{"id":"5cfe71c5-b5a1-4b5b-9554-ddc917b5ff1a","arxiv_id":"2506.20407","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"It reports that fusing radiomic and deep features estimates gestational age with 8.0 days error, but the label is computed from the same head contours that feed the radiomics.","lead":"This paper combines hand-crafted radiomic features and deep learning image features to estimate fetal gestational age from ultrasound images, reporting a mean absolute error of 8 days. The approach is meant to reduce reliance on manual fetal measurements, which may help settings with few expert sonographers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 8.0-day MAE is contaminated by label leakage: the GA target is computed from the head-circumference contour in the same ROI mask that supplies radiomic shape features.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the ground-truth GA is derived from the same ROI contour that provides radiomic shape features, making the evaluation self-referential. My independent reading confirms this and sharpens it: the leakage is not just conceptual but mechanistic, because the label is a monotone function of contour pixel count, and contour perimeter is a standard radiomic shape feature. The central claim as stated, that the framework estimates GA with 8.0-day MAE and outperforms current ML methods at these gestational ages, is therefore not supported as a claim about clinical GA accuracy. The paper does provide useful engineering components, publicly available code, and ablation studies, and the cross-attention fusion may be a valid contribution, but the evaluation protocol invalidates the headline performance figure. Because my concern coincides with the reader's, the existing REJECT verdict stands without modification.","tokens_in":7960,"tokens_out":2990,"duration_ms":37465,"concrete_test":"Use the released code to rerun the full pipeline with all shape-based radiomic features removed (or with only first-order and texture features retained), keeping the label definition unchanged. If the MAE rises from 8.0 days toward or above the image-only baseline of about 8.6 days, the reported result is driven by the leaked contour perimeter. As a complementary check, compute the MAE of a trivial regressor that predicts GA directly from the mask perimeter via the INTERGROWTH formula without any image features: near-zero error would demonstrate the leakage directly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 defines the target y^(i) by first computing HC = pnum × psize, where pnum is the number of pixels along the edge of the ROI mask X_ROI, and then applying the INTERGROWTH-21st formula GA = exp[0.05970 × (log_e(HC))^2 + 0.000000006409 × (HC)^3 + 3.3258]. The same X_ROI is used as input to the radiomics pipeline in Section 2.3, which extracts 95 features including 2D shape/perimeter descriptors. The label is therefore a deterministic monotone function of a quantity directly present in the model's input feature set. This is label leakage, not merely self-reference: a trivial algorithm that measures contour length from the provided mask and applies the formula would achieve near-perfect MAE on this target. Consequently, the reported 8.0-day MAE does not establish clinically meaningful gestational-age estimation; it establishes that the fusion model can reproduce the HC-derived formula from the provided contour. The abstract's claim of estimating GA 'without any measurement information' is also undermined, because the ROI masks are expert annotations that are used as inputs, and the comparison against image-only DL baselines is not apples-to-apples. The load-bearing assumption that formula-derived HC is a valid ground-truth for this evaluation fails because the same contour that defines the ground truth also defines a dominant predictive feature.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a feature-fusion framework for gestational age (GA) estimation from 2D fetal head ultrasound images, combining deep representations from a ConvNeXt backbone with 95 pyradiomics features extracted from an associated ROI mask, fused through a cross-attention module. The authors evaluate on the ES-TT and HC18 datasets, computing each image's GA label from head circumference (HC) obtained as the pixel count along the ROI contour times the pixel size, followed by the INTERGROWTH-21st formula. They report a best mean absolute error (MAE) of 8.0 days and claim superiority over image-only deep-learning baselines and radiomics-only machine-learning baselines, with code publicly available.","tokens_in":8260,"tokens_out":6821,"duration_ms":78675,"significance":"If the evaluation were valid, the framework could offer an interpretable, automatic alternative to manual biometry. The paper ships public code, compares several CNN/ViT backbones, and includes an ablation of the cross-attention module. However, the ground-truth label is constructed from the same ROI contour that generates the radiomic shape features used as input. The central quantitative claim is therefore not supported: the reported MAE largely measures how well the model can exploit a contour-derived measurement to invert a deterministic formula, rather than how well it estimates true clinical GA. The comparison to image-only baselines is also confounded by the additional mask input. These issues make the headline result uninterpretable as a statement about estimating gestational age in practice.","major_comments":[{"comment":"The GA label is defined by y = exp[0.05970 × (log_e(HC))^2 + 0.000000006409 × HC^3 + 3.3258] with HC = pnum × psize, where pnum is the number of pixels along the edge of the ROI mask X_ROI. The same X_ROI is used in Section 2.3 to extract 95 radiomic features that include Shape2D/perimeter descriptors. Thus the target is derived from the same contour that supplies a dominant input-feature family. Even though psize is not part of the explicit feature vector, the model has direct access to the contour from which HC is computed, and a model that reads contour length can reproduce the label once it learns the per-image scalar. The paper reports no control experiment with perimeter-only features, with all shape features withheld, or with masks withheld. The 8.0-day MAE therefore cannot be interpreted as evidence of clinically meaningful GA estimation; it is at least in part a test of whether the model can decode the HC-derived formula from the annotation.","section":"Section 3.1"},{"comment":"The baseline deep-learning models are trained on ultrasound images only, whereas the proposed framework receives both the image and an expert-annotated ROI mask, from which radiomics are computed. This is not an apples-to-apples comparison: part of the reported improvement over ResNet18 (from 9.9 to 8.0 days) may simply reflect access to the mask rather than the fusion mechanism. The authors should evaluate image-only, mask-only, and image-plus-mask variants of the same fusion pipeline, and ideally an automatic-segmentation setting. Without such matched baselines, the claim of superiority over image-based methods is unsubstantiated. The abstract's statement that the method works 'without any measurement information' also conflicts with the use of expert ROI masks as inputs.","section":"Section 3.3, Table 1"},{"comment":"The paper's across-trimesters claim is based on formula-derived GA from ROI contours, not on any externally established gestational age. For HC18, the dataset provides HC values, but the authors recompute HC from the contour and use that to form the target rather than validating against the provided HC or against a clinical reference such as first-trimester crown-rump length or last menstrual period. Consequently, the conclusion that the method 'outperforms current machine learning-based methods at these gestational ages' is not supported for actual pregnancy dating; it is a statement about fitting the INTERGROWTH-21st formula to contour lengths. An external evaluation with independently dated pregnancies is required before any clinical claim can be made.","section":"Section 3.1 and Section 3.3"}],"minor_comments":[{"comment":"The paragraph preceding Table 2 says the ML models are evaluated 'without using the radiomic features,' but the experimental setup in Section 2.5 and the table itself indicate that these models use radiomic features exclusively; this wording should be corrected.","section":"Section 3.3, Table 2"},{"comment":"The P-value column contains entries only for a few rows, with no description of the statistical test, the number of repeated runs, or multiple-comparison correction; the table should either present complete test information or omit the column.","section":"Section 3.3, Table 1"},{"comment":"The abbreviations IDMN, SDHGE, SAGLE, SRHGLE, and HGLE are not defined in the caption or in the text; they should be expanded or referenced to the pyradiomics documentation.","section":"Fig. 3 caption"},{"comment":"Since HC18 provides HC values in millimeters, the manuscript should state whether the contour-derived HC used for the label matches the provided HC values, and if not, why the recomputation is preferable.","section":"Section 3.1, HC18 dataset"}],"recommendation":"reject","confidential_remarks":"The proposed cross-attention fusion architecture is reasonable and the authors provide public code, but the evaluation design has a load-bearing circularity: the label is computed from the same ROI contour that provides the radiomic shape features. Fixing this requires independent gestational-age labels and matched baselines, which would constitute a substantially different study. I see no evidence of misconduct, but the current manuscript's central claim cannot be supported by the presented experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe headline number in this paper—8.0 days MAE—is not a valid estimate of gestational age. The target is computed from the head-circumference contour in the same ROI mask that feeds the radiomic shape features. Section 3.1 defines GA via INTERGROWTH-21st with HC = pnum × psize, where pnum is the number of pixels along the ROI edge. The radiomics pipeline (Section 2.3) extracts 95 features from that same mask, including 2D shape descriptors. The label is thus a monotone function of an input feature. A model that simply measures the contour and applies the formula would nearly nail this target. The reported MAE is internal consistency, not clinical accuracy.\n\nWhat is new and worth credit: the specific fusion of pyradiomics features with ConvNeXt deep representations through a cross-attention module for fetal head ultrasound. The module is plug-and-play, the code is public, and the authors run a reasonable set of backbones and ML baselines. The ablation and GradCAM/attention visualizations are thoughtful. The engineering is clean.\n\nThe soft spot is load-bearing. The abstract's claim of estimating GA 'without any measurement information' is misleading because the ROI masks are expert annotations that carry the head-circumference measurement. The comparison against image-only DL baselines is not apples-to-apples: those baselines do not receive the mask, while the proposed method does. The ML baselines in Table 2, which use radiomic features alone, report 21–27 days MAE, which is puzzling given the leak; but that does not rescue the central result. Even if the ML baselines fail to exploit the shape feature, the fusion model can and probably does. The 'outperforming' claim is therefore unsupported.\n\nThe paper should not be accepted as is. It is, however, a coherent and well-written manuscript with a fixable flaw. A serious referee could ask for re-evaluation with an external GA standard (e.g., first-trimester crown-rump length or last menstrual period) and a fair comparison that either excludes the masks or treats them honestly. As it stands, the central claim does not hold.\n\nI'd send this to peer review only with the expectation of major revision, not desk reject, because the topic is relevant and the code will let the authors fix the experiment. If you want a cautionary tale for a reading group on label leakage, this is a good one.\n\nBest,\nYour name","headline":"The 8-day MAE is an artifact of label leakage—the GA target is computed from the same ROI contour that supplies the radiomic shape features—so the headline result does not establish clinically valid GA estimation.","tokens_in":8785,"tokens_out":3429,"would_cite":false,"duration_ms":37397,"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":"Fusing radiomic features with deep image representations estimates fetal gestational age with a mean absolute error of 8.0 days.","keywords":["gestational age estimation","fetal ultrasound","radiomics","deep learning","feature fusion","cross-attention","head circumference","interpretability"],"falsifier":"Calculate the mean absolute error of the framework after removing all shape-based radiomic features; if the 8.0-day accuracy substantially degrades, the result is largely explained by the model reading the contour that defines the label. A second falsifier is to evaluate against first-trimester or known-conception gestational age: if the 8.0-day error does not survive an independent dating standard, the claimed gestational age accuracy is an artifact of the formula-derived target.","tokens_in":7803,"feed_emoji":"🤰","tokens_out":5593,"duration_ms":52627,"temperature":0.7,"pith_summary":"This paper tries to show that gestational age can be estimated automatically from fetal head ultrasound images and their region-of-interest masks by fusing two kinds of features: hand-crafted radiomic features that describe shape, intensity, and texture, and deep representations learned by a convolutional network. The authors report a mean absolute error of 8.0 days across three trimesters on two public datasets, which they claim outperforms current machine-learning-based methods for these gestational ages. The practical goal is an operator-independent dating tool that does not require manual biometric measurements and that offers a degree of interpretability through the radiomic features.","feed_headline":"Fetal age hits 8.0-day error with deep-radiomic fusion","feed_subtitle":"Automated dating from ultrasound images and masks works without manual biometry across two public datasets.","key_machinery":"The central mechanism is the cross-attention fusion module, which projects the standardized radiomic feature vector into queries and the deep image representation into keys and values, then applies scaled dot-product attention to produce a fused representation that a two-layer MLP regresses to gestational age. The radiomic features consist of 95 shape, first-order, and texture descriptors extracted from the fetal head region of interest, and the deep representation comes from a ConvNeXt convolutional network pre-trained on ImageNet and fine-tuned on the ultrasound data. The cross-attention module is the component that the ablation study credits for the accuracy gain: without it, concatenation gives no improvement over the image-only baseline, while adding it lowers the mean absolute error from 8.6 to 8.0 days.","core_discovery":"The paper's central claim is that a cross-attention fusion of deep image features and radiomic features yields more accurate gestational age estimates than either feature type alone, simple concatenation, or the underlying CNN by itself. In their experiments, the ConvNeXt-based image model fused with 95 radiomic features through cross-attention reaches a mean absolute error of 8.0 days, improving on the same model without fusion by 0.6 days and on the ResNet18 baseline by 1.9 days, and the cross-attention fusion outperforms concatenation across all tested architectures. The authors interpret this as evidence that radiomic features, which are inherently interpretable, complement deep representations for fetal brain maturation.","pith_inferences":["Because the gestational age labels are computed from the same head-circumference contour that feeds the radiomic shape features, a direct test would be to remove the shape features and see whether the fusion advantage persists; if it does not, part of the 8.0-day accuracy may be the model recovering the label formula rather than learning image-based dating.","A stronger clinical validation would compare the framework's estimates against first-trimester crown-rump length dating or known conception dates, which would tell whether the method adds information beyond reproducing the INTERGROWTH-21st curve.","The cross-attention improvement might also be tested with an ablation that replaces radiomic queries with random features to check how much of the gain comes from the specific radiomic content rather than the attention mechanism itself."],"forward_implications":["Automated gestational age dating becomes feasible without manual biometry, potentially reducing operator dependence in ultrasound screening.","The cross-attention fusion module is plug-and-play and improves accuracy across several CNN backbones, so it could be transferred to other fetal planes or other imaging tasks.","Because radiomic features drive part of the prediction, the model offers a route to explainable dating, with attention weights indicating which shape or texture properties matter.","The authors state the framework can be extended to other fetal anatomies such as the abdomen and femur, using the same fusion strategy."],"supporting_citations":[{"why":"Supplies the INTERGROWTH-21st formula that converts head circumference to gestational age, defining the regression target.","marker":"[21]"},{"why":"Provides the HC18 dataset with head circumference values and images, one of the two evaluation datasets.","marker":"[12]"},{"why":"Provides the ES-TT multi-center Spanish dataset with fetal head ROI annotations, the other evaluation dataset.","marker":"[4]"},{"why":"The ConvNeXt architecture is the deep backbone whose features are fused with radiomics and achieves the best reported result.","marker":"[17]"},{"why":"Provides the radiomics feature extraction used to compute the 95 hand-crafted features.","marker":"[9]"},{"why":"The image-based deep learning baseline whose ResNet18 result is the direct comparison for the reported improvement.","marker":"[15]"}],"fun_headline_variants":["Deep-radiomic fusion cuts fetal age error to 8 days","AI fuses radiomics and deep features for 8-day fetal dating","Fetal gestational age estimated to 8-day MAE via fusion","Cross-attention fusion: fetal age in 8 days without biometry"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that gestational age computed from head circumference via the INTERGROWTH-21st formula, with the head circumference derived from the annotated ROI contour in the same image, is a valid ground truth for measuring gestational age estimation accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Deep-radiomic fusion cuts fetal age error to 8 days","AI fuses radiomics and deep features for 8-day fetal dating","Fetal gestational age estimated to 8-day MAE via fusion","Cross-attention fusion: fetal age in 8 days without biometry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1515,"prompt_tokens":869,"completion_tokens":646,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":569}},"tokens_in":485,"tokens_out":646,"duration_ms":6832,"temperature":1.0,"reasoning_tokens":569,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:49:10.686166+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Calculate the mean absolute error of the framework after removing all shape-based radiomic features; if the 8.0-day accuracy substantially degrades, the result is largely explained by the model reading the contour that defines the label. A second falsifier is to evaluate against first-trimester or known-conception gestational age: if the 8.0-day error does not survive an independent dating standard, the claimed gestational age accuracy is an artifact of the formula-derived target.","supporting_citations":[{"cited_title":"Ultrasound in Obstetrics & Gynecology48(6), 719–726 (2016)","cited_arxiv_id":null,"evidence_quote":"Supplies the INTERGROWTH-21st formula that converts head circumference to gestational age, defining the regression target."},{"cited_title":"PLOS ONE 13(8), 1–20 (Aug 2018)","cited_arxiv_id":null,"evidence_quote":"Provides the HC18 dataset with head circumference values and images, one of the two evaluation datasets."},{"cited_title":"Scientific Reports10(1), 10200 (Dec 2020)","cited_arxiv_id":null,"evidence_quote":"Provides the ES-TT multi-center Spanish dataset with fetal head ROI annotations, the other evaluation dataset."},{"cited_title":"In: proceedings of Conference on Computer Vision and Pattern Recognition (CVPR) (2022)","cited_arxiv_id":null,"evidence_quote":"The ConvNeXt architecture is the deep backbone whose features are fused with radiomics and achieves the best reported result."},{"cited_title":"Cancer Research77(21), e104–e107 (Oct 2017)","cited_arxiv_id":null,"evidence_quote":"Provides the radiomics feature extraction used to compute the 95 hand-crafted features."},{"cited_title":"npj Digital Medicine6(1), 36 (Mar 2023)","cited_arxiv_id":null,"evidence_quote":"The image-based deep learning baseline whose ResNet18 result is the direct comparison for the reported improvement."}],"review_version":1}