{"id":"8f25a3ac-32e6-4251-ada9-9548b3d2ad59","arxiv_id":"2509.00872","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Pose-based DRF with PAV-guided attention reaches 86.0% accuracy and 80.8% F1 on Scoliosis1K, outperforming silhouette and raw-pose baselines.","lead":"This study introduces a new dataset of 2D pose keypoints from walking adolescents and a deep-learning framework that uses postural asymmetry cues to screen for scoliosis. It reports state-of-the-art accuracy on the Scoliosis1K benchmark, suggesting pose-based screening could be a privacy-preserving alternative to silhouette or torso-image methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PAV min-max normalization is computed 'across the dataset'; if it includes test sequences, DRF's reported gains are contaminated by leakage.","rationale":"The reader's verdict is CONDITIONAL, and their rationale already lists 'potential test-set leakage' as one reason. However, their weakest_assumption focuses on ViTPose keypoint accuracy and whether PAV preserves discriminative asymmetry. I think the normalization leakage is more load-bearing because it directly threatens the quantitative evidence for the central claim, whereas the keypoint-accuracy concern is partially mitigated by the ablation results: if ViTPose noise destroyed the asymmetry signal, PAV guidance would not outperform random or learnable priors as it does in Table 2b. The leakage concern is concrete, cheap to resolve, and if confirmed would require rerunning all experiments. I therefore keep the verdict at CONDITIONAL rather than moving it, but the condition should explicitly require training-only PAV normalization. I disagree slightly with the reader's emphasis: the keypoint accuracy issue is real but secondary; the normalization ambiguity is the first thing to verify.","tokens_in":7758,"tokens_out":6458,"duration_ms":84992,"concrete_test":"Obtain the released code and dataset split. Verify whether the min/max for PAV normalization are computed on the 745 training sequences only or on all 1,493 sequences. Then re-run the full DRF and the Table 2b ablation with training-only min/max. If accuracy/F1 drops by more than ~1 point (or the PAV-vs-learnable gap narrows materially), the reported SOTA and the clinical-prior conclusion are partially artifacts of leakage; if results are essentially unchanged, the concern is cleared.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In §3.1 (Sequence-Level PAV Refinement, step 3), the authors state: 'Min-Max scaling is applied across the dataset for all P × M dimensions.' Since evaluation uses a fixed train/test split (745/748 sequences), 'dataset' is ambiguous. If min/max are fitted on the union of train and test PAV vectors, the test PAVs given to the PGA module (§3.2) are not truly unseen: their normalization constants encode information about the test distribution. This transductive leakage can inflate the central comparison in Table 1 (DRF vs. ScoNet-MTske: +3.5% accuracy, +4.2% F1) and the Table 2b ablation showing PAV guidance outperforming learnable/random priors. The paper does not state that normalization was fit on the training split only, and the released code is not analyzed. Because the SOTA claim and the clinical-prior conclusion rest on this gap, the numbers need re-verification before accepting the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pose-based scoliosis screening framework. It introduces Scoliosis1K-Pose, a 2D keypoint annotation set for the Scoliosis1K dataset, and the Dual Representation Framework (DRF), which combines a continuous Gaussian skeleton map with a hand-crafted Postural Asymmetry Vector (PAV) that encodes vertical, midline, and angular asymmetries between paired keypoints. A PAV-Guided Attention (PGA) module uses the PAV as a clinical prior to recalibrate features extracted from the skeleton map. On the fixed 745/748 train/test split of Scoliosis1K, the method reports 86.0% accuracy and 80.8% macro F1, outperforming silhouette-based ScoNet-MT and pose-coordinate baselines. Ablations indicate that both channel and spatial attention branches contribute, and that PAV guidance outperforms self-attention, fixed vectors, and a learnable prior.","tokens_in":8085,"tokens_out":2399,"duration_ms":30645,"significance":"If the reported numbers are sound, the paper makes a useful contribution: it introduces a new pose dataset for a clinically relevant screening task, proposes a clinically interpretable prior that is not derived from labels, and provides public code and data. The PAV construction is transparent and the ablation design is thoughtful. The main risk is methodological: the normalization step in §3.1 may leak test information, and the reported improvements are not accompanied by any uncertainty quantification. Because the central SOTA claim and the conclusion that PAV guidance is effective rest on the Tables 1 and 2 comparisons, these issues are load-bearing.","major_comments":[{"comment":"The PAV normalization is described as 'Min-Max scaling is applied across the dataset for all P × M dimensions.' The evaluation uses a fixed train/test split (745/748 sequences). If the min/max statistics are computed on the union of training and test PAVs, then the test PAVs input to the PGA module in §3.2 are not unseen; the normalization constants encode information about the test distribution. This can inflate the DRF versus ScoNet-MTske gains in Table 1 (accuracy +3.5%, F1 +4.2%) and the ablation comparisons in Table 2b. The paper must state explicitly that the normalization statistics were fitted on the training split only, and ideally report results with normalization fitted per split. Without this clarification, the central comparison is unverifiable.","section":"§3.1 (Sequence-Level PAV Refinement, step 3) and §4.1 (Evaluation Protocol)"},{"comment":"All reported numbers appear to come from a single run on a single split. The differences that support the main claims are modest: +3.5% accuracy and +4.2% F1 in Table 1, and similar gaps in the Table 2 ablations. No error bars, standard deviations, or multiple-seed results are provided. Since the test set is small and the class distribution is imbalanced, these gaps could fall within run-to-run variation. The authors should report mean ± std across at least 3–5 random seeds, or a statistical significance test (e.g., paired bootstrap over test sequences).","section":"Tables 1 and 2"},{"comment":"The PAV is computed from ViTPose 2D keypoints, and the method assumes that these keypoints are sufficiently accurate to capture subtle left-right asymmetries in natural walking videos. The paper does not independently validate this premise. A concrete test would be to compare pose-derived asymmetries (e.g., shoulder height difference, pelvic tilt) with radiological or clinical measures on a subset of subjects, or to inject synthetic keypoint noise and measure DRF's performance sensitivity. If the PAV is dominated by pose-estimation noise, the reported improvement could be an artifact rather than a clinically meaningful prior.","section":"§2 and §3.1 (ViTPose keypoint dependence)"}],"minor_comments":[{"comment":"The Gaussian heatmap parameter σ is introduced but its numerical value is not given in the implementation details. Please specify σ (or state that it follows the skeleton-map literature).","section":"§3.1, Eq. (1) and Eq. (2)"},{"comment":"The IQR outlier removal is described only as 'statistical filtering.' Please state the IQR multiplier (e.g., 1.5×) and what happens when a value is removed, for reproducibility.","section":"§3.1, step 3"},{"comment":"The combined loss Ltotal = Lce + Ltriplet is mentioned, but the triplet loss setup (margin, sampling strategy, how anchors/positives/negatives are defined) is not described. Please add details or cite the exact protocol from ScoNet-MT.","section":"§3.2, Training Objective"},{"comment":"The feature response heatmaps would be more convincing if the scale or normalization of the heatmap was identical across all three panels. Please state the visualization protocol or ensure a common color scale.","section":"Figure 3"},{"comment":"OF-DDNet and GPGait are adapted to 2D pose inputs, but the adaptation details are not given. Please specify how the adaptation was performed (e.g., input format, network changes) so readers can assess fairness.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is plausible and the dataset/code release is a strength. However, the normalization leakage concern in §3.1 is directly testable with the released code; if the code shows cross-dataset min/max fitting, the reported SOTA numbers would need to be recomputed. I would recommend the editor ask the authors to confirm the training-only normalization and provide uncertainty estimates before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something real. It contributes a new dataset, Scoliosis1K-Pose (447,900 frames of ViTPose keypoints from 1,050 adolescents), and a clean framework, DRF, that combines continuous skeleton maps with a hand-crafted Postural Asymmetry Vector (PAV). The PAV-guided attention (PGA) module is the key piece: it uses clinical asymmetry descriptors to modulate feature extraction. The ablations are genuinely informative—PAV beats self-attention, all-ones, random, and learnable priors by solid margins (Table 2b). That is a good demonstration that the prior itself carries signal, not just the attention machinery. The visualizations also match the claim: the model attends to both shoulders and pelvis rather than one body side. I believe the central idea is sound and the dataset will be useful to the community.\n\nThe soft spots are real but not fatal. The stress-test concern about PAV min-max normalization \"applied across the dataset\" is legitimate: it is ambiguous whether the scaling constants were fitted on the training split only. If they include test sequences, that is transductive leakage, and the 3.5% accuracy gain over ScoNet-MTske could be inflated. The fix is simple—clarify and re-run with training-only statistics. Even without leakage, the absence of error bars in Tables 1 and 2 is a weakness for a medical screening claim. One run of a deep network can move by a couple of points; the reported differences are plausible but not yet backed by variance. Also, the comparative evaluation is thin: only three baselines, all on one dataset. That is acceptable for a new benchmark, but it means the \"state-of-the-art\" claim is really \"best on this benchmark so far.\" Finally, the PAV relies entirely on ViTPose keypoints. The paper does not independently check whether the extracted asymmetries correlate with radiological measures. IQR filtering and temporal averaging mitigate noise, but a sanity check would strengthen the clinical-prior story.\n\nOverall: this deserves a serious referee. I would accept with requested revisions, mainly the normalization clarification, error bars, and ideally a small keypoint-noise analysis. The paper is clearly written, the ablation design is above average, and the dataset alone is a contribution. My recommendation: send it to peer review and ask for those re-verifications before trusting the headline numbers.","headline":"Solid, well-ablated paper on pose-based scoliosis screening with a new dataset and a useful clinical-prior attention module; the headline gain over baselines needs re-verification because of an ambiguous normalization step and missing error bars.","tokens_in":8491,"tokens_out":1871,"would_cite":true,"duration_ms":25186,"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":"Clinical pose prior lifts scoliosis screening to 86% accuracy","keywords":["scoliosis screening","2D pose estimation","postural asymmetry","clinical prior","attention mechanism","ViTPose","Scoliosis1K","gait analysis"],"falsifier":"Compute the PAV descriptors for the same adolescents and correlate them with Cobb angles from radiographs or with clinical Adam's forward bend test scores; if the PAV shows no correlation with curve severity, or if adding the PGA module to the skeleton-map baseline stops helping when keypoints are artificially corrupted, the clinical-prior claim would be undermined.","tokens_in":7736,"feed_emoji":"🦴","tokens_out":3144,"duration_ms":36322,"temperature":0.7,"pith_summary":"This paper claims that 2D pose keypoints, when enriched with explicit clinical asymmetry descriptors, can outperform silhouette-based methods for scoliosis screening from ordinary walking videos. The authors introduce Scoliosis1K-Pose, a 447,900-frame pose annotation set for 1,050 adolescents, and a Dual Representation Framework (DRF) that combines a continuous skeleton map with a discrete Postural Asymmetry Vector (PAV). A PAV-Guided Attention module uses the PAV as a clinical prior to direct feature extraction. On the Scoliosis1K benchmark, DRF reaches 86.0% accuracy and 80.8% macro F1, beating the silhouette-based ScoNet-MT (82.0% accuracy, 75.4 F1) and pose-coordinate baselines (67.0–67.5 F1). If the result holds, it would make privacy-preserving, interpretable scoliosis screening feasible using only natural gait videos.","feed_headline":"Clinical pose prior lifts scoliosis screening to 86% accuracy","feed_subtitle":"A dual-representation network using shoulder and pelvic asymmetry beats silhouette-based methods on 1,050 adolescents' walking videos.","key_machinery":"The Postural Asymmetry Vector (PAV) is a matrix over anatomically symmetric keypoint pairs (e.g., shoulders, hips, knees) with three metrics: vertical deviation, midline deviation, and angular deviation, refined by interquartile-range outlier removal, temporal mean aggregation, and min-max normalization. The PAV-Guided Attention (PGA) module takes the PAV as input and learns channel-wise and spatial-wise attention weights via a linear layer and a 1D convolution, then applies them to the skeleton-map features. The PAV supplies the clinical prior; the PGA enforces it on the learned representation.","core_discovery":"The central claim is that modeling clinically defined postural asymmetries—vertical, midline, and angular deviations between paired keypoints—as an explicit prior improves pose-based scoliosis classification. The Dual Representation Framework converts per-frame 2D keypoints into a two-channel Gaussian skeleton map and a temporally aggregated PAV matrix. The PAV, rather than raw coordinates or learned features, guides the attention module: it generates channel and spatial weights that recalibrate the skeleton-map features before classification. Experiments show the full model outperforms the same architecture without attention, with only the skeleton map, or with alternative guidance sources","pith_inferences":["The PAV descriptions might transfer to other asymmetry-related conditions, such as Parkinson's disease gait or stroke-induced hemiparesis, where left-right deviations are diagnostic.","The framework's reliance on ViTPose keypoints means its gains could shrink in settings with poor pose estimation; testing the same pipeline with other detectors or with 3D pose could reveal the extent of that dependency.","Because the paper does not compare PAV descriptors with radiological measurements, an obvious next step is to regress Cobb angle or curve severity from the PAV, which would strengthen the clinical grounding of the method.","Combining the skeleton map and PAV with the original silhouette stream in a three-branch fusion might further improve accuracy, since silhouettes and skeletons carry complementary shape and asymmetry cues."],"forward_implications":["Pose data alone can serve as a sufficient input for scoliosis screening, avoiding the need for silhouettes or exposed-torso imagery.","The PAV's three asymmetry metrics capture clinically meaningful cues that can be computed cheaply from any 2D pose estimator, making the approach easy to deploy in school or community screening.","Adding the PAV-guided attention module to an existing skeleton-map pipeline yields consistent gains, suggesting explicit clinical priors can complement learned features in medical image analysis.","The Scoliosis1K-Pose dataset provides a ready benchmark for future pose-based scoliosis research, including methods that fuse silhouette and pose modalities."],"supporting_citations":[{"why":"Supplies the original Scoliosis1K dataset, the ScoNet-MT architecture, and the standard train/test split that all comparisons use.","marker":"[21]"},{"why":"ViTPose is the 2D pose estimator used to generate the 447,900 keypoint frames of Scoliosis1K-Pose.","marker":"[17]"},{"why":"Defines the MS-COCO 17-keypoint format that determines which keypoint pairs are treated as anatomically symmetric in the PAV.","marker":"[11]"},{"why":"A pose-coordinate gait-based baseline that the paper adapts to 2D pose inputs and compares against to show coordinate-only methods lag behind.","marker":"[12]"},{"why":"A pose-based gait recognition method (GPGait) used as a second pose-coordinate baseline to benchmark the new framework.","marker":"[7]"},{"why":"Introduces the skeleton-map representation that DRF's continuous input is built upon.","marker":"[6]"},{"why":"The clinical practice guideline that identifies shoulder imbalance and pelvic tilt as key screening indicators, motivating the PAV's asymmetry metrics.","marker":"[3]"}],"fun_headline_variants":["Scoliosis screening sees straight with pose asymmetry prior","Pose asymmetry prior straightens scoliosis screening accuracy","Dual pose representations hit 86% in scoliosis screening","Clinical asymmetry cues guide pose-based scoliosis detection","PAV-guided attention boosts scoliosis screening to 86%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The framework assumes the 2D pose keypoints produced by ViTPose are accurate enough across natural walking videos to capture subtle left-right asymmetries that survive temporal averaging, and the paper does not verify this by comparing pose-derived asymmetry with X-ray or clinical measurements.","fun_headline_variants_meta":{"raw":{"variants":["Scoliosis screening sees straight with pose asymmetry prior","Pose asymmetry prior straightens scoliosis screening accuracy","Dual pose representations hit 86% in scoliosis screening","Clinical asymmetry cues guide pose-based scoliosis detection","PAV-guided attention boosts scoliosis screening to 86%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001,"raw_usage":{"total_tokens":4088,"prompt_tokens":780,"completion_tokens":3308,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":3229}},"tokens_in":524,"tokens_out":3308,"duration_ms":25457,"temperature":1.0,"reasoning_tokens":3229,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:06:20.166870+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the PAV descriptors for the same adolescents and correlate them with Cobb angles from radiographs or with clinical Adam's forward bend test scores; if the PAV shows no correlation with curve severity, or if adding the PGA module to the skeleton-map baseline stops helping when keypoints are artificially corrupted, the clinical-prior claim would be undermined.","supporting_citations":[{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Supplies the original Scoliosis1K dataset, the ScoNet-MT architecture, and the standard train/test split that all comparisons use."},{"cited_title":"Advances in Neural Information Processing Systems35, 38571–38584 (2022)","cited_arxiv_id":null,"evidence_quote":"ViTPose is the 2D pose estimator used to generate the 447,900 keypoint frames of Scoliosis1K-Pose."},{"cited_title":"In: Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13","cited_arxiv_id":null,"evidence_quote":"Defines the MS-COCO 17-keypoint format that determines which keypoint pairs are treated as anatomically symmetric in the PAV."},{"cited_title":"In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part III","cited_arxiv_id":null,"evidence_quote":"A pose-coordinate gait-based baseline that the paper adapts to 2D pose inputs and compares against to show coordinate-only methods lag behind."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"A pose-based gait recognition method (GPGait) used as a second pose-coordinate baseline to benchmark the new framework."},{"cited_title":"In: Proceedings of the AAAI conference on artificial intelligence","cited_arxiv_id":null,"evidence_quote":"Introduces the skeleton-map representation that DRF's continuous input is built upon."},{"cited_title":"Chinese Journal of Orthopaedics 40(23), 1574–1582 (2020)","cited_arxiv_id":null,"evidence_quote":"The clinical practice guideline that identifies shoulder imbalance and pelvic tilt as key screening indicators, motivating the PAV's asymmetry metrics."}],"review_version":1}