{"id":"03abc278-0aa3-46d9-afc0-35d9d6907bf9","arxiv_id":"2608.09182","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A prototype-guided iterative refinement model achieves small but consistent accuracy gains over baseline methods for cephalometric, fetal heart, and intrapartum ultrasound landmark localization.","lead":"This paper presents PPOC-LL, a lightweight deep-learning model that finds anatomical landmarks in X-ray and ultrasound images by iteratively correcting an initial guess using small image patches and learned prototype patterns. The authors report slightly better accuracy than prior state-of-the-art methods on three medical datasets, with a similar or smaller model size.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"State-of-the-art claim rests on unverifiable baseline comparisons: Table 1 is corrupted and the stated 'ResNet18 backbone for all models' likely handicaps native-architecture baselines like HRNet, SHG, and DSNT.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the superiority claim depends on fair and accurately transcribed baseline comparisons. My read adds specificity about why 'ResNet18 as backbone for all models' is not neutral for HRNet, SHG, and DSNT, and notes the absence of significance testing for the close margins. Because these issues are unresolved but not evidence of internal inconsistency, the appropriate verdict remains CONDITIONAL, matching the reader's assessment. I found no additional concern that would shift the verdict to ACCEPT or REJECT: the method is described in sufficient detail, ablations support the contribution of each component, and the reported gains are internally consistent with the design. The concrete test of re-running native baselines would settle whether the state-of-the-art claim survives fair comparison.","tokens_in":9143,"tokens_out":2566,"duration_ms":26753,"concrete_test":"Run the original official implementations of HRNet, SHG, DSNT, and NFDP on the same CE-Test1/CE-Test2 splits and FHU test split with their native architectures and identical preprocessing, then recompute Table 1's MRE and SDR values. If any native baseline matches or beats PPOC-LL, the 'state-of-the-art' claim must be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that PPOC-LL surpasses the state of the art on CE and FHU, with MRE gains over NFDP of only 0.05 mm, 0.03 mm, and 0.23 px. These differences are small enough that comparison fairness and transcription accuracy are load-bearing. The manuscript states 'ResNet18 is used as the backbone for all models' (Section 3, Implementation Details), but several baselines are not backbone-plus-head models: HRNet is a multi-resolution parallel network, SHG is a repeated hourglass stack, and DSNT has a specific fully-convolutional coordinate regression design. Rebuilding them on ResNet18 changes their inductive biases and can materially reduce accuracy; no code or configuration files are provided to show the adaptations were neutral. The provided Table 1 is visibly corrupted, with duplicated labels and values (e.g., repeated 'PPOC-LL' cells and repeated SDR columns), so the exact compared numbers cannot be verified from the manuscript. In addition, no significance testing or error bars for baselines are reported, so the 0.05 mm advantage over NFDP on CE-Test1 could be within run-to-run variation. The method itself has internal support (ablations, three-seed averaging, plausible parameter count), but the external superiority claim is under-supported until the comparison protocol is verifiable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PPOC-LL, a prototype-learning-based progressive offset correction framework for medical landmark localization. The method combines a multi-scale dynamic perception strategy for coarse-to-fine patch refinement, similarity-driven prototype learning for offset prediction, and an error-aware reliability regularizer that supervises per-step matching confidence. Experiments are reported on two public datasets (cephalograms, CE; intrapartum ultrasound, IUGC) and one private fetal heart ultrasound dataset (FHU). The authors claim that PPOC-LL achieves MREs of 1.09 mm, 1.43 mm, and 7.34 px on CE-Test1, CE-Test2, and FHU, respectively, surpassing the previous state-of-the-art NFDP while maintaining a favorable model size. Ablations on CE attribute consistent gains to the progressive refinement paradigm and the reliability regularization.","tokens_in":9570,"tokens_out":3699,"duration_ms":32571,"significance":"The proposed architecture is parameter-economic and the design of progressive offset correction with prototype-based local matching is well motivated and internally coherent. The ablation study in Table 2 shows monotonic improvements from each proposed component, which supports the internal validity of the method. The evaluation spans two imaging modalities and three datasets, and the reported model-size/accuracy trade-off in Fig. 3 is attractive. However, the central claim of state-of-the-art performance currently rests on comparison tables whose formatting is corrupted in the provided version, and on baseline adaptations whose fairness is not documented. If the tables are corrected and the comparison protocol is made verifiable, this would be a solid and practical contribution to medical landmark localization.","major_comments":[{"comment":"Table 1 is severely corrupted in the provided manuscript: the PPOC-LL rows contain repeated labels and duplicated numeric values across the SDR columns (e.g., 'PPOC-LL PPOC-LL PPOC-LL PPOC-LL 1.09 1.09 1.09 1.09 0.87 88.49 88.49 88.49 88.49 ...'), and the NFDP row shows the same value repeated multiple times (e.g., '7.57 6.62 6.62 6.62 6.62 22.19 22.19 22.19 22.19 32.83 51.93 66.85'). Similar duplication artifacts also appear in Table 2 (e.g., '1.26 1.26 1.26 1.26 82.14 82.14 ...') and Table 3 (e.g., '1.09 1.09 1.09 1.09 0.87 0.87 0.87 0.87 88.49 ...'). Because the paper's central claim is the reported superiority over NFDP by margins as small as 0.05 mm in MRE, the exact compared numbers must be verifiable. The authors should provide a clean, correctly formatted version of all tables, and ideally release the underlying per-run results.","section":"Section 3, Table 1"},{"comment":"The statement 'ResNet18 is used as the backbone for all models' is problematic for several baselines. HRNet is a high-resolution multi-branch architecture, SHG is a stacked hourglass network, and DSNT is a fully convolutional coordinate regression design; these are not backbone-plus-head architectures, and replacing their native encoders with ResNet18 changes their inductive biases and can materially affect accuracy. Since the reported gains over NFDP are only 0.05 mm, 0.03 mm, and 0.23 px on CE-Test1, CE-Test2, and FHU, comparison fairness is load-bearing. The authors need to describe exactly how each baseline was adapted to ResNet18, provide configuration files or code, or use the baselines' original architectures for the comparison.","section":"Section 3, Implementation Details"},{"comment":"No significance testing or seed-wise variance is reported for the comparison in Table 1. The text says PPOC-LL and ablation variants were repeated with three random seeds and averaged, but the small values shown next to the MREs are described as 'standard deviation of radial errors on the test set' (i.e., across test images), not across training runs. For the baseline methods, no per-seed results or variance measures are provided at all. A MRE difference of 0.05 mm between PPOC-LL and NFDP on CE-Test1 could plausibly fall within run-to-run variation. The authors should report per-seed breakdowns or appropriate significance tests (e.g., paired tests or bootstrap confidence intervals) for the main comparisons.","section":"Section 3, Implementation Details and Table 1"}],"minor_comments":[{"comment":"The word 'progressivly' is a typo and should be 'progressively'.","section":"Section 2, first paragraph"},{"comment":"The phrase 'centered at the xt' should be 'centered at xt' (remove the definite article before the variable) for formal clarity.","section":"Section 2.1, paragraph 3"},{"comment":"The caption says 'The best results are shown in bolded'; this should be 'shown in bold'.","section":"Section 3, Table 1 caption"},{"comment":"The explanation that 'the small values shown in the right-down corner indicate the standard deviation of radial errors on the test set' is ambiguous, because in Tables 1–3 these values appear immediately after the MRE in the same cell or column (e.g., '1.09 0.87') rather than in a corner. Please clarify the notation or provide a separate column for standard deviation.","section":"Section 3, Implementation Details"},{"comment":"The formatting corruption that affects Table 1 also appears to affect Tables 2 and 3, with repeated numeric values across SDR columns. These tables should be reformatted so that each cell contains a single value.","section":"Section 3, Table 2 and Table 3"}],"recommendation":"major_revision","confidential_remarks":"The corrupted table formatting may be a rendering artifact of the arXiv PDF-to-text conversion, but as presented in the submitted manuscript the main comparison table is unverifiable. The baseline adaptation protocol is a substantive fairness concern, especially because the reported margins over NFDP are very small. I would ask the editor to require the authors to provide a clean version of all tables, detailed baseline adaptation descriptions, and code/configuration files as supplementary material before the paper can be considered further."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before reading: the method is a sensible integration of known building blocks and the ablations support every component, but the headline \"state of the art\" claim rests on baseline comparisons I cannot verify. Table 1 as provided is corrupted—duplicated labels and cells—and the paper says ResNet18 was used for all models, which for HRNet, SHG, and DSNT is not a neutral choice. Those are not backbone-plus-head architectures; re-implementing them on a ResNet18 base changes their design.\n\nWhat's actually new: the specific combination of multi-scale dynamic patch sampling with a coarse-to-fine schedule, prototype-based similarity matching for offset prediction, and the tolerance-based reliability regularizer. None of these is new in isolation, but the integration is clean and the ablation shows each piece helps: MRE goes from 1.37 to 1.26 mm on CE. The gamma sensitivity table is honest—performance stays in 1.09–1.16 mm across gamma 2–10, which is a good sign the regularizer is not a knife's-edge trick. Three-seed averaging for their own method is better than usual.\n\nThe soft spots are exactly where the stress-test lands. The gains over NFDP are tiny: 0.05 mm on CE-Test1, 0.03 mm on CE-Test2, 0.23 px on FHU. Those margins are within plausible run-to-run variation, and there are no significance tests or baseline error bars. The corrupted table alone prevents me from trusting the exact numbers as printed. The baseline adaptation protocol is the load-bearing question: if HRNet and SHG were rebuilt on ResNet18 without their native multi-resolution or hourglass structures, their reported numbers may be unfairly low. Without code, configs, or a detailed adaptation description, the central claim is under-supported.\n\nNone of this is fatal to the method itself. The internal evidence is coherent, the loss design is not circular, and the writing is straightforward. The paper's value is the parameter-efficient progressive refinement recipe with a decent validation across three datasets. If the authors release code and clarify the baseline setups, I'd believe the SOTA claim; as-is, I'd treat it as plausible but unverified.\n\nWho this is for: people working on medical landmark localization, especially cephalometrics and fetal ultrasound, who want a lightweight alternative to cascade ensembles. The reliability regularizer idea might transfer to other coordinate regression tasks.\n\nRecommendation: send it to peer review. A serious referee can ask for the corrected table, the baseline details, and significance testing. The method deserves that round of scrutiny, and the subfield would benefit from a verifiable comparison.","headline":"The method is a clean integration of known ideas with honest ablations, but the state-of-the-art claim rests on baseline comparisons that are currently unverifiable: a corrupted Table 1 and a likely unfair 'ResNet18 for all models' protocol.","tokens_in":9971,"tokens_out":1944,"would_cite":false,"duration_ms":18057,"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":"PPOC-LL, a compact landmark-localization network, claims to beat the previous state of the art on three medical-imaging benchmarks by correcting landmark positions stepwise through prototype-guided patch matching.","keywords":["landmark localization","prototype learning","progressive offset correction","multi-scale feature pyramid","cephalometric landmarks","fetal ultrasound","reliability regularization","heatmap regression"],"falsifier":"Re-run the two best methods on the same test images with several random seeds and the paper's stated training settings; if the reported CE-Test1 mean radial error difference between PPOC-LL and the previous best does not reproduce within about 0.03 mm, the ranking claim is not stable.","tokens_in":1666,"feed_emoji":"🎯","tokens_out":1979,"duration_ms":60030,"temperature":0.7,"pith_summary":"The paper tries to show that costly multi-stage refinement for medical landmark localization can be replaced by a compact single-network pipeline that corrects landmark positions step by step. It introduces PPOC-LL, which first produces rough landmarks from a feature pyramid, then iteratively samples small patches around each current estimate and reads off a displacement from prototype-guided matching. On two public and one private dataset, the authors report mean radial errors of 1.09 mm, 1.43 mm, and 7.34 px, each better than the previous best method. The practical point is that accurate landmarks drive cephalometric, intrapartum, and fetal-heart measurements, and a smaller model is more plausible for clinical use. The authors present the gain as a favorable balance of accuracy and model complexity rather than a large accuracy jump.","feed_headline":"Step-by-step landmark correction beats previous best","feed_subtitle":"Compact CNN with prototype matching tops X-ray and ultrasound landmark benchmarks while staying small.","key_machinery":"The load-bearing object is the re-centered, multi-scale local patch paired with a per-landmark learnable prototype. The patch gives the model a changing receptive field as refinement proceeds, from coarse to fine, and the prototype supplies a semantic reference that makes local matching robust to texture changes and anatomical similarity. The offset is not regressed directly; it is the expectation of a fixed offset field under the softmax of cosine similarities between the prototype and patch features. The third mechanism is the error-aware reliability regularizer, which converts the current Euclidean error into a soft target and supervises the peak of the matching distribution, so confidence is tied to actual localization error.","core_discovery":"The central claim is that progressive offset correction with prototype-based local matching offers a better trade-off than global single-pass heatmap regression or heavy cascaded refinement. In PPOC-LL, each refinement step re-centers a patch on the current predicted location, selects a pyramid level and patch size from a hand-set radius schedule, computes a cosine-similarity map between the patch features and one learnable prototype per landmark, softmax-normalizes it, and updates the coordinate by the expectation under a fixed offset field. A reliability regularizer predicts the current error from the peak of the matching distribution, which stabilizes training. The authors report PPOC-LL as best or second-best across all tested thresholds, with lower standard deviations than the previous best method and the best fully-supervised result on the IUGC benchmark.","pith_inferences":["The prototype-matching readout is not tied to the three tested anatomies and could be applied to other tasks with repetitive local structure, such as spine or retinal keypoints, as a direct extension of the same expectation-over-offset-field idea.","The predicted reliability from the regularizer is a natural candidate for an uncertainty or quality-control signal at deployment, though the paper does not evaluate it in that role.","The hand-set radius schedules could be replaced by a learned search-range policy; the reported insensitivity to iteration count suggests such a policy would mainly help early steps.","Because the reported gains over the previous best are near 0.05-0.23 units, an independent reimplementation with identical evaluation protocol is the quickest way to confirm the ordering; the larger margin may be parameter economy rather than a big accuracy lead."],"forward_implications":["On the CE, FHU, and IUGC benchmarks, the stated mean radial errors are lower than the previous best method's, so if the numbers hold, PPOC-LL is the best reported fully-supervised landmark localizer on these test sets.","The accuracy plateau at five or six refinement steps means the schedule is not sensitive to extra iterations, so deployment can choose the step count by compute budget.","The small mean-error differences and lower standard deviations imply at least as much of a practical advantage from consistency and model size as from raw accuracy.","On IUGC, PPOC-LL uses no unlabeled data yet is reported comparable to semi-supervised methods, suggesting the refinement mechanics recover some of the benefit of additional data.","The tolerance parameter for the reliability regularizer is stable across a wide range, so the method does not require precise per-dataset tuning of that term."],"supporting_citations":[{"why":"The state-of-the-art fully-supervised baseline NFDP, whose mean radial errors PPOC-LL claims to beat on CE, FHU, and IUGC.","marker":"[11]"},{"why":"APR, a cascaded refinement baseline whose two-stage design motivates the progressive offset-correction idea.","marker":"[29]"},{"why":"Provides the cephalometric benchmark dataset, its official split, and the 19-landmark annotation protocol used for evaluation.","marker":"[25]"},{"why":"Provides the IUGC intrapartum ultrasound benchmark, its landmarks, and the MRE and angle-of-progression evaluation protocol.","marker":"[2]"},{"why":"Multi-Reg, a global-and-local encoder baseline against which the patch-based progressive scheme is compared on CE.","marker":"[14]"},{"why":"DSNT, a soft-argmax numeric regression baseline included in the FHU comparison.","marker":"[21]"},{"why":"HRNet, a high-resolution representation baseline used to situate PPOC-LL's accuracy-versus-parameter trade-off.","marker":"[26]"},{"why":"CenterNet, a point-detection baseline included in the CE comparison as a regression-type anchor.","marker":"[31]"}],"fun_headline_variants":["Prototype-based progressive correction shrinks landmark error","Compact CNN with prototype offsets tops landmark benchmarks","Progressive offset correction: accurate, lightweight landmark fit","PPOC-LL: prototype-driven refinement for precise landmarks","Small model, big precision: prototype-guided landmark shifts"],"cache_read_input_tokens":12032,"weakest_assumption_plain":"The central result depends on the comparison methods in Table 1 being set up fairly and their reported numbers being correct, since the headline gains over the previous best are as small as 0.05 mm on one test set.","fun_headline_variants_meta":{"raw":{"variants":["Prototype-based progressive correction shrinks landmark error","Compact CNN with prototype offsets tops landmark benchmarks","Progressive offset correction: accurate, lightweight landmark fit","PPOC-LL: prototype-driven refinement for precise landmarks","Small model, big precision: prototype-guided landmark shifts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000302,"raw_usage":{"total_tokens":1717,"prompt_tokens":899,"completion_tokens":818,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":744}},"tokens_in":515,"tokens_out":818,"duration_ms":7425,"temperature":1.0,"reasoning_tokens":744,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:16:06.818579+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the two best methods on the same test images with several random seeds and the paper's stated training settings; if the reported CE-Test1 mean radial error difference between PPOC-LL and the previous best does not reproduce within about 0.03 mm, the ranking claim is not stable.","supporting_citations":[{"cited_title":"Medical image analysis31, 63–76 (2016)","cited_arxiv_id":null,"evidence_quote":"Provides the cephalometric benchmark dataset, its official split, and the 19-landmark annotation protocol used for evaluation."},{"cited_title":"Neurocomputing470, 182–189 (2022)","cited_arxiv_id":null,"evidence_quote":"Multi-Reg, a global-and-local encoder baseline against which the patch-based progressive scheme is compared on CE."},{"cited_title":"IEEE transactions on pattern analysis and machine intelligence43(10), 3349–3364 (2020)","cited_arxiv_id":null,"evidence_quote":"HRNet, a high-resolution representation baseline used to situate PPOC-LL's accuracy-versus-parameter trade-off."}],"review_version":2}