{"id":"2206ae1c-53c9-4aaf-b6af-642343982fd6","arxiv_id":"2608.06126","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A CT-based synthetic data generation framework improves AI-based patient pose assessment for ankle radiography by up to 11 percentage points when used for pretraining.","lead":"This paper builds a pipeline that turns existing CT scans into synthetic depth images and X-ray-like images of ankle joints, then uses them to train an AI that judges whether a patient is positioned well enough for a diagnostic X-ray. Pre-training on the synthetic data improved pose assessment accuracy on real ankle data by up to 11 percentage points.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 11 pp headline is built on a two-subject dataset with seeds, not subjects, as the statistical unit; the claimed effect size and its significance are not established.","rationale":"Good-faith reading: the paper proposes a practical framework, releases data, and reports consistent improvements across two real datasets. The strongest claim concerns the 11 pp gain on the two-preparation dataset. What must be true: the improvement is real and not an artifact of the chosen seed, configuration, or the idiosyncrasies of two anatomical preparations. That condition is least secure because the design has only two subjects in the headline experiment, and the significance test does not address subject-level variability. The reader's weakest assumption about synthetic-radiograph label validity is plausible but secondary: the transfer results themselves (Table 2 and linear probing in Table 4) suggest that depth features learned with synthetic labels are useful on real labels, so even imperfect synthetic labels do not obviously invalidate the transfer. The statistical issue, however, directly threatens the numeric headline. The clinical dataset (18 subjects) provides some replication for the general direction, but not for the specific 11 pp figure. I therefore agree with the conditional verdict, but for a different primary reason than the reader's stated weakest assumption. Credit: the authors disclose limitations in Section 7, release the datasets, and report a permutation test; the issue is the choice of independent unit, not a hidden flaw in the method itself.","tokens_in":17157,"tokens_out":9807,"duration_ms":90284,"concrete_test":"Using the released code/data, recompute Table 2's Accuracy and Diag. Acc. for the from-scratch and the best synthetic-pretrained model separately for each reciprocal preparation fold (prep A to B and B to A) across all 10 seeds. If the 11 pp improvement appears in only one fold, the headline effect is subject-specific and should be reported as such. Also rerun the Section 6.1 permutation test on Accuracy with multiplicity adjustment across the 8 pretraining configurations; if the headline configuration is no longer significant after adjustment, the 'up to 11 pp' claim should be downgraded to an exploratory finding.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract: 'up to 11 percentage points') rests on Table 2, Experiment 2a, where finetuning is on the anatomical preparations dataset containing only two subjects (Section 5.2: 'we trained on one preparation and tested on the other, and vice versa'). The reported statistical test (Section 6.1) treats the 10 repeated seeds as independent observations and tests MAE only, not the Accuracy metric that produces the headline 11 pp gain (79.25% to 90.45%). Moreover, the winning configuration (camera-view-specific, no augmentation) is selected from 8 synthetic-pretraining conditions in Table 2 without multiple-comparison correction. Because the two reciprocal folds yield only two subject-level observations, any seed-level permutation test cannot establish that the improvement generalizes to new subjects. If the 11 pp gain is not consistently present in both reciprocal folds, it could be driven by a single anatomical preparation. The secondary concern in Section 3.4 – that synthetic radiograph labels were only 'validated' by expert visual inspection, without quantitative agreement or comparison to real-radiograph labels – compounds the uncertainty, but the statistical/subject-size issue alone is sufficient to make the headline claim premature.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a framework that generates synthetic depth-image/radiograph pairs from CT scans, using Marching Cubes surface extraction, point-cloud augmentation, scene composition, and Monte Carlo radiograph simulation, to create a pretraining dataset of 3,077 upper-ankle poses from 10 CTs. The authors evaluate the synthetic data by training EfficientNet-B0 regressors and then transfer to two real datasets: a two-subject anatomical preparations dataset with radiograph-derived quality labels and a weakly labeled 18-subject clinical dataset. The central claim is that synthetic pretraining improves real-data pose assessment by up to 11 percentage points in Accuracy.","tokens_in":17409,"tokens_out":3820,"duration_ms":31815,"significance":"The framework addresses a real data-acquisition bottleneck in medical imaging, and the public release of the synthetic and real-world clinical datasets is a useful contribution. If the transfer results are reliable, the approach could reduce radiation exposure by supporting pose feedback before radiograph acquisition. The paper's strengths include a physically based radiograph synthesis pipeline that avoids deep-learning-based label leakage, a clear experimental setup with multiple pretraining variants, and an honest discussion of the weak labels in the clinical dataset. The main uncertainty lies in the statistical support for the headline effect size and in the validation of synthetic-radiograph labels.","major_comments":[{"comment":"The statistical analysis in Section 6.1 does not support the headline 'up to 11 pp' claim. The permutation test is applied to MAE only, whereas the abstract's 'up to 11 percentage points' refers to Accuracy (Table 2: 79.25% vs. 90.45%). Furthermore, the test treats the 10 random seeds as independent observations for a fixed dataset of two anatomical preparations. With only two subjects and reciprocal folds, the effective number of subject-level observations is two, so a seed-level test cannot establish that the improvement generalizes to new subjects. Please report the per-fold (per-preparation) Accuracy values and perform an analysis that respects the subject-level design, or revise the claim to reflect the strength of the evidence.","section":"6.1, Table 2"},{"comment":"The 'camera view specific without augmentation' configuration (90.45% Accuracy) is selected from nine pretraining conditions in Table 2, but no multiple-comparison correction or pre-registered selection procedure is described. Because the best configuration is identified on the same two-subject dataset on which it is then evaluated, the reported gain is at risk of selection bias. Moreover, this exact configuration (without augmentation) is not among the configurations tested on the independent real-world clinical dataset in Table 3, so the confirmation there does not directly validate the best Table 2 result. Please acknowledge this limitation, report all configurations with per-subject results, or validate the selected configuration on the clinical dataset.","section":"5.2, Table 2"},{"comment":"The synthetic radiographs are the sole source of labels for the synthetic pretraining data, but the manuscript reports only that 'expert radiologists have validated that the visual quality is suitable' (Section 3.4). No quantitative comparison between synthetic and real radiograph labels, no inter-rater agreement (e.g., Cohen's kappa or ICC), and no analysis of whether joint-space visibility is rendered equivalently in synthetic and real images is provided. If synthetic radiographs obscure or exaggerate joint space differently from real radiographs, the pretraining labels would be systematically biased. Please add a label-consistency experiment or explicitly restrict the claim to the specific rendering validated by the visual inspection.","section":"3.4"}],"minor_comments":[{"comment":"The grouping of the augmentation columns under 'unified camera view' and 'camera view specific' is visually ambiguous because the header line does not clearly separate the two strategies from the 'wo/ aug' column; please adjust the formatting for readability.","section":"Table 2"},{"comment":"The text says 'subject-level three-fold cross-validation' but does not specify how the left and right feet of the same subject are assigned to folds; since both feet of a subject are captured, please clarify whether they are kept together in the same fold.","section":"5.2, Experiment 2b"},{"comment":"The notation uses N both for the number of samples in Accuracy and for the number of subject-foot pairs in Pose Delta; please use distinct symbols to avoid confusion.","section":"5.4, Pose Delta"},{"comment":"The one-sided paired permutation test is described as exact, but the p-values are reported only as p=0.0049 and p=0.0020; please state the number of permutations used or confirm that an exact enumeration over the 2^10 possible sign flips was performed.","section":"6.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's strongest evidence for transfer is the anatomical preparations dataset with only two subjects, and the seed-level permutation tests in Section 6.1 are not a substitute for subject-level evidence. I recommend that the authors be asked to either add a subject-level analysis or substantially temper the 'up to 11 pp' claim. The paper otherwise appears technically sound and within the journal's scope, and the released datasets are a valuable contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you care about synthetic training data for medical imaging. The genuine news here is a pipeline that turns existing CTs into paired depth images and radiographs — combining marching cubes surface extraction, Monte Carlo DRR, scene composition, and augmentation — and the authors release the resulting datasets. They then show that pretraining on the synthetic pairs helps downstream pose assessment on two real datasets, with gains over both from-scratch and ImageNet initialization. That transfer result is the paper's centerpiece, and it's plausible: the effect is consistent across most metrics in Tables 2 and 3, and the linear probing experiment strengthens the case that the features are useful. The authors also disclose the main weaknesses themselves, which is a good sign.\n\nThe soft spot is the 11 percentage point headline. It comes from the anatomical preparations dataset, which has exactly two subjects — the finetuning is done on one preparation and tested on the other, and vice versa. That gives two subject-level observations. The reported significance test instead treats the 10 random seeds as paired observations, which inflates the evidence; seeds on the same fold are not independent. The test is also on MAE, while the 11 pp claim is about Accuracy. And the configuration chosen (camera-view-specific, no augmentation) is one of several compared in Table 2, without multiple-comparison correction. So the effect size is not established, even though the direction of the effect is believable.\n\nA secondary concern: the synthetic radiograph labels were validated only by expert visual inspection, not by quantitative comparison to real radiographs or inter-rater reliability. The real-world clinical dataset only has weak surrogate labels. Both are acknowledged in the text. These are limitations, not fatal flaws.\n\nAll in all, this is a solid, honest piece of engineering research. It deserves a serious referee — the pipeline and datasets are valuable, and the transfer claim is testable. But the authors should be asked to either temper the headline or back it with a subject-level analysis. I'd bring it to a reading group focused on medical synthetic data.","headline":"Plausible, well-engineered pipeline for synthetic depth/radiograph pairs from CT, but the headline 11 pp transfer gain is built on two subjects and seed-level statistics.","tokens_in":17951,"tokens_out":2960,"would_cite":true,"duration_ms":22167,"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":"A CT-based framework that generates synthetic depth-image/radiograph pairs from existing scans can train pose-assessment networks that transfer to real patients, improving accuracy by up to 11 percentage points.","keywords":["Patient Pose Assessment","Synthetic Data Generation","Diagnostic Quality","CT Scan","Time-of-Flight Cameras","Radiography","Deep Learning"],"falsifier":"Measure inter-rater agreement between expert quality ratings of synthetic radiographs and real radiographs of the same anatomical pose; near-chance agreement, or a fine-tuning experiment in which synthetic pretraining performs worse than from-scratch training, would falsify the transfer claim.","tokens_in":16988,"feed_emoji":"🩻","tokens_out":10858,"duration_ms":69301,"temperature":0.7,"pith_summary":"The paper claims that pairs of depth images and radiographs needed to train an AI pose assessor can be produced synthetically from existing CT scans, and that this synthetic data transfers to real patients. Using 10 CTs of ankles, the framework generates 3,077 pose-labeled pairs; pretraining on these images before fine-tuning on real depth images raises pose-assessment accuracy on anatomical preparations from 79.25 percent to 90.45 percent, an 11 percentage-point gain. Gains also appear on a new weakly labeled dataset of 18 living subjects, and synthetic-pretrained features beat ImageNet features in linear probing. If correct, the approach removes a major data-acquisition bottleneck: it creates clinically relevant training labels, including deliberately bad poses, without exposing new subjects to radiation.","feed_headline":"Synthetic CT data lifts ankle pose accuracy by 11 points","feed_subtitle":"Pretraining on 3,077 generated depth/radiograph pairs beats ImageNet for judging real X-ray poses.","key_machinery":"The load-bearing object is the data-generation pipeline that creates paired depth images and radiographs from a single CT. Surface extraction via Marching Cubes at -500 Hounsfield units yields a patient point cloud; normal-vector shifts and Gaussian bumps simulate body-shape variation; scene composition aligns the point cloud onto a recorded X-ray table so depth images look like real time-of-flight camera frames; and a Monte Carlo simulation using voxel material and mass-density maps renders the paired radiograph, with expert radiologist labels for joint-space visibility defining ground truth. These paired synthetic images are then used for task-specific pretraining of two EfficientNet-B0 regressors, one per camera view, with camera-view-specific weight initialization.","core_discovery":"The central claim is that a CT-based synthetic data-generation framework can substitute for scarce real depth-image/radiograph pairs in training pose-assessment networks. From each CT, the framework extracts the body surface as a point cloud, augments it to vary body shape and simulate edema-like bumps, places it on a recorded X-ray-room table, rotates it through 181 medial poses per foot, renders depth images from two time-of-flight camera views, and produces a matching synthetic radiograph by Monte Carlo X-ray transport with material-specific voxels. Radiologists label each synthetic radiograph's diagnostic quality from 1 to 3, and this label is attached to all corresponding depth images. Pretraining two EfficientNet-B0 regressors on these synthetic images, then fine-tuning on real data, improves accuracy on the two-preparation dataset to 90.45 percent versus 79.25 percent from scratch and 77.37 percent with ImageNet pretraining; the same recipe improves pose ranking on a living-subject dataset with surrogate labels. The paper reads this as evidence that synthetic CT-derived labels are usable proxies for real radiographic quality in pose assessment.","pith_inferences":["The paper's own suggestion that quality labels could be assigned automatically, for example with an existing radiograph-quality network, points to a fully automated loop: CT in, labeled depth-image training set out; the reported experiments do not test whether an automated labeler would preserve the 11-point gain.","Because the largest gain comes from camera-view-specific pretraining and task-specific size augmentation, the mechanism may be domain-adapted feature learning rather than radiograph realism; an ablation that swaps Monte Carlo radiographs for cheaper ray-traced projections would separate photorealism from label-transfer value.","The real-world clinical dataset uses surrogate positioning labels rather than radiographic ground truth, so the actual clinical benefit of fewer retakes and less radiation remains untested until a prospective study measures retake rates."],"forward_implications":["Pretraining on synthetic CT-derived depth images improves downstream pose assessment on real data by up to 11 percentage points over training from scratch, and by more than ImageNet pretraining.","Depth images alone can predict radiograph diagnostic quality, reaching 90.45 percent accuracy compared with the 93 percent reported for assessing real radiographs directly.","Retrospective CT scans can be turned into large training sets with deliberately non-diagnostic poses, avoiding the ethical and regulatory problem of radiographing subjects without indication.","Synthetic-pretrained features transfer to a weakly labeled living-subject dataset, improving pose margin accuracy by about 8 percentage points over from-scratch training."],"supporting_citations":[{"why":"Establishes the premise that patient pose is the dominant factor in radiograph diagnostic quality, motivating the pose-assessment task.","marker":"Little et al. (2017)"},{"why":"Supplies the anatomical preparations dataset of real depth images and radiographs and the baseline pose-assessment results that synthetic pretraining improves.","marker":"Laufer et al. (2024)"},{"why":"Provides the radiograph diagnostic-quality assessment used as reference accuracy (93 percent) and as the suggested automatic-labeling route for synthetic radiographs.","marker":"Mairhöfer et al. (2021)"},{"why":"Provides the MCGPU Monte Carlo X-ray transport implementation that renders synthetic radiographs from CT material and density voxels.","marker":"Badal and Badano., 2011"},{"why":"Supplies the PENELOPE material property files used by the Monte Carlo simulation to model photon-matter interaction.","marker":"Salvat et al. (2006)"},{"why":"Provides the Marching Cubes algorithm used to extract the patient surface point cloud from CT volumes.","marker":"Lorensen and Cline (1987)"},{"why":"Defines the ImageNet pretraining baseline that the paper compares against to show synthetic pretraining is more useful than generic natural-image features.","marker":"Deng et al. (2009)"},{"why":"Defines the EfficientNet-B0 architecture used for both depth-image regressors in all experiments.","marker":"Tan and Le (2019)"}],"fun_headline_variants":["CT-synthetic data boosts ankle pose accuracy by 11 points","Synthetic CT images improve ankle X-ray pose estimation","AI pose assessment gains 11 points from CT-generated synthetic data","CT-based synthetic radiographs train better ankle pose AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that expert quality labels assigned to synthetic radiographs are valid for real patient poses; if the synthetic joint-space appearance differs systematically from real radiographs, the learned transfer could be wrong.","fun_headline_variants_meta":{"raw":{"variants":["CT-synthetic data boosts ankle pose accuracy by 11 points","Synthetic CT images improve ankle X-ray pose estimation","AI pose assessment gains 11 points from CT-generated synthetic data","CT-based synthetic radiographs train better ankle pose AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000267,"raw_usage":{"total_tokens":1590,"prompt_tokens":898,"completion_tokens":692,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":625}},"tokens_in":514,"tokens_out":692,"duration_ms":5300,"temperature":1.0,"reasoning_tokens":625,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:25:07.487773+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure inter-rater agreement between expert quality ratings of synthetic radiographs and real radiographs of the same anatomical pose; near-chance agreement, or a fine-tuning experiment in which synthetic pretraining performs worse than from-scratch training, would falsify the transfer claim.","supporting_citations":[{"cited_title":"and Reiser, Ingrid and Liu, Lili and Kinsey, Tiffany and S","cited_arxiv_id":null,"evidence_quote":"Establishes the premise that patient pose is the dominant factor in radiograph diagnostic quality, motivating the pose-assessment task."},{"cited_title":"Medical Imaging 2024: Image Processing , editor =","cited_arxiv_id":null,"evidence_quote":"Supplies the anatomical preparations dataset of real depth images and radiographs and the baseline pose-assessment results that synthetic pretraining improves."},{"cited_title":"Chapter 50 - Fast Simulation of Radiographic Images Using a Monte Carlo x-Ray Transport Algorithm Implemented in","cited_arxiv_id":null,"evidence_quote":"Provides the MCGPU Monte Carlo X-ray transport implementation that renders synthetic radiographs from CT material and density voxels."}],"review_version":1}