{"id":"b241e2b1-609d-4b44-af69-b789923c020c","arxiv_id":"2507.07131","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A model trained on CT-derived simulated X-rays segments ten wrist bones in radiographs, with average per-view Dice 0.80 to 0.92 on simulated tests but only qualitative evidence on real X-rays.","lead":"The authors trained a segmentation model on thousands of simulated X-rays generated from CT scans, avoiding the need for manually labeled wrist radiographs. The model outlines ten wrist bones across different view angles, but quantitative validation currently rests on simulated images and only qualitative checks on real X-rays.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim is not quantitatively validated on real X-rays; simulated Dice alone cannot establish transfer to real radiographs.","rationale":"The paper is a clearly presented application of CT-based X-ray simulation to multi-view wrist segmentation, with a practical motivation and an honest admission that real-image evaluation is qualitative. The simulated Dice results and qualitative examples are encouraging, but the load-bearing claim about real X-ray performance lacks quantitative support. The reader's weakest assumption points to the sim-to-real gap; the concrete missing piece is a quantitative real-image evaluation that would directly test this gap. My concern therefore partially overlaps with the reader's but is framed as missing evidence rather than a specific failure mechanism. The proposed test - expert annotations on a subset of the real radiographs - would settle whether the transfer claim holds. If the test reveals a large drop in real-image Dice, the central claim would be weakened; if the drop is modest, the method would be validated as intended. I recommend keeping the CONDITIONAL verdict because the idea is sound but the evidence is incomplete, which is exactly the reader's conclusion.","tokens_in":6217,"tokens_out":4407,"duration_ms":51005,"concrete_test":"Select 30 of the 121 real radiographs spanning view angles and bone overlap difficulty, have two expert radiologists independently annotate the 10 bones, compute Dice and ASD between the model output and the consensus masks, and compare per-bone and per-angle results with Table 1. If average real-image Dice at matched angles is materially lower (e.g., more than 0.15) than simulated Dice, the sim-to-real transfer claim would not be supported. If public wrist X-ray datasets with ground truth become available, an additional comparison with existing methods (e.g., Kang et al.) on AP views would further quantify the gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a model trained on CT-based simulated X-rays segments real wrist radiographs across view angles. Quantitative support comes only from simulated test images produced by the same pipeline (Section 3.2, Table 1), which mainly shows self-consistency with the training distribution. Real-image evaluation (Section 3.3) is explicitly qualitative, with no ground truth; the conclusion defers quantitative real-image evaluation to future work. The simulation model is heavily simplified (orthographic, scatter-free, monoenergetic, Eq. 1) and includes an ad hoc tissue-reduction step with thresholds set by trial and error. These choices can plausibly distort bone contrast and edges enough that real-image performance differs substantially from the reported Dice scores of 0.80-0.92. Without quantitative validation on real radiographs, the paper does not yet establish the transfer claim. This is a missing-evidence concern, not an internal inconsistency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes to train a deep-learning segmentation model (nnU-Net) on simulated wrist X-ray images generated from CT volumes, with the goal of automatically segmenting eight carpal bones plus the distal radius and ulna in real plain radiographs across view angles from -70 to 70 degrees. The simulation uses orthographic projection with simplified attenuation physics, and the corresponding masks are obtained by projecting the CT bone segmentations. The authors report Dice scores of 0.80-0.92 on a simulated test set, and present qualitative segmentation results on 121 real wrist radiographs without ground truth.","tokens_in":6359,"tokens_out":2875,"duration_ms":35304,"significance":"If the transfer from simulated to real X-rays were quantitatively established, the approach would offer a valuable solution to the label-scarcity problem for multi-view wrist segmentation, potentially enabling downstream applications such as fracture detection and bone age assessment. The paper's strengths include the use of a publicly available training pipeline (code and model), a relatively large synthetic dataset covering multiple angles, and the inclusion of all eight carpal bones. The central claim, however, rests on a self-referential evaluation: the quantitative metrics are obtained only on simulated images generated by the same pipeline used to create the training data, while the real-image evaluation is explicitly qualitative. The significance of the reported Dice scores therefore depends entirely on the unvalidated assumption that the simulation distribution is close enough to real clinical radiographs.","major_comments":[{"comment":"The central claim that the model segments real X-ray images is not quantitatively supported. The abstract states that the method was evaluated using both simulated and real images, but Section 3.3 presents only qualitative examples for real radiographs, and Section 4 explicitly defers quantitative real-image evaluation to future work. Since the entire quantitative evidence (Table 1) comes from simulated test images produced by the same pipeline as the training data, the reported Dice/ASD scores demonstrate self-consistency of the synthetic generation procedure rather than generalization to real X-rays. Please provide a quantitative evaluation on real radiographs with expert-annotated ground truth for at least a subset of the 121 images, or alternatively use a public wrist X-ray segmentation benchmark, in order to substantiate the abstract's claim.","section":"Section 3.3 and Section 4"},{"comment":"The realism of the simulation is load-bearing for the paper's transfer claim, yet the simulation is heavily simplified and includes ad hoc adjustments. Specifically, the orthographic, scatter-free, monoenergetic model in Eq. (1) ignores key physical effects, and the tissue-reduction step uses thresholds 'determined by trial and error' (setting intensities below the 20th percentile to the 10th percentile). These choices can plausibly distort bone contrast and edges relative to real radiographs. To gauge the sim-to-real gap, the paper should include at least a quantitative comparison of intensity statistics or a small calibration experiment on real images with labels; without such evidence, the simulated Dice scores alone do not establish that the model will perform on real X-rays.","section":"Section 2.2, Eq. (1)"},{"comment":"The statement 'No existing method could segment wrist radiographs in a variety of view angles, hence only our method was evaluated' is not a valid justification for omitting comparison or quantitative evaluation. Even if no prior method covers all the claimed view angles, existing wrist-segmentation methods (e.g., Kang et al. [4]) could be compared on the AP subset, or the authors could report a reader study on the real radiographs. At a minimum, the real-image segmentation results should be quantified with metrics that do not require dense ground truth, such as boundary agreement between two expert annotators or overlap-based scores on a small manually labeled subset.","section":"Section 3.1"},{"comment":"The training and test labels are not fully independent of the method being evaluated. The multi-label masks for all CT scans are obtained using the interactive segmentation tool [6] plus the shape and kinematic models from [7]; the same masks are used to generate both the training labels and the simulated test-set ground truth. Any systematic error in this label-propagation pipeline will inflate the reported Dice scores because the model is trained and tested on the same label definitions. The paper should report validation of the propagated masks against manual annotations on non-neutral poses, or at least quantify inter-observer variability of the segmentation masks.","section":"Section 2.1"}],"minor_comments":[{"comment":"The bone name 'Treapezium' appears to be a typo for 'Trapezium' in Table 1 and in Section 3.2; please correct it.","section":"Table 1"},{"comment":"The abstract states 'Dice scores ranging from 0.80 to 0.92' without specifying that this range corresponds to average Dice per view angle; the per-bone Dice at 70° can be substantially lower (e.g., Capitate 0.67 in Table 1). Please clarify the range in the abstract to avoid overstating the results.","section":"Abstract"},{"comment":"The term 'nnUnet' should be consistently written as 'nnU-Net' to match the reference [10] and the standard notation.","section":"Section 2.3"},{"comment":"The sentence 'the results of both negative and positive of the same angle are combined' is unclear; please specify whether this refers to averaging left and right wrists or to combining positive and negative rotation angles, and how the averaging is performed.","section":"Section 3.1"},{"comment":"The statement that the code and model 'link will be provided upon acceptance' is non-standard for a journal submission; please provide a stable repository or detailed instruction for obtaining the materials in the final version.","section":"Abstract and Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely problem and the synthetic-data approach is potentially valuable, but the current evidence does not support the headline claim of successful real-X-ray segmentation. The authors themselves acknowledge the lack of quantitative real-image evaluation in Section 4, and the self-referential nature of the simulated test set makes the reported Dice scores difficult to interpret. I would be willing to reconsider after the authors add a rigorous quantitative evaluation on real radiographs, even on a modest subset with expert annotations, and address the sim-to-real gap analysis. The paper is not ready for acceptance in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely useful: it trains a nnU-Net to segment all eight carpal bones plus the distal radius and ulna in wrist X-rays across view angles from -70 to 70 degrees, using 9240 simulated radiographs generated from CT volumes. That is a real advance over prior work, which either used manual labels on AP views only (Kang et al.) or excluded the carpals entirely (Lee et al.). The writing is clear, the motivation is sound, and the authors are honest about the limits of their evaluation.\n\nThe core weakness is that the quantitative evidence for the central claim comes entirely from simulated test images produced by the same pipeline used to generate training data. Table 1 reports Dice and ASD on 912 simulated images from held-out CT scans, but those images share the same projection model, the same tissue-reduction thresholds, and the same augmentation scheme as the training set. So the scores mostly show self-consistency, not transfer to real radiographs. The real-image evaluation, Section 3.3, is explicitly qualitative: 121 X-rays, no ground truth, selected examples, and no error rate or failure analysis beyond noting that partial hands fail. That is a missing-evidence problem, not an internal inconsistency, but it leaves the load-bearing claim—that simulation-trained segmentation transfers to clinical X-rays—unvalidated.\n\nThere are also smaller concerns. The simulation itself is deliberately simplified (orthographic, scatter-free, monoenergetic), and the tissue-reduction step uses thresholds determined by trial and error; these choices could plausibly distort bone contrast and edges in ways that matter for real images. The paper does not compare against any existing method, although it is fair to say that no prior method handles all ten bones across view angles. And despite the abstract promising the trained model and simulation code \"freely available... link will be provided upon acceptance,\" nothing is public at the moment, which limits independent verification.\n\nOn the positive side, the paper is not overclaiming in its conclusion: it explicitly says the method was evaluated quantitatively on simulated images and qualitatively on real images, and it defers quantitative real-image validation to future work. The related work is cited appropriately, including the authors' own prior work on CT segmentation and 3D statistical models, which is directly relevant and not self-citation for its own sake.\n\nWho gets value: researchers working on sim-to-real transfer in medical imaging, or on wrist or hand segmentation, will find this a useful proof-of-concept and a clear statement of the evaluation gap. It deserves a serious referee: the question is real, the method is reasonable, and the paper is honest. But as it stands, the central claim is not yet established. My recommendation is to send it to peer review with the expectation of major revision, requiring quantitative validation on real X-rays with reference labels, a baseline comparison, and public release of code and data.","headline":"Simulation-trained multi-view wrist segmentation is a solid proof-of-concept, but the paper's quantitative evidence is self-referential and the real-X-ray transfer claim remains unvalidated.","tokens_in":6895,"tokens_out":2940,"would_cite":false,"duration_ms":29928,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Using synthetic X-rays made from CT scans, a deep learning model segments eight wrist bones plus radius and ulna in real radiographs across view angles from -70 to 70 degrees.","keywords":["wrist bone segmentation","X-ray image segmentation","CT-based X-ray simulation","synthetic training data","sim-to-real transfer","carpal bones","deep learning","multi-view radiographs"],"falsifier":"Have a radiologist manually label the eight carpal bones and two forearm bones on the 121 real radiographs (or register the CT volumes to them to get ground truth) and compute Dice scores; if the real-image scores fall far below the simulated range of 0.80 to 0.92, the central claim of sim-to-real transfer fails.","tokens_in":6010,"feed_emoji":"🦴","tokens_out":4856,"duration_ms":49396,"temperature":0.7,"pith_summary":"This paper tries to show that a deep learning model can learn to segment the eight carpal bones plus the distal radius and ulna in ordinary wrist X-rays without any hand-labeled X-rays for training. Instead, the authors generate thousands of synthetic radiographs by projecting CT volumes at angles from -70 degrees to 70 degrees, using the CT bone labels as ground truth masks. On simulated test images the model reaches mean Dice scores of about 0.92 near the front-back view and 0.80 at 70 degrees; on 121 real radiographs the results are judged qualitatively, not by numbers. If the sim-to-real transfer holds, this removes the main bottleneck of expert annotation for multi-view wrist segmentation.","feed_headline":"Fake X-rays from CT train a model that works on real wrist images","feed_subtitle":"No manual X-ray labels needed: the model segments eight carpal bones plus radius and ulna from -70 to 70 degrees.","key_machinery":"The load-bearing mechanism is an orthographic X-ray simulation that models each pixel as the exponential of the negative summed attenuation along a ray through the CT volume, treating the CT values as absorption coefficients and ignoring scatter and energy dependence. Ground-truth masks come from 3D multi-label bone segmentations of the CT scans, obtained for one neutral pose and propagated to all poses by a statistical shape-and-kinematic wrist model; projected labels are resolved to the closest bone when overlap occurs. A tissue-reduction step clipped low intensities to boost bone contrast, with thresholds chosen by trial and error, and augmentation by rotation, translation, zoom, and horizontal flip widens the simulated distribution. Training then uses a self-configuring encoder-decoder network with combined Dice and cross-entropy loss.","core_discovery":"The central claim is that segmentation trained on CT-based simulated X-rays generalizes to real X-rays for ten wrist bones across a wide range of view angles. The paper reports that a single model, trained on 9240 synthetic images from 88 CT volumes, produces average Dice scores of 0.92 at 0 degrees and 0.80 at 70 degrees on a simulated test set, with per-bone values ranging roughly from 0.67 to 0.95. On real radiographs, the same model visually produces acceptable outlines and preserves anatomical overlap relations even in lateral views, though it fails when only part of the hand is visible. The authors state that no existing method automatically segments all eight carpal bones plus the two forearm bones in arbitrary views.","pith_inferences":["If the sim-to-real transfer is as good as the qualitative real-image results suggest, the same recipe could be applied to other joints or to the metacarpal bones, wherever CT volumes with bone labels exist.","A quantitative real-image benchmark, produced by expert annotation or by 3D-to-2D registration of the CT volumes to the radiographs, would be the natural next test and would tell how much of the simulated Dice survives the gap.","Adding perspective projection, scatter, and beam-hardening effects to the simulator, rather than handling them with augmentation and clipping, may specifically improve the high-angle and overlap-heavy cases where Dice drops below 0.8.","The dependence of the tissue-reduction thresholds on trial and error suggests a sensitivity analysis: reporting how segmentation varies as those thresholds move would show how robust the pipeline is."],"forward_implications":["Wrist segmentation training no longer needs manually annotated X-rays; CT-derived masks can supply arbitrarily many training pairs.","The same model covers view angles from -70 degrees to 70 degrees, including lateral views where carpal bones overlap heavily, something prior methods restricted to front-back views.","The released model and simulation code let others generate synthetic wrist X-rays for downstream tasks such as fracture detection, bone age assessment, and arthritis evaluation.","Performance degrades with view angle, with average Dice falling from 0.92 near the front-back view to 0.80 at 70 degrees, so high-angle lateral views remain the weak end.","Left and right hands are handled through horizontal flip augmentation, removing a manual mirroring step."],"supporting_citations":[{"why":"Supplies the statistical shape-and-kinematic wrist model used to propagate bone masks from neutral to all other CT poses.","marker":"[7]"},{"why":"Supplies the self-configuring encoder-decoder segmentation network and training scheme with Dice plus cross-entropy loss.","marker":"[10]"},{"why":"Grounds the orthographic 3D-to-2D X-ray simulation approach in prior image registration work.","marker":"[8]"},{"why":"Gives the attenuation equation used to compute image intensity from summed CT values.","marker":"[9]"},{"why":"Provides the interactive tool that produced the initial multi-label bone segmentation on neutral CT scans.","marker":"[6]"},{"why":"Defines the prior CNN method for ten wrist bones that was limited to manually labeled front-back images and serves as the comparison point.","marker":"[4]"},{"why":"Represents the alternative semi-automatic annotation route that excludes carpal bones.","marker":"[5]"}],"fun_headline_variants":["Simulated X-rays train wrist bone segmentation for real images","CT-based fake X-rays train model for real wrist segmentation","No manual X-ray labels needed: CT-simulated training works on real wrists","Single model segments 10 wrist bones from synthetic to real X-rays"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach depends on the assumption that the simplified simulated X-rays---parallel rays, no scatter, no energy dependence, and tissue intensities clipped by trial-and-error thresholds---look enough like real clinical radiographs that a model trained on them transfers to real images.","fun_headline_variants_meta":{"raw":{"variants":["Simulated X-rays train wrist bone segmentation for real images","CT-based fake X-rays train model for real wrist segmentation","No manual X-ray labels needed: CT-simulated training works on real wrists","Single model segments 10 wrist bones from synthetic to real X-rays"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000806,"raw_usage":{"total_tokens":3534,"prompt_tokens":932,"completion_tokens":2602,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2540}},"tokens_in":548,"tokens_out":2602,"duration_ms":20177,"temperature":1.0,"reasoning_tokens":2540,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:05:15.853559+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have a radiologist manually label the eight carpal bones and two forearm bones on the 121 real radiographs (or register the CT volumes to them to get ground truth) and compute Dice scores; if the real-image scores fall far below the simulated range of 0.80 to 0.92, the central claim of sim-to-real transfer fails.","supporting_citations":[{"cited_title":"Automatic Segmentation of Ulna and Radius in Forearm Radiographs","cited_arxiv_id":null,"evidence_quote":"Supplies the statistical shape-and-kinematic wrist model used to propagate bone masks from neutral to all other CT poses."},{"cited_title":"J Pers Med., 12(5):776, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the self-configuring encoder-decoder segmentation network and training scheme with Dice plus cross-entropy loss."},{"cited_title":"Multimed","cited_arxiv_id":null,"evidence_quote":"Grounds the orthographic 3D-to-2D X-ray simulation approach in prior image registration work."},{"cited_title":", Automatic inference and measurement of 3D carpal bone kinematics from single view fluoroscopic sequences","cited_arxiv_id":null,"evidence_quote":"Gives the attenuation equation used to compute image intensity from summed CT values."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the interactive tool that produced the initial multi-label bone segmentation on neutral CT scans."},{"cited_title":"8 carpal bones and 2 distal forearm bones)","cited_arxiv_id":null,"evidence_quote":"Defines the prior CNN method for ten wrist bones that was limited to manually labeled front-back images and serves as the comparison point."},{"cited_title":"Ethical approval was obtained for the use of this data","cited_arxiv_id":null,"evidence_quote":"Represents the alternative semi-automatic annotation route that excludes carpal bones."}],"review_version":1}