 46.47-mm translation error motivates the coarse-to-fine separation.","marker":"[Aoki et al., 2019]"},{"why":"Prior automatic CBCT–IOS registration via deep pose regression, providing the clinical context for the task.","marker":"[Chung et al., 2020]"},{"why":"PointNet++ is used to segment IOS tooth surfaces, a required input preprocessing step.","marker":"[Qi et al., 2017]"},{"why":"nnU-Net is used to segment CBCT tooth surfaces, the other required input preprocessing step.","marker":"[Isensee et al., 2021]"}],"fun_headline_variants":["Anatomy-guided CBCT–IOS fusion reaches submillimeter accuracy","Dental-arch priors steer registration to 0.87 mm Chamfer error","Automatic jaw scan alignment: no manual error, 0.87 mm","Coarse-to-fine with anatomy priors beats baselines in jaw registration","Submillimeter CBCT–IOS matching via dental-arch guidance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim assumes that the reported evaluation metrics were computed exactly as defined, and in particular that a translation error of 13.10 mm is consistent with a pointwise surface error of 1.00 mm; if the metric definitions, the evaluation code, or the table entries do not match, the first-place ranking is not established.","fun_headline_variants_meta":{"raw":{"variants":["Anatomy-guided CBCT–IOS fusion reaches submillimeter accuracy","Dental-arch priors steer registration to 0.87 mm Chamfer error","Automatic jaw scan alignment: no manual error, 0.87 mm","Coarse-to-fine with anatomy priors beats baselines in jaw registration","Submillimeter CBCT–IOS matching via dental-arch guidance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000714,"raw_usage":{"total_tokens":3218,"prompt_tokens":961,"completion_tokens":2257,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":2158}},"tokens_in":577,"tokens_out":2257,"duration_ms":16739,"temperature":1.0,"reasoning_tokens":2158,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:50:56.675736+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute all six metrics from the authors' released predicted transforms and ground-truth transforms using the paper's metric definitions. If the translation component of the relative transform is indeed 13.10 mm on average, a simple geometric calculation implies the mean Chamfer distance cannot be 0.87 mm; a discrepancy would settle that the reported metrics are not mutually consistent.","supporting_citations":[{"cited_title":"Clinical Oral Investigations , volume =","cited_arxiv_id":null,"evidence_quote":"Supplies the MICCAI STSR 2025 benchmark dataset, train/validation/test splits, and the comparison protocol."},{"cited_title":"IEEE Transactions on Medical Imaging , volume =","cited_arxiv_id":null,"evidence_quote":"Prior automatic CBCT–IOS registration via deep pose regression, providing the clinical context for the task."},{"cited_title":"and Kohl, Simon A","cited_arxiv_id":null,"evidence_quote":"nnU-Net is used to segment CBCT tooth surfaces, the other required input preprocessing step."}],"review_version":1}