{"id":"c21e1818-ac57-4a61-8af8-3ed2a6a31e6f","arxiv_id":"2411.15799","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A dual-path network that compares a back photo with its mirror image and uses ordinal binary classifiers achieves state-of-the-art scoliosis severity classification on a new private dataset of 1,898 natural back images.","lead":"A new deep-learning system estimates scoliosis severity from ordinary photos of the back, reaching 95.11% accuracy for four broad levels and 81.46% for ten fine-grained levels. The work is an attempt to enable cheap, radiation-free home screening, but the results rest on a private dataset with no released code or data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 95.11% accuracy may be inflated because the five-fold split is image-level on 1,898 images from only 1,067 patients, so same-patient photos likely appear in both training and test folds.","rationale":"The reader's weakest-assumption analysis correctly identifies the image-level versus patient-level split as the central threat. My independent reading of Sec. 4.1.1 and Sec. 4.1.2 confirms that no patient-level stratification is described. Because the dataset has roughly 1.78 images per patient on average, the expected number of cross-fold same-patient pairs is large, and convolutional models are known to exploit such identity-level shortcuts, especially when global appearance is highly correlated with the condition. I therefore agree with the CONDITIONAL verdict: the architecture and ablations are internally plausible, but the headline accuracy and the SOTA comparison are unverified until a patient-disjoint evaluation is reported. I would also note a secondary reporting issue: Figs. 4 and 5 appear to contain identical confusion matrices despite being captioned for different tasks, which further weakens the human-performance comparison; however, that is not the single most load-bearing issue. The concrete test of a patient-level split directly settles whether the central accuracy claim holds.","tokens_in":15777,"tokens_out":2202,"duration_ms":22247,"concrete_test":"Re-run the five-fold cross-validation with patient-level containment: use the 1,067 patient IDs to group images, assign all images of each patient to one fold only (e.g., StratifiedGroupKFold), and retrain the proposed model and at least the strongest baseline (TransNeXt-S) under identical settings. If the patient-level general accuracy drops materially below 95.11%, or below TransNeXt-S's 94.58%, the reported margin and screening claim are not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 4.1.1 states the dataset contains 1,898 natural back images from 1,067 patients, and Sec. 4.1.2 says USTC&SYSU-Scoliosis is randomly divided into five folds. The paper does not state that folds are stratified by patient, nor that all images of a patient are confined to one fold. With 1,898 images and 1,067 patients, many patients contribute multiple images; a random image-level split will, with high probability, place images of the same patient in both training and test folds. A network can then memorize patient-specific appearance (clothing, skin texture, pose, background) rather than learn generalizable back-asymmetry cues, inflating the reported 95.11% general-level and 81.46% fine-grained accuracies. The same leakage affects every baseline in Table 2, so the claim of outperforming SOTA methods is also not a fair comparison. The screening framing in the abstract and Sec. 5 requires generalization to unseen patients, not merely unseen images, so this is a load-bearing threat to the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a dual-path deep network for estimating scoliosis severity from natural (non-radiographic) back images. Two modules are introduced: a symmetric feature matching module (SFMM) that compares features of an image and its horizontal flip, and an ordinal regression head (ORH) that converts the multi-class severity estimation into K-1 binary classification sub-problems. The method is evaluated on a newly collected dataset of 1,898 natural back images from 1,067 patients, with five-fold cross-validation, and is reported to achieve 95.11% accuracy for general severity levels, 81.46% for fine-grained levels, and to outperform state-of-the-art methods and human experts. The paper includes ablations, statistical analyses, and visualizations supporting the contributions.","tokens_in":15929,"tokens_out":5618,"duration_ms":49222,"significance":"If the reported results hold, the work would be a useful step toward low-cost, wide-range scoliosis screening using ordinary photographs, avoiding radiation exposure. The SFMM gives a principled way to exploit the symmetric/asymmetric structure of the back, and the ORH appropriately respects the ordinal nature of severity levels. The paper is also commendable for collecting a natural-image dataset paired with X-ray Cobb-angle labels and for including human expert comparisons, ablation studies, and per-fold results. However, the evaluation protocol has a load-bearing weakness: the random split is at the image level, not the patient level, so identity leakage could substantially inflate the headline accuracies. The lack of confidence intervals and the duplicated human-comparison figures further weaken the central claims as currently presented.","major_comments":[{"comment":"The dataset contains 1,898 images from only 1,067 patients, and Section 4.1.2 states that the data are \"randomly divided into five folds\" without any mention of patient-level stratification. With an average of about 1.78 images per patient, a random image-level split will with high probability place images from the same patient in both the training and test folds. This allows the model to memorize patient-specific appearance (clothing, background, skin texture) rather than generalizable asymmetry cues, inflating the reported 95.11% and 81.46% accuracies and the SOTA comparison in Table 2. The authors should re-run the experiments using a patient-level split, or at minimum report the fraction of patients whose images appear in both training and test folds and analyze the impact on the results.","section":"Sections 4.1.1 and 4.1.2"},{"comment":"The reported accuracies are averages over five folds without confidence intervals or significance tests. The margin over TransNeXt-S is only 0.53 percentage points (95.11% vs 94.58%), while per-fold results in Table 5 vary from 93.16% to 97.88%. Without a paired per-fold statistical test (e.g., Wilcoxon signed-rank test or corrected paired t-test) or confidence intervals, the claim of state-of-the-art performance is not statistically supported. Please add the missing statistical measures or temper the claim accordingly.","section":"Table 2 and Table 5"},{"comment":"The confusion matrices shown in Figures 4 and 5 appear to be identical, despite the fact that the two figures are supposed to represent different tasks: general severity level estimation vs fine-grained severity level estimation. Since the reported accuracies for these two tasks differ substantially (95.11% vs 81.46%), the duplicated figures are internally inconsistent and cannot be correct as presented. This undermines the human-comparison experiment and the claim of outperforming human performance. The authors should provide the correct figures and describe the human evaluation protocol more rigorously, including inter-observer variability and results across more than one fold.","section":"Section 4.3, Figures 4 and 5"},{"comment":"The Spinecube and ScolioNets baselines are reimplemented by the authors because the original code is not released. The text states that \"we utilize the original settings in their code or papers\" and also that \"for a fair comparison, the networks of these methods are trained up to the same 610 epochs as our method.\" These two statements are in tension: if the original methods used fewer training epochs, forcing 610 epochs may disadvantage them (e.g., through overfitting), or conversely give them an advantage. The reimplementation fidelity is not validated. Please clarify the exact training protocol, and if possible, compare against the original papers' reported metrics on comparable data, or release the reimplementation code to allow verification.","section":"Section 4.2"}],"minor_comments":[{"comment":"The text states that \"when lambda_general is higher, the model performs better in estimating the general severity level,\" but the 2:1 ratio yields 95.02% accuracy, which is lower than the 1:1 ratio's 95.11%. Please correct the narrative or explain why the 1:1 ratio is considered optimal despite this apparent contradiction.","section":"Table 4"},{"comment":"The human comparison is performed on only the fifth fold and with only two experts, which limits the generality of the \"outperforms human performance\" claim. Adding more experts and reporting inter-observer agreement would strengthen this comparison.","section":"Section 4.3"},{"comment":"The notation Y_k1 and Y_k2 is used but the second component Y_k2 is never formally introduced; please define both components of the label vector for clarity.","section":"Equation (3)"},{"comment":"The FLOPs and #Params reported for the proposed method include both the general and fine-grained branches, while the comparison methods are implemented as single-task models. For a fair complexity comparison, please also report the complexity of a single branch of the proposed network.","section":"Section 4.2"},{"comment":"There is a typo in the sentence \"we use a popular Grad-CAM [38] technique to generate heatmaps for general severity level estimation branch of our method\"; the phrase \"for general severity level estimation branch\" should read \"for the general severity level estimation branch.\"","section":"Section 4.6"}],"recommendation":"major_revision","confidential_remarks":"The patient-level leakage concern in Section 4.1 is the key issue: if the split is indeed at the image level, the headline accuracies could be substantially inflated, and the method may not generalize to unseen patients as claimed. The duplicated figures in Section 4.3 also suggest an error in figure preparation rather than a fundamental flaw. I believe the paper can be revised to address these concerns, but the experiments need to be re-run with a patient-level split and the statistical analysis needs to be added before the claims can be accepted. Given the journal's standards, major revision is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a plausible engineering paper for natural-image scoliosis screening, but the headline accuracy claim is not yet trustworthy because the five-fold split is image-level on 1,898 images from 1,067 patients, and the human comparison figures are duplicated. The method itself is a legitimate combination: weight-shared backbone on input plus horizontal flip, symmetric feature matching with cat-conv and self-attention, and an ordinal regression head that decomposes K labels into K-1 binary tasks. That is not a huge conceptual leap, since Spinecube and ScolioNets already did natural-image screening, but the particular pairing is sensible and the ablations show each component adds roughly a point on the private USTC&SYSU-Scoliosis set. I credit the authors for using real paired X-ray labels, a five-fold protocol, a broad backbone comparison, and an honest limitations paragraph about compute and Cobb-angle range prediction.\n\nThe soft spots are concentrated in the evaluation. First, the split. With 1,898 images from only 1,067 patients, a random image-level five-fold split will very likely put multiple photos of the same patient in both train and test. The paper never says folds are patient-exclusive. That alone can inflate the 95.11% and 81.46% numbers through appearance memorization, and it also makes Table 2's SOTA comparison unfair in the same direction for every method. This is not a nitpick; it undercuts the screening framing, which is about unseen patients, not unseen images. Second, Figs. 4 and 5 are identical. Both show the same six confusion matrices, yet the text claims different expert performance for general and fine-grained tasks. That looks like a copy-paste error, and until it is corrected the \"outperforms human performance\" claim has no visible support. Third, the margin over the strongest baseline is only about half a point, with no confidence intervals or significance tests across folds. Fourth, data and code are not released; Spinecube and ScolioNets are reimplemented without their code, so the comparisons are hard to audit.\n\nNone of these are fatal to the core idea. If the authors add a patient-level split or external validation, release artifacts, report variance, and fix the figures, the paper would be a useful contribution for the medical-image-screening community. The right move is to send it to peer review with a request for major revision, not to desk reject.","headline":"Reasonable idea and sensible modules, but the headline accuracy is not yet trustworthy because the image-level split likely leaks patient identity and the human-comparison figures are duplicated.","tokens_in":16527,"tokens_out":2176,"would_cite":false,"duration_ms":20722,"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":"The paper argues that scoliosis severity can be graded from ordinary back photographs using symmetry matching and ordinal regression, reporting 95.11% general and 81.46% fine-grained accuracy on a new dataset.","keywords":["scoliosis detection","natural image","symmetric feature matching","ordinal regression","severity level estimation","human back","Cobb angle","dual-path network"],"falsifier":"A patient-exclusive five-fold split of the collected dataset, keeping all images of each patient in a single fold, would settle the matter: if accuracy falls materially below 95.11%, the current split is leaking identity information.","tokens_in":15523,"feed_emoji":"🩺","tokens_out":6860,"duration_ms":55940,"temperature":0.7,"pith_summary":"This paper argues that scoliosis severity can be graded directly from ordinary photographs of a patient's back, without X-rays, by exploiting the fact that a healthy back is roughly symmetric and disease makes it asymmetric. The proposed network feeds both the photo and its horizontal mirror image through a shared backbone, uses a symmetric feature matching module to compare the two sides, and replaces flat multi-class classification with an ordinal regression head that converts the severity scale into a series of binary comparisons. On a collected dataset of 1,898 back images from 1,067 patients, the method reports 95.11% accuracy for the four-level general severity task and 81.46% for the ten-level fine-grained task, above the compared state-of-the-art models and above two spine surgeons asked to grade the same photographs. The practical goal is cheap, radiation-free, wide-ranging screening for adolescent scoliosis that could happen outside specialist clinics.","feed_headline":"Back-photo AI grades scoliosis at 95% accuracy","feed_subtitle":"No X-rays needed: network matches back asymmetry and ordinal severity labels, beating doctors on the test set.","key_machinery":"The load-bearing mechanism is the symmetric feature matching module: original and horizontally flipped back images go through a weight-sharing backbone, their features are concatenated and convolved into a fused feature map, and self-attention uses that fused map as key and value while the two branch features act as queries, so the network can directly compare left and right sides of the back. The ordinal regression head then rewrites the $K$-class severity problem as $K-1$ binary classification subproblems, using the label ordering to sharpen class boundaries instead of treating severity levels as independent categories.","core_discovery":"The central claim is that bilateral symmetry of the human back is a usable visual signal for scoliosis severity and that explicit symmetry matching plus ordinal regression turns that signal into accurate severity labels from natural images. The network takes the original back photo and its horizontal flip, extracts features with a shared visual attention backbone, fuses the two feature maps by concatenation-convolution followed by self-attention to model symmetric correspondence, and then sends the fused features through an ordinal regression head that learns $K-1$ binary classifiers, each asking whether the true severity rank exceeds $k$. The predicted level is obtained by summing the binary answers. With this design the paper reports state-of-the-art results on its collected dataset, including near-perfect recall for the severe class, and concludes that the approach is a viable economic screening tool.","pith_inferences":["The headline accuracies may be optimistic because the split is random at the image level, not the patient level; since the dataset has 1,898 images from only 1,067 patients, the same patient can appear in both training and test folds, and identity memorization could inflate the numbers. A patient-exclusive split would be the natural stress test.","The symmetry mismatch between original and flipped feature maps could be turned into a continuous asymmetry score that regresses the Cobb angle directly, something the paper explicitly says its current method cannot do because it only predicts severity ranges.","Two spine surgeons were asked to grade photographs without the aid of corresponding X-rays; a real deployment would need to define how the tool is used and what follow-up it triggers, since the paper does not address referral workflows."],"forward_implications":["A radiation-free screening pipeline could be built from ordinary back photographs, lowering the barrier for mass or home-based checks of adolescent scoliosis.","The dual-path symmetry-aware design could transfer to other bilateral medical surfaces where asymmetry signals disease, such as facial palsy or limb swelling.","Because the ordinal head combines binary answers into a rank, the same network can report both coarse and fine severity levels through two parallel branches sharing the backbone.","At the reported 95.11% general accuracy, the model would misplace roughly 5% of images in a four-level task, with most remaining confusion between the minor and moderate classes rather than across distant levels."],"supporting_citations":[{"why":"Defines the natural-image back screening task and supplies the primary field baseline the paper must beat.","marker":"[56]"},{"why":"Smartphone-photograph scoliosis classification method used as a comparison baseline.","marker":"[59]"},{"why":"Visual attention network used as the shared backbone that extracts features from both the image and its flipped version.","marker":"[12]"},{"why":"Self-attention formulation used inside the symmetric feature matching module to relate original and flipped features.","marker":"[49]"},{"why":"Defines the Cobb angle, the clinical measurement from X-rays that determines the severity-level ground truth labels.","marker":"[4]"},{"why":"Supplies the four-level general severity scale (normal, minor, moderate, severe) used to define the coarse classification task.","marker":"[57]"}],"fun_headline_variants":["Back symmetry AI detects scoliosis without X-rays","Ordinal regression reads back photos for scoliosis","AI matches back symmetry to grade scoliosis severity","Back-photo model beats doctors on scoliosis grading","No-radiation scoliosis screening via symmetry network"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation's random five-way split of images assumes that the same patient's photographs never appear in both the training and test folds; if they do, the reported accuracy is inflated by the model recognizing familiar patients rather than generalizing to new backs.","fun_headline_variants_meta":{"raw":{"variants":["Back symmetry AI detects scoliosis without X-rays","Ordinal regression reads back photos for scoliosis","AI matches back symmetry to grade scoliosis severity","Back-photo model beats doctors on scoliosis grading","No-radiation scoliosis screening via symmetry network"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00015,"raw_usage":{"total_tokens":1199,"prompt_tokens":947,"completion_tokens":252,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":180}},"tokens_in":563,"tokens_out":252,"duration_ms":2822,"temperature":1.0,"reasoning_tokens":180,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:54:22.637887+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A patient-exclusive five-fold split of the collected dataset, keeping all images of each patient in a single fold, would settle the matter: if accuracy falls materially below 95.11%, the current split is leaking identity information.","supporting_citations":[{"cited_title":"JAMA Network Open 6(8), e2330,617– e2330,617 (2023)","cited_arxiv_id":null,"evidence_quote":"Smartphone-photograph scoliosis classification method used as a comparison baseline."},{"cited_title":"Communications Biology 2(1), 390 (2019)","cited_arxiv_id":null,"evidence_quote":"Defines the natural-image back screening task and supplies the primary field baseline the paper must beat."},{"cited_title":"Computational Visual Media 9(4), 733–752 (2023)","cited_arxiv_id":null,"evidence_quote":"Visual attention network used as the shared backbone that extracts features from both the image and its flipped version."},{"cited_title":"In: Advances in Neural Information Processing Systems, pp","cited_arxiv_id":null,"evidence_quote":"Self-attention formulation used inside the symmetric feature matching module to relate original and flipped features."},{"cited_title":"Instructional Course Lecture (1948)","cited_arxiv_id":null,"evidence_quote":"Defines the Cobb angle, the clinical measurement from X-rays that determines the severity-level ground truth labels."},{"cited_title":"Beijing China: People’s Health Publishing House (2015)","cited_arxiv_id":null,"evidence_quote":"Supplies the four-level general severity scale (normal, minor, moderate, severe) used to define the coarse classification task."}],"review_version":1}