{"id":"9d2ef082-e021-46b1-a94d-ed1e33e49ba3","arxiv_id":"2607.18638","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"A weighted ensemble of ConvNeXt, ViT, and MaxViT estimates adult sex, age, height, and weight from coronal CT-derived DRRs with sex accuracy 0.997 and MAEs of 3.57 years, 2.59 cm, and 3.40 kg on an institution-external test set.","lead":"This paper trains an ensemble of three deep learning models to estimate a patient's sex, age, height, and weight from a two-dimensional projection (DRR) generated from their CT scan, reaching sex accuracy above 99% and age error near 3.6 years on an external hospital test set. It matters because large medical imaging databases often lack these metadata, and the model could fill missing values for research, dose optimization, and body-size normalization.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DICOM metadata label reliability is the main residual risk, but it does not invalidate the recorded-attribute claim; no verdict change needed.","rationale":"I read the paper in good faith. The ensemble is trained and evaluated on a large multi-institutional J-MID sample, with institution-level splitting, one examination per patient in tuning/test, patient-level bootstrap CIs, public code/weights, and additional cross-domain datasets. These are substantive supports for the main quantitative claim. The central weak point is exactly the one the reader identified: DICOM metadata are treated as ground truth for age, height, and weight. The Discussion explicitly lists this as a limitation. I considered other potential concerns—possible cross-institution patient overlap, clipping of standardized labels at ±3 SD, and the qualitative nature of the BSA proof-of-concept—but none is as directly tied to the headline accuracy numbers as label quality. The DICOM-label concern is real but bounded: for the stated purpose of imputing missing metadata in databases, agreement with recorded metadata is the relevant performance metric, and the paper reports that honestly. Therefore I agree with the reader's ACCEPT verdict and would not change it, though a small external label audit would materially strengthen the interpretation.","tokens_in":12633,"tokens_out":7765,"duration_ms":87861,"concrete_test":"For a random sample of 200–500 J-MID test patients, independently re-extract age, height, and weight from source electronic health records or same-date clinical measurements, without revealing model predictions. Recompute sex accuracy and MAEs against these verified labels and compare with Table 3. If the verified-label MAEs differ by more than about 0.5 cm (height) or 0.5 kg (weight), or if classification accuracy shifts more than 0.002, the reported estimates need to be reframed as database-record agreement rather than biological accuracy. If the verified-label metrics are consistent, the DICOM-label concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline numbers (test MAEs of 3.57 years, 2.59 cm, 3.40 kg; sex accuracy 0.997) measure agreement with DICOM metadata, not with independently verified biological measurements. The Discussion explicitly acknowledges that height/weight may be self-reported, measured at a different time, or transcribed inconsistently. If labels are noisy, reported MAEs include label noise; if labels are systematically biased, the model may be partly learning to reproduce a database artifact. This is genuinely load-bearing because the stated downstream use—imputing missing metadata and reproducing BSA-corrected organ-volume trends—is framed around replacing recorded values. Notably, for that use case, agreement with the database record is the correct target, so the limitation does not undermine the paper's core empirical claim. What is missing is an independent label audit quantifying how much of the reported error is model error versus label error. The authors flag this as a limitation but do not test it; that is the single most important gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops and validates a multitask deep learning ensemble (ConvNeXt-Base, ViT-Base/16, MaxViT-Base) that estimates adult sex, age, height, and weight from coronal digitally reconstructed radiographs (DRRs) generated from diagnostic CT. The models were trained on 114,147 examinations from seven Japanese institutions, tuned on 4,305 examinations from one institution, and tested on 10,169 examinations from another institution. The weighted ensemble achieved a sex-classification accuracy of 0.997 (95% CI, 0.996–0.998) and MAEs of 3.57 years, 2.59 cm, and 3.40 kg for age, height, and weight on the external test set. Performance was generally better for examinations with broader anatomical coverage. A proof-of-concept showed that BSA computed from estimated versus recorded height/weight reproduces age-related heart and liver volume trends. Cross-domain validation on ENHANCE.PET and RATIC is reported, with continued fine-tuning reducing height error on non-Japanese data. The authors make model weights and inference code publicly available.","tokens_in":12918,"tokens_out":10997,"duration_ms":111730,"significance":"If the results hold, this work provides a scalable tool for imputing missing demographic and anthropometric metadata in large retrospective CT databases, with implications for radiation dose estimation, body-size normalization, and epidemiological analyses. The paper has notable strengths: a very large multi-institutional training set, institution-level disjoint tuning and test splits, patient-level bootstrap confidence intervals, per-anatomical-region performance reporting, explicit cross-domain evaluation, and public release of code and weights. The claims are carefully bounded in most of the manuscript, and the main limitation—use of DICOM metadata as the reference standard—is disclosed. These practices are exemplary and increase confidence in the reported performance.","major_comments":[],"minor_comments":[{"comment":"The dependence on DICOM metadata is acknowledged, but the manuscript repeatedly refers to 'true' age, height, and weight (e.g., Figure 3, Figure 5, and the BSA analyses). This conflates agreement with the clinical record with agreement with independently measured biological values. Please replace 'true' with 'reference' or 'recorded' throughout, and add a sentence in the Abstract or Conclusion clarifying that reported errors are relative to DICOM metadata. This is especially important because height/weight may be self-reported or measured at a different time.","section":"Discussion, Limitations (paragraph 4)"},{"comment":"The flowchart appears to display the two exclusion criteria in reverse order. The narrative states that eligibility required available anthropometric attributes and age ≥18, with subsequent exclusion for insufficient vertebral coverage, but the boxes list 'Examinations with ≥1 axial image showing ≥3 vertebrae: 136,272' above 'Cases with available anthropometric information: 169,024.' The implied arithmetic (283,132 − 114,108 = 169,024; then − 32,752 = 136,272) should be clarified by reordering the boxes or adding explicit arrows.","section":"Materials and Methods: Case selection (Figure 2)"},{"comment":"The text says 'two non-Japanese datasets' were used, but RATIC provides sex and age only, while height and weight are not evaluated. For clarity, indicate in Table 3 or the main text which attributes are evaluated on each cross-domain dataset.","section":"Materials and Methods: Dataset (cross-domain)"},{"comment":"Table 2 reports the ensemble's performance on the tuning set, which was also used for early stopping and for optimizing the ensemble weights. This is not an unbiased performance estimate. It would be helpful to label this as a development-set result and note that the test set is the primary external validation.","section":"Results: Table 2"},{"comment":"The canvas size of 640 × 320 pixels at 1.5 mm/pixel corresponds to 960 × 480 mm. Please clarify whether the full body is contained within this canvas for the largest patients. If cropping occurs, describe how the 'real-world-scale-preserving' property interacts with the fixed canvas and how height estimation remains robust when the body extends beyond the field of view. A sentence in the text or a supplementary figure showing the canvas relative to the full DRR extent would resolve this ambiguity.","section":"Materials and Methods: DRR Generation and Preprocessing"},{"comment":"Given that the downstream use case is imputing missing metadata, agreement with the database record is the relevant target. However, the absence of an independent label audit means that the proportion of the reported MAE attributable to label noise is unknown. Even a small subsample audit, or at least a quantitative discussion of how self-reported height/weight errors would propagate through the reported MAEs, would strengthen the interpretation.","section":"Discussion, Limitations (paragraph 4)"}],"recommendation":"minor_revision","confidential_remarks":"The paper is technically sound and the central empirical claim is well supported. The DICOM-metadata limitation is transparently disclosed and does not invalidate the results for the stated application of metadata imputation. The requested revisions are primarily clarifying language and a few presentation details; no additional experiments are required for acceptance. I suggest the editors encourage the authors to review all uses of 'true' and to confirm that the flowchart matches the case-selection narrative."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, well-executed application paper that deserves a serious referee. The headline numbers are credible because the evaluation is properly separated from training and the authors report per-region and cross-domain results. The main caveat—DICOM metadata treated as ground truth—is real but doesn't sink the central claim, since the paper is really about reproducing the recorded attributes, not biological truth.\n\nWhat's new is the scale and scope: four attributes from diagnostic CT coronal DRRs, trained on 114k exams from seven institutions, tuned and tested on disjoint institutions, plus two non-Japanese public datasets. That's a meaningful step beyond prior single-attribute or scout-image work. The per-region breakdown (Torso vs Chest vs Abdomen etc.) is a nice touch and shows the model behaves sensibly—broader coverage gives better estimates. The BSA proof-of-concept is a reasonable demonstration of downstream utility, and the public code/weights make it reproducible.\n\nThe soft spots are mostly acknowledged in the paper. The DICOM-label issue is the biggest: we don't know if height/weight were measured or self-reported, or when recorded relative to the scan. That adds label noise to the reported MAEs and could partly explain cross-dataset gaps. The authors flag it but don't quantify it, so an independent label audit would be the most valuable addition. Weight underestimation above 100 kg is visible in the scatter plots and likely due to target clipping and the skewed training distribution; again, acknowledged. Cross-population generalization is uncertain, but the continued fine-tuning experiment is an honest attempt to address it.\n\nThe circularity concern doesn't land: the ensemble weights are fixed on a separate tuning set, and the BSA analysis fits nothing. No self-citation inflation. The paper is careful about separating tuning and test, and the bootstrap CIs are done at the patient level.\n\nWho's this for? Anyone working on metadata imputation in large CT databases, dose optimization, or body-size normalization. It's a practical contribution rather than a methodological breakthrough, but it's carefully done and honestly reported. I'd send it to peer review—it deserves referee time, and the main revision request should be an effort to audit or at least subsample DICOM labels against source records.","headline":"Solid, well-validated application paper; the DICOM-label caveat is real but the central claim, agreement with recorded metadata, holds.","tokens_in":13365,"tokens_out":1844,"would_cite":true,"duration_ms":18510,"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":"One coronal X-ray-like projection of a CT scan, passed through an ensemble of three neural networks, predicts an adult's sex, age, height, and weight—with 99.7% sex accuracy and mean absolute errors of 3.6 years, 2.6 cm, and 3.4 kg.","keywords":["computed tomography","digitally reconstructed radiograph","multitask deep learning","demographic estimation","anthropometric estimation","sex classification","age estimation","body surface area normalization"],"falsifier":"Measure height and weight on the same day as a CT scan in a new cohort spanning the full weight range, run the model, and compare its outputs to those measured values; if the mean absolute errors are much larger than 2.59 cm and 3.40 kg, or the systematic underestimation above 100 kg remains, the central claim that the model recovers true body size from CT would be refuted.","tokens_in":12595,"feed_emoji":"🩻","tokens_out":7134,"duration_ms":62808,"temperature":0.7,"pith_summary":"The paper aims to show that a patient's sex, age, height, and weight can be recovered from diagnostic CT images alone, using a single synthetic X-ray projection per scan. The authors train a three-model deep-learning ensemble on more than 128,000 CT examinations from adults across nine institutions and test it on an institution never seen during training. The ensemble achieves near-perfect sex classification and small errors in age, height, and weight, with the best results on scans covering the full torso. If this holds, it would allow researchers to fill in missing or inconsistent patient metadata in large retrospective imaging databases and to perform body-size normalization of organ measurements without recorded height or weight.","feed_headline":"One CT projection yields sex, age, height, and weight","feed_subtitle":"Near-perfect sex detection and small errors in age, height, and weight could restore missing metadata in imaging databases.","key_machinery":"The central mechanism is the digitally reconstructed radiograph (DRR): a single coronal, parallel-projection X-ray-like image computed from each 3D CT volume using Beer–Lambert line integrals, with voxel spacing preserved so that real-world scale (and thus body-size information) is retained in the 2D image. This 640×320-pixel projection is fed to three deep encoders—a convolutional network and two vision transformer variants—sharing a multitask head that predicts sex as a classification and age, height, and weight as regressions. The three models' outputs are combined by weighted averaging, with the ensemble weights optimized on a separate tuning set. Preserving physical scale in the DRR is","core_discovery":"The paper's central claim is that a weighted ensemble of three deep image encoders—fine-tuned on coronal digitally reconstructed radiographs (DRRs) generated from diagnostic CT—can simultaneously estimate adult sex, age, height, and weight. On an institution-external test set of 10,169 examinations, sex-classification accuracy reached 0.997 (95% CI, 0.996–0.998), and mean absolute errors were 3.57 years for age, 2.59 cm for height, and 3.40 kg for weight. Performance was best for examinations covering the chest through pelvis (Torso), where sex accuracy hit 1.000 and errors dropped to 3.15 years, 2.28 cm, and 3.18 kg. The authors further demonstrate that body surface area computed from the m","pith_inferences":["- Because the model relies only on a coronal projection, adding a lateral projection or full 3D context could reduce weight error at the high end, where the paper already observes systematic underestimation above 100 kg; this is a testable architecture extension the paper leaves implicit.","- The near-ceiling sex classification implies CT images encode demographic information that can be read without patient metadata; this has privacy implications for anonymized imaging databases, which the paper itself flags, and suggests that de-identification of images alone does not hide sensitive attributes.","- The method could serve as a quality-control tool for imaging databases: large discrepancies between stored metadata and model estimates might flag transcription errors or incorrect patient records, a use the paper does not develop.","- The finding that height transfer degrades across populations with different average statures suggests the model learns population-specific height priors; a straightforward test would be to train on a multi-ethnic dataset and measure whether the height MAE drops for all groups."],"forward_implications":["- If the accuracy holds, researchers can recover sex, age, height, and weight from CT archives even when those fields are missing or corrupted, enabling secondary studies on existing scans.","- The performance gradient with anatomical coverage means full-torso scans are the most reliable source for these estimates; partial-coverage scans still provide useful but less precise estimates.","- Body surface area calculated from model-estimated height and weight can replace measured BSA in normalizing organ volumes for age-related trends, at least at the population level.","- On new populations with different body-size distributions, direct transfer degrades particularly for height, but continued fine-tuning from the original model substantially reduces this error, suggesting the pretrained model is a strong starting point for adaptation."],"fun_headline_variants":["Single CT projection yields sex, age, height, and weight","One CT-derived X-ray predicts four body metrics","Deep learning estimates sex, age, height, weight from 2D CT","AI predicts sex, age, height, weight from a single DRR"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The study's numbers assume that the age, height, and weight stored in each CT file are the ground truth; if those records are self-reported, outdated, or inconsistently entered, the reported errors measure agreement with the database rather than with true measurements.","fun_headline_variants_meta":{"raw":{"variants":["Single CT projection yields sex, age, height, and weight","One CT-derived X-ray predicts four body metrics","Deep learning estimates sex, age, height, weight from 2D CT","AI predicts sex, age, height, weight from a single DRR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001224,"raw_usage":{"total_tokens":4977,"prompt_tokens":962,"completion_tokens":4015,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":706,"completion_tokens_details":{"reasoning_tokens":3942}},"tokens_in":706,"tokens_out":4015,"duration_ms":38639,"temperature":1.0,"reasoning_tokens":3942,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T14:45:28.585991+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure height and weight on the same day as a CT scan in a new cohort spanning the full weight range, run the model, and compare its outputs to those measured values; if the mean absolute errors are much larger than 2.59 cm and 3.40 kg, or the systematic underestimation above 100 kg remains, the central claim that the model recovers true body size from CT would be refuted.","supporting_citations":[],"review_version":1}