{"id":"c68ec1ac-1506-4af9-a057-7dc9caf5e9fd","arxiv_id":"1908.08329","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A sequential extension of Mask R-CNN that labels and segments ribs in chest X-rays achieves a mean segmentation Dice of 0.733 and is claimed to be the first to do both tasks.","lead":"The authors adapt a standard computer vision network (Mask R-CNN) to find, label, and outline the ribs in chest X-ray images, using a step-by-step process where the network first sees the rib above the one it is detecting. A smart generalist might read this because automated rib labeling could help doctors assess X-ray quality, locate findings from written reports, and improve lung nodule detection by suppressing rib shadows.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Anchor boxes computed from all ground-truth boxes leak test-fold information into the five-fold cross-validation, so the reported Dice gains may be overstated.","rationale":"The reader identified annotation quality and rib visibility as the weakest assumption. That is a legitimate concern, but it is not directly testable from the paper and may be accurate. The anchor-box leakage is more clearly load-bearing because it is explicitly described and affects all reported numerics. The reader's rationale did mention that 'anchor boxes are fitted to ground truth boxes of the entire dataset' as a soundness issue, so there is partial agreement. However, the reader's headline weakest_assumption was about annotations, not leakage. The proposed check would settle whether the leakage materially changes the results. If it does, the central quantitative claim weakens; if not, the conditional verdict remains appropriate. I therefore recommend keeping the reader's CONDITIONAL verdict rather than moving to ACCEPT or REJECT, because the remaining reproducibility and comparison issues are still unresolved.","tokens_in":2863,"tokens_out":4813,"duration_ms":51013,"concrete_test":"Rerun the five-fold cross-validation with anchor boxes estimated separately for each fold using only that fold's training ground-truth boxes, and recompute Table 1. Compare the per-fold detection and segmentation Dice to the original values. If the mean Dice drops by more than about 1-2 points, or if the reported 5%/23% improvement over the unmodified Mask R-CNN baseline shrinks materially, the original numbers are partly leakage-induced. Reporting per-fold anchor statistics and per-fold Dice would make the magnitude of the bias assessable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 states that the 30 anchor boxes were computed by Mean Shift clustering from all ground-truth bounding boxes of the entire dataset. Section 3 then reports a five-fold cross-validation. Under this design, the ground truth of the fold being evaluated enters the definition of the anchor boxes used by that fold's model. Because the modified RPN 'determines shifts on the basis of anchor boxes' without non-maximum suppression, the anchors are not a neutral hyperparameter; they encode the test fold's rib locations and sizes. The reported detection Dice (0.846) and segmentation Dice (0.733), and the claimed 5%/23% improvements over the original implementation, therefore reflect a favorable bias of unknown size. The central claim that the method is 'comparable or even better than existing methods' depends on these numbers; if the bias is material, the claim is unsupported. This is the most load-bearing concern because it is a concrete methodological problem in the only quantitative evidence, independent of whether the hand annotations are accurate.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an adaptation of Mask R-CNN for simultaneous detection, anatomical labeling, and segmentation of ribs in PA chest X-ray images. The method introduces three extensions: rib-specific networks, dedicated anchor boxes computed by Mean Shift clustering from ground-truth bounding boxes, and a sequential processing scheme in which the segmentation of the upper neighbor rib is provided as an additional input channel for ribs 2-9. The authors report a five-fold cross-validation on 174 images, achieving mean Dice values of 0.846 for bounding-box detection and 0.733 for segmentation, and they claim improvements of more than 5% and 23% over the original Mask R-CNN implementation, as well as better performance than an atlas-based rib detection method.","tokens_in":2990,"tokens_out":2581,"duration_ms":25610,"significance":"If the reported results hold, this would be the first method to provide both rib segmentation and anatomical labeling in chest radiographs, which has practical value for rib counting, report generation, and follow-up correspondence. The paper builds on a publicly available Mask R-CNN implementation, uses a clearly described dataset, and reports quantitative results from cross-validation. The strengths are the explicit architectural extensions (rib-specific anchors and sequential neighbor-mask input) and the direct comparison to a baseline implementation. However, the central empirical claims depend on the soundness of the evaluation protocol, and the current manuscript leaves two methodological points unresolved: anchor-box information leakage across cross-validation folds, and the training/inference protocol for the sequential network input. These issues must be addressed before the reported gains can be considered reliable.","major_comments":[{"comment":"The anchor boxes are computed by Mean Shift clustering from all ground-truth bounding boxes of the entire dataset, as stated in Section 2. In the five-fold cross-validation of Section 3, the ground-truth boxes of the test fold therefore contribute to the anchor geometry used by the model evaluated on that fold. Because the modified RPN determines shifts directly from these anchors without non-maximum suppression, the anchors are not a neutral training hyperparameter; they encode test-fold rib locations and sizes. This introduces a favorable bias of unknown magnitude into the reported Dice values (0.846 detection, 0.733 segmentation) and into the claimed >5% and >23% improvements over the original implementation. The authors should recompute the anchor boxes using only each training fold's ground truth, or otherwise demonstrate that the evaluation is insensitive to this leakage.","section":"Section 2 (Data and Method) and Section 3 (Results)"},{"comment":"The text states that for ribs 2 to 9, 'the output of the segmentation of the above rib is used for the third channel,' but it does not specify whether this output is the predicted mask from the network or the ground-truth mask during training. If ground-truth masks are used in training and predicted masks in testing, the network sees perfect upper-rib segmentations during training, creating a train-test mismatch that can inflate the reported segmentation accuracy. If predicted masks are used, the propagation order and the handling of errors in upstream ribs need to be described. This ambiguity is load-bearing because the sequential input is a central novelty of the method.","section":"Section 2, paragraph on sequential processing"},{"comment":"The comparison with the atlas-based method of Candemir et al. reports accuracy, sensitivity, and specificity values, but the text does not define the units, the reference standard, or the spatial extent over which these metrics are computed (e.g., pixel-level, rib-level, or image-level). Without a precise definition, and without confidence intervals or a statistical significance test, the statement that the proposed algorithm 'achieves better results in all categories' is not supported. The authors should specify the evaluation protocol used for this comparison and report variability across the cross-validation folds.","section":"Section 3, comparison with Candemir et al."}],"minor_comments":[{"comment":"The abstract describes the images as 'anterior-posterior' while Section 2 states 'posterior-anterior'; the terminology should be made consistent.","section":"Abstract and Section 2"},{"comment":"The table heading 'Detection' should clarify that the reported value is the Dice coefficient for bounding boxes, and the spacing in '0 .841' should be corrected.","section":"Section 3, Table 1"},{"comment":"The comparison to the 'original implementation' of Mask R-CNN would be more informative if the anchor configuration and training details of that baseline were described, since the claimed improvement depends on the baseline setup.","section":"Section 3, baseline comparison"},{"comment":"The claim of 'very fast run-times that enable realtime analysis' is not supported by any timing measurement in the paper; a statement of inference time per image would be needed.","section":"Section 4, Conclusion"},{"comment":"The reference for He et al. contains a formatting error: 'InComputer Vision (ICCV)' should be 'In Computer Vision (ICCV)'.","section":"References"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is that they do rib segmentation and anatomical labeling in one pass, which nobody else seems to have done for chest X-rays. The per-rib sequential networks using the upper rib's mask as an extra input channel is a sensible idea, and they credit Lessmann et al. for the sequential concept. That part is worth paying attention to.\n\nThe evaluation is the problem. The paper computes 30 anchor boxes by Mean Shift clustering from all ground truth bounding boxes in the entire 174-image dataset, then runs five-fold cross-validation. The test fold's rib locations and sizes therefore enter the proposal geometry of the trained model. That is exactly the kind of circularity reviewers are supposed to catch. It does not invalidate the method, but it does mean the reported Dice values (0.846 detection, 0.733 segmentation) and the claimed 5%/23% gains over the original Mask R-CNN are likely optimistic, perhaps substantially so. The comparison to Candemir et al. also uses accuracy/sensitivity/specificity on what may be a different dataset, and there is no comparison to Mader et al., which is the relevant labeling baseline. Those are soft spots that a full paper could address with a properly split anchor calculation and head-to-head evaluation.\n\nThe paper is an extended abstract, so missing details like training hyperparameters and exact fold splits are forgivable, but the anchor leak is not a detail. That said, the central architectural contribution is coherent and the first-to-combine claim holds up. The hand-contoured annotations from von Berg et al. are a reasonable ground truth for visible rib shadows. The work is useful for bone suppression, quality assessment, and report localization, even if it is a niche task.\n\nMy take: the method deserves serious referee attention, but not on the strength of these numbers. A revised version should recompute anchors per training fold only and report per-fold variability. If the improvements shrink, the paper still has a publishable core—just a weaker quantitative claim. I would not cite the current numbers in my own work, but I would point to the method idea.\n\nRecommendation: send it to peer review, with the evaluation protocol as the primary required revision.","headline":"First genuine combination of rib segmentation and labeling in chest X-ray, but the reported numbers are compromised by anchor boxes fitted to the whole dataset before cross-validation.","tokens_in":3560,"tokens_out":1762,"would_cite":false,"duration_ms":21245,"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":"The paper reports the first approach to simultaneously segment and anatomically label ribs in chest X-rays, with mean Dice of 0.846 for bounding boxes and 0.733 for masks.","keywords":["rib segmentation","rib labeling","chest X-ray","Mask R-CNN","instance segmentation","anatomical labeling","sequential processing","anchor boxes"],"falsifier":"On a hold-out set of posterior-anterior chest X-rays with rib masks produced from CT bone segmentations, retrain and evaluate per-label Dice; if the sequential channel ceases to improve labeling accuracy over the non-sequential baseline, or if ribs 7–9 labels are frequently confused, the central claim is not supported.","tokens_in":2633,"feed_emoji":"🩻","tokens_out":8104,"duration_ms":73726,"temperature":0.7,"pith_summary":"The paper sets out to show that one neural network can do what previously required separate tools: cut out individual rib shadows from a chest X-ray and simultaneously say which anatomical rib each shadow is. The authors modify an instance-segmentation network so that each rib gets its own branch and the branch for a rib also receives the segmentation of the rib above it as a hint; anchor boxes are computed from the data by clustering actual rib locations. In five-fold cross-validation on 174 images, the method achieves mean overlap scores of 0.846 for rib-box detection and 0.733 for rib-mask segmentation, improving on the unmodified network by more than 5% and 23%, respectively. If this holds, radiologists and automated pipelines could obtain labeled rib anatomy in a single fast pass, relevant for rib counting, quality assessment, and localizing findings.","feed_headline":"A single Mask R-CNN labels and segments ribs in one pass","feed_subtitle":"Rib-specific anchors and sequential upper-rib context hit Dice 0.733, beating the base model by 23 percent.","key_machinery":"The load-bearing idea is sequential, rib-specific instance segmentation. For each label 1 through 9, a Mask R-CNN branch is trained with its own set of anchor boxes; these boxes come from Mean Shift clustering of all ground-truth rib boxes normalized to image size, producing 30 prototypical locations. For ribs 2 to 9, the predicted mask of the rib above is inserted as the third input channel, with the grayscale X-ray replicated in the other two channels, so the network can use the upper neighbor's shape and position to disambiguate the current rib. That sequential context is what gives the labels anatomical consistency rather than just separating arbitrary instances.","core_discovery":"The central claim is that a Mask R-CNN, split into rib-specific networks and fed the segmentation of the upper neighboring rib as contextual input, is the first method to both segment ribs and assign anatomical labels in chest radiographs. The reported results—mean Dice 0.846 for bounding-box detection and 0.733 for segmentation on a five-fold cross-validation of 174 posterior-anterior chest X-rays—are said to be comparable to or better than published methods that only segment, and an improvement of over 5% and 23% over the base Mask R-CNN implementation. The authors also report higher accuracy, sensitivity, and specificity than an atlas-based rib-detection baseline. The claim matters because rib labeling and rib segmentation have previously been separate problems, and solving them together removes a preprocessing step and enables automatic rib counting for inhalation assessment.","pith_inferences":["The upper-neighbor input channel is a transferable design: any elongated, serially ordered anatomy—vertebrae, teeth, bowel segments—could be labeled by feeding the previous segment's mask back into the next detection head.","Because the reference masks cover only visible rib shadows, the reported Dice measures fidelity to visible anatomy rather than bone recovery; a CT-derived ground truth would likely be a sterner test and might reveal systematic under-segmentation.","The normalization of anchor boxes by image size implies the same architecture should carry over to X-rays of different resolutions without retuning the proposal geometry, as long as the anatomy is similarly positioned in the frame."],"forward_implications":["A single forward pass through this network yields both rib masks and rib labels, so chest X-ray pipelines could replace separate bone-suppression and rib-labeling stages.","The large jump from the base architecture suggests that domain-specific anchors and sequential context are the main levers for instance segmentation of repetitive anatomy, not just more training data.","The reported speed advantage over atlas-based registration opens the possibility of real-time rib counting for assessing inspiratory effort in bedside or screening settings."],"supporting_citations":[{"why":"Defines the Mask R-CNN architecture that the paper modifies for rib instance segmentation.","marker":"He et al., 2017"},{"why":"Supplies the 174 posterior-anterior chest X-rays and the hand-contoured rib masks used for training and evaluation.","marker":"von Berg et al., 2016"},{"why":"Provides the sequential neighbor-context processing idea the authors adapt, originally developed for vertebra segmentation.","marker":"Lessmann et al., 2019"},{"why":"Provides the Mean Shift clustering algorithm used to compute the 30 dedicated rib anchor boxes.","marker":"Comaniciu and Meer, 2002"},{"why":"Supplies the ResNet50 backbone used in the feature extractor.","marker":"He et al., 2016"},{"why":"Supplies the Feature Pyramid Network used for multi-scale feature extraction.","marker":"Lin et al., 2017"},{"why":"Gives the atlas-based rib detection baseline that the paper claims to outperform on accuracy, sensitivity, and specificity.","marker":"Candemir et al., 2016"},{"why":"Motivates rib labeling by linking rib counting to chest X-ray quality assessment.","marker":"Mader et al., 2018"}],"fun_headline_variants":["First to label and segment ribs in one Mask R-CNN","Rib-specific anchors and neighbor context beat base model by 23%","Sequential rib labeling and segmentation improves Dice to 0.733","Single network does rib labeling and segmentation, beating atlas baseline","Rib labeling and segmentation unified with Mask R-CNN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy of the hand-contoured rib masks in the 174 training images, and the consistent visibility of ribs 1–9, are assumed; errors there propagate into every anchor, segmentation target, and reported Dice value.","fun_headline_variants_meta":{"raw":{"variants":["First to label and segment ribs in one Mask R-CNN","Rib-specific anchors and neighbor context beat base model by 23%","Sequential rib labeling and segmentation improves Dice to 0.733","Single network does rib labeling and segmentation, beating atlas baseline","Rib labeling and segmentation unified with Mask R-CNN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000256,"raw_usage":{"total_tokens":1527,"prompt_tokens":847,"completion_tokens":680,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":593}},"tokens_in":463,"tokens_out":680,"duration_ms":6793,"temperature":1.0,"reasoning_tokens":593,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:41:40.222400+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a hold-out set of posterior-anterior chest X-rays with rib masks produced from CT bone segmentations, retrain and evaluate per-label Dice; if the sequential channel ceases to improve labeling accuracy over the non-sequential baseline, or if ribs 7–9 labels are frequently confused, the central claim is not supported.","supporting_citations":[{"cited_title":"A novel bone suppression method that improves lung nodule detection","cited_arxiv_id":null,"evidence_quote":"Supplies the 174 posterior-anterior chest X-rays and the hand-contoured rib masks used for training and evaluation."},{"cited_title":"Iterative fully convolutional neural networks for automatic vertebra segmentation and identification","cited_arxiv_id":null,"evidence_quote":"Provides the sequential neighbor-context processing idea the authors adapt, originally developed for vertebra segmentation."},{"cited_title":"Feature pyramid networks for object detection","cited_arxiv_id":null,"evidence_quote":"Supplies the Feature Pyramid Network used for multi-scale feature extraction."},{"cited_title":"Atlas-based rib-bone detection in chest X -rays","cited_arxiv_id":null,"evidence_quote":"Gives the atlas-based rib detection baseline that the paper claims to outperform on accuracy, sensitivity, and specificity."},{"cited_title":"Localization and labeling of posterior ribs in chest radiographs using a CRF -regularized FCN with local refinement","cited_arxiv_id":null,"evidence_quote":"Motivates rib labeling by linking rib counting to chest X-ray quality assessment."}],"review_version":1}