{"id":"7212f1fe-c3d3-42ea-abf8-a46a29595d90","arxiv_id":"2505.05004","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An automated CT pipeline measures rib length and rib-start morphology, and classifies short thoracolumbar stump ribs with F1 0.84 using only the first 22 mm of each rib.","lead":"This paper builds an automated CT pipeline that segments ribs, measures each rib's length, and flags short 'stump' ribs, then shows that the shape of the first 22 millimeters of a rib can identify stump ribs even when the rest of the rib is outside the scan. A generalist reader would care because this could automate detection of spine numbering anomalies that affect surgical planning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The morphological classifier's F1=0.84 is built on labels from the authors' own 38 mm length threshold, with no independent anatomical gold standard; Appendix A.2 suggests the features may just track rib length, so the central claim needs external validation.","rationale":"I read the paper's strongest claim as the morphological classifier result, since this is what enables stump-rib detection in partial FOV. The load-bearing condition is that the stump-rib labels are valid and that the F1 is not an artifact of choosing a configuration on the test data. Neither is established. The paper is honest about the missing gold standard, and the public release of code and weights is a plus, but the threshold experiment in Appendix A.2 undercuts the anatomical status of the 38 mm label. This is exactly the reader's weak point; I add the Appendix A.2 detail. The verdict remains conditional: acceptance should require independent-label validation and a pre-specified, unbiased evaluation protocol.","tokens_in":12541,"tokens_out":4765,"duration_ms":49719,"concrete_test":"Pre-register the frozen linear SVM with 4-PPR+DRC features and its decision threshold. On 50-100 new CT subjects from an independent site, have two radiologists, blinded to any automatic measurements, independently classify the lowest two ribs as stump or normal using full 3D CT, and record inter-rater agreement. Then apply the frozen SVM to these subjects and compute F1 against each radiologist's labels and against the 38 mm RLMA rule. If F1 falls below 0.75, or if radiologists' agreement with the 38 mm rule is poor, the morphological classification claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that morphological features classify thoracolumbar stump ribs requires that the 'stump rib' labels are valid anatomical categories. In Section 2.4, a rib is labeled SR if the RLMA-computed length is <=38 mm; the Likert review in Section 2.6 only verifies that segmentation errors did not change that computed length, not that the length corresponds to true stump-rib anatomy. The SVM features (DRC, n-PPR) are derived from the same rib start point and path-point machinery that produces the length, so the classifier is effectively trained to predict the length threshold from geometric proxies. The paper's own Appendix A.2 is a red flag: classifier F1 changes smoothly with the threshold and has no peak at 38 mm, the pattern expected if the features separate 'shorter ribs' from 'longer ribs' rather than a discrete anatomical entity. In addition, Table 5 reports the best configuration chosen across kernels, feature sets, and n values on the same data used to estimate F1; with multiple seeds and splits, this selection can inflate the reported 0.84. The limitations section acknowledges there is no gold standard for rib length, but the current design does not let us distinguish a genuine morphological signal from an artifact of label construction and model selection.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a complete pipeline for automated thoracolumbar stump rib detection in CT. The authors train a high-resolution nnU-Net for whole-rib segmentation using a mix of private in-house and public data, introduce an iterative rib-length measurement algorithm (RLMA), extract morphological features (DRC, n-PPR, volume-to-length ratio), and train SVMs to classify stump ribs from regular ribs using only the first roughly 22 mm of rib. On 45 manually corrected test subjects the segmentation model achieves Dice 0.997 versus 0.751--0.779 for public baselines, the RLMA is reported to achieve a 98.2% success rate on expert review, and a linear SVM with 4-PPR and DRC reaches an average F1 of 0.84. Code, model weights, and segmentation masks are publicly released.","tokens_in":12798,"tokens_out":7013,"duration_ms":72999,"significance":"If the results hold, the segmentation model, length-measurement algorithm, and released masks are valuable public resources for rib analysis in CT, and the partial-field-of-view morphological classifier would be clinically useful for surgical planning and vertebral enumeration. The segmentation comparisons on manually corrected test data are convincing and reproducible, and the authors are transparent about releasing code and data. However, the central classification claim currently depends on a self-referential length threshold rather than an independent anatomical gold standard, and on selecting the best SVM configuration from many alternatives on the same test data. The claimed morphological signal 'beyond the length' is not yet established, so the overall significance is conditional on additional validation or a more modest reframing.","major_comments":[{"comment":"The SR/non-SR labels used to train and evaluate the SVM are obtained by comparing the RLMA-computed rib length with the 38 mm threshold from Wigh [3], and the Likert review in Section 2.6 only checks that segmentation errors did not alter the computed length. There is no independent anatomical rib-length standard. Therefore the reported F1=0.84 measures agreement between morphological features and the RLMA length threshold, not agreement with a clinical ground truth. The paper should obtain an independent length or anatomy reference on at least a subset, or explicitly reframe the classifier as a length-proxy model for the 38 mm definition.","section":"Section 2.4, Section 2.6, Table 5"},{"comment":"The threshold-shift experiment shows no peak in SVM F1 at the 38 mm threshold and a smooth decay with increasing threshold, which the authors themselves interpret as evidence that either length-based labeling is suboptimal or that the features smoothly track rib length. This directly contradicts the abstract's claim that the morphological features are 'beyond the length' and classify stump ribs as a discrete entity. This load-bearing evidence requires either independent anatomical validation of a discrete stump-rib class or a revised conclusion presenting the method as a length-proxy classifier for partial-field-of-view scans.","section":"Appendix A.2, Figure A.9"},{"comment":"The best configuration (4-PPR+DRC, linear kernel) is selected from a grid of 14 feature/kernel combinations and evaluated on the same 70/30 subject-wise splits across 10 seeds. Selecting the maximum F1 over this grid biases the reported 0.84 upward, and no nested resampling or multiple-comparison correction is described. The authors should report selection-corrected estimates, use a separate validation split for model selection, or pre-specify the configuration.","section":"Section 2.6, Table 5"},{"comment":"The 98.2% 'correct measure' rate is obtained by summing ratings (5), (4), and (3), but rating (3) explicitly indicates that the rib is cut off at the image border and the measured length is that of the visible section only. For such ribs, the full rib length is unknown unless the visible length already exceeds 38 mm, so counting all rating-(3) cases as correct full-length measurements inflates the success rate. The authors should separate truly complete measurements from truncated ones and report the rate of valid stump-rib labeling accordingly.","section":"Table 4, Section 3.2"},{"comment":"The Wilcoxon rank-sum tests compare 2464 ribs from 648 subjects as if every rib were an independent observation, despite ribs being clustered within subjects. Ignoring within-subject correlation can inflate significance. The reported p-values are very small and the group differences are large, so the qualitative conclusions would probably survive a corrected analysis, but the authors should repeat the feature comparisons with subject-level aggregation or a mixed-effects model to confirm.","section":"Section 2.8, Section 3.3"}],"minor_comments":[{"comment":"The heading 'Statisticial Analysis' contains a typo and should read 'Statistical Analysis'.","section":"Section 2.8"},{"comment":"The definition of ASSD divides by |X|+|Y|, but X and Y are the full rib volumes while A and B are the boundary surfaces; the standard average symmetric surface distance should divide by |A|+|B|, the total number of boundary points. Please correct the formula or clarify the notation.","section":"Equation 2"},{"comment":"Table 3 reports evaluation on 20 manually corrected test subjects from RibFrac, while Section 2.2 describes 45 randomly chosen test subjects from VerSe and RibFrac combined. Please clarify how the 20-subject RibFrac subset relates to the 45-subject test set.","section":"Table 3"},{"comment":"The scatter plot in Figure 6 shows substantial overlap between the two classes, while the text describes a 'clear trend'; consider adding marginal distributions or a decision boundary to make the group separation more visible.","section":"Figure 6"},{"comment":"The name 'TotalSegmentor' is misspelled in the text and should be 'TotalSegmentator'.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about the lack of a rib-length gold standard and the threshold-shift analysis in Appendix A.2 is commendable, but the self-referential label construction is a central weakness. The segmentation contribution alone may be incremental for a specialized medical imaging venue, so the classification claim needs either external validation or a careful reframing as a length-proxy method. The use of the authors' own tools (panoptica, Btrfly) in the evaluation pipeline is not problematic, but the model-selection procedure should be made more rigorous before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should read this paper for the rib segmentation and length-measurement pipeline, which is genuinely useful and well evaluated. The authors train an nnU-Net that beats TotalSegmentator and RibSeg by a large margin (Dice 0.997 vs 0.75–0.78), release code, weights, and masks, and show a 98.2% success rate on automatic rib length measurement against internal expert review. They also automate stump rib detection and quantitatively describe morphological differences between stump and regular ribs. That is new as far as I can tell, and the comparison against public baselines is honest and convincing.\n\nThe soft spots are real but localized. The main issue is that the morphological classifier's ground-truth labels are derived from the authors' own length measurement thresholded at 38 mm. The segmentation is strong, so the length measurement is probably mostly correct, but there is no independent anatomical gold standard. The SVM features are computed from the same rib path machinery that produces the length, so the F1 of 0.84 may simply be predicting 'short rib' from the first 22 mm rather than detecting a discrete morphological entity. The paper's own Appendix A.2 is the red flag: the F1 varies smoothly with the length threshold and has no peak at 38 mm. The authors note this and offer two interpretations, but the more parsimonious one is that the morphological features track length continuously. That weakens the claim of 'morphological features beyond the length.' The model-selection issue is secondary: the best SVM configuration is chosen from many kernels, feature sets, and n values on the same test splits, so the reported 0.84 is probably optimistic, though the paper does average over 10 seeds and report standard deviations.\n\nThese do not sink the paper. The segmentation and measurement tools are a contribution on their own, and the partial-FOV classification is clinically plausible even if it is a proxy for length. The limitations section is refreshingly open about the lack of a gold standard. For a spine-imaging audience, this is a practical step forward. I would send it to peer review with a request for external validation of the length labels (e.g., manual rib-length measurement on a subset) and a nested or pre-registered model-selection procedure. The interpretational claim should be softened unless that validation changes the picture.\n\nVerdict: deserving of a serious referee, despite the load-bearing caveat.","headline":"Solid segmentation and measurement pipeline with a new stump-rib classification task, but the morphological classifier's labels come from the length threshold it claims to go beyond, and Appendix A.2 undercuts the 'beyond length' interpretation.","tokens_in":13364,"tokens_out":1433,"would_cite":true,"duration_ms":16790,"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":"Thoracolumbar stump ribs can be detected, measured, and classified automatically from CT scans, using the first roughly 22 mm of rib direction and position to reach F1 0.84 even on partially visible ribs.","keywords":["thoracolumbar stump rib","rib segmentation","computed tomography","deep learning","rib length measurement","morphological features","transitional vertebrae","support vector machine"],"falsifier":"Take 100 CT scans with fully visible lowest ribs, have two independent radiologists measure the same ribs with a standardized curved reformatting protocol, and compare those measurements to the algorithm's lengths; if the automated lengths disagree systematically, especially within 10 mm of the 38 mm threshold, the ground-truth labels and the F1 0.84 result are called into question.","tokens_in":12346,"feed_emoji":"🦴","tokens_out":8643,"duration_ms":76715,"temperature":0.7,"pith_summary":"This paper aims to make the detection of thoracolumbar stump ribs—the short, malformed ribs at the bottom of the thoracic spine that mark transitional vertebrae or numbering anomalies—fully automatic in CT scans. It trains a deep-learning segmentation model that captures the whole rib, including the part near the spine that public tools miss, then measures each rib with an iterative path-following algorithm, reporting that 98.2% of lowest ribs are measured correctly. The authors show that stump ribs are not just shorter; their first few centimeters point more downward, sideways, and posteriorly, and the rib is proportionally thinner. A simple support-vector machine using these early-direction and start-position features separates stump ribs from regular ribs with an average F1 of 0.84 using only the first roughly 22 mm of a rib. If this is right, radiologists could flag transitional vertebrae and spine-numbering anomalies in ordinary or narrow-field CT scans without manual rib tracing.","feed_headline":"Stump ribs can be detected in CT from just 22 mm of rib","feed_subtitle":"A whole-rib segmentation and length pipeline reaches F1 0.84 on partially visible stump ribs, aiding spine numbering.","key_machinery":"The load-bearing mechanism is a three-stage pipeline: a high-resolution deep-learning segmentation model trained at 0.8 mm isotropic resolution on 55 manually corrected CT volumes, which recovers the rib head near the vertebra that public segmentation tools omit; the rib-length measurement algorithm (RLMA), an iterative path walk that averages candidate points within a 14.5 to 15.5 mm circular shell of the current path point, projects each step back onto the segmentation surface, and sums the path; and morphological features derived from that path—the posterior distance between the rib start and the vertebra corpus center (PDRC) and the first n direction vectors between consecutive path points (n-PPR)—which are fed to a support-vector machine. The conceptual move is using only the first four path points, about 22 mm, plus the start-vertebra relation to classify a stump rib without needing the full rib length.","core_discovery":"The paper claims that thoracolumbar stump ribs—ribs of 38 mm or less at the lowest thoracic level—are automatically detectable from CT, and that their distinguishing signal is not only length. Starting from the vertebra corpus, the first roughly 22 mm of a stump rib is angled more downward, sideways, and posteriorly than a full rib, and the rib has a lower volume-to-length ratio. The paper's segmentation model produces dense whole-rib masks (Dice 0.997 on its test set versus 0.751 to 0.779 for public baselines), an iterative path-point algorithm measures rib length with 98.2% success, and a linear support-vector machine on the first four path-point directions plus the rib-start/vertebra-center distance separates stump ribs from regular ribs with an average F1 of 0.84 using only the initial 22 mm of the rib. This implies stump ribs can be flagged even when the scan's field of view cuts the rib short.","pith_inferences":["Beyond the paper: because the authors' own threshold-shift experiment shows no peak at 38 mm but a small peak near 41 mm, the fixed length cutoff may not carve anatomy at a natural boundary; a population study relating rib length to vertebral level could test whether the cutoff should be revised.","Beyond the paper: the same 22 mm directional signature might help identify transitional vertebrae in MRI or low-dose scans, where whole-rib segmentation is unreliable—this is untested here.","Beyond the paper: since the authors note that some non-stump ribs with large posterior start distances had enlarged foveae or corpus anomalies, adding vertebra morphometry to the feature set could further separate the groups, a hypothesis they leave for future work.","Beyond the paper: a pre-registered test on whole-body CT with ribs truncated at varying levels could establish how robust the F1 0.84 is when the visible rib length varies continuously, and whether morphological classification can be embedded in automated spine-labeling pipelines."],"forward_implications":["Stump ribs can be discovered automatically in routine CT, reducing manual measurement and helping with vertebral numbering and detection of transitional anomalies.","Because classification needs only the first 22 mm of a rib, limited field-of-view CT scans—where whole ribs are not visible—become usable for stump-rib assessment.","The measured morphological features offer quantitative descriptors of stump-rib shape that could supplement the length-only definition in clinical reporting.","The released model weights and masks let other groups produce whole-rib segmentations and length measurements on their own CT data without retraining.","Accurate whole-rib segmentation with high Dice and low surface distance supports downstream biomechanical or surgical-planning analyses of the thoracolumbar junction."],"supporting_citations":[{"why":"Supplies the 38 mm length threshold that defines a stump rib and anchors the ground-truth labels.","marker":"[3]"},{"why":"Provides the segmentation architecture and training recipe used to build the whole-rib model.","marker":"[9]"},{"why":"Publicly available rib segmentation baseline that the paper's model outperforms on Dice and surface distance.","marker":"[18]"},{"why":"Supplies the VerSe CT cohort and expert vertebra segmentations used for instance assignment and analysis.","marker":"[19]"},{"why":"Provides the RibFrac CT cohort with rib fractures, used as public training and test data.","marker":"[20]"},{"why":"Provides the RibFrac challenge context and associated rib-fracture instance segmentation data used for evaluation.","marker":"[21]"},{"why":"Provides the RibSeg rib annotations used as an additional segmentation baseline on RibFrac.","marker":"[22]"},{"why":"Supplies vertebra instance segmentation for RibFrac, enabling per-rib assignment via spatial proximity.","marker":"[24]"},{"why":"The support-vector machine used to test whether morphological features separate stump ribs from regular ribs.","marker":"[25]"}],"fun_headline_variants":["AI finds stubby ribs in CT from merely 22 mm","Stump rib detection from 22 mm of rib in CT scans","CT rib stubs spotted via AI using just 22 mm","Automated stump rib detection: 22 mm is enough","Rib stubs in CT: AI needs only 22 mm to flag them"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ground-truth labels for stump ribs come from the paper's own automated length measurement crossing a fixed 38 mm cutoff, with no independent or clinically validated rib-length standard; if those length estimates are systematically biased by segmentation errors, the reported accuracy may partly reflect the error pattern rather than true anatomy.","fun_headline_variants_meta":{"raw":{"variants":["AI finds stubby ribs in CT from merely 22 mm","Stump rib detection from 22 mm of rib in CT scans","CT rib stubs spotted via AI using just 22 mm","Automated stump rib detection: 22 mm is enough","Rib stubs in CT: AI needs only 22 mm to flag them"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000281,"raw_usage":{"total_tokens":1688,"prompt_tokens":996,"completion_tokens":692,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":602}},"tokens_in":612,"tokens_out":692,"duration_ms":6742,"temperature":1.0,"reasoning_tokens":602,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:15:52.600989+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take 100 CT scans with fully visible lowest ribs, have two independent radiologists measure the same ribs with a standardized curved reformatting protocol, and compare those measurements to the algorithm's lengths; if the automated lengths disagree systematically, especially within 10 mm of the 38 mm threshold, the ground-truth labels and the F1 0.84 result are called into question.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 38 mm length threshold that defines a stump rib and anchors the ground-truth labels."},{"cited_title":"Wasserthal, H.-C","cited_arxiv_id":null,"evidence_quote":"Publicly available rib segmentation baseline that the paper's model outperforms on Dice and surface distance."},{"cited_title":"Sekuboyina, M","cited_arxiv_id":null,"evidence_quote":"Supplies the VerSe CT cohort and expert vertebra segmentations used for instance assignment and analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the RibFrac CT cohort with rib fractures, used as public training and test data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the RibSeg rib annotations used as an additional segmentation baseline on RibFrac."},{"cited_title":"Sekuboyina, M","cited_arxiv_id":null,"evidence_quote":"Supplies vertebra instance segmentation for RibFrac, enabling per-rib assignment via spatial proximity."},{"cited_title":"Cortes, V","cited_arxiv_id":null,"evidence_quote":"The support-vector machine used to test whether morphological features separate stump ribs from regular ribs."}],"review_version":1}