{"id":"b4dbd82b-44ae-4795-a7db-7ec9c322042e","arxiv_id":"2506.11852","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A comparison of a custom region-growing method versus TotalSegmentator on MRI and CT concludes the custom method is better on MRI, but the accuracy metric used is cross-modality surface consistency rather than ground truth.","lead":"Two skin segmentation approaches, a graphics-based region growing algorithm from the authors' laboratory and the deep learning tool TotalSegmentator, are compared on MRI and CT volumes. The paper reports that the graphics method handles MRI better while TotalSegmentator is more automated and removes patient beds in CT, but the quantitative evidence is weakened by an evaluation metric that does not measure segmentation accuracy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quantitative evidence for 'graphics-based performs better on MRI' compares CT and MR skin surfaces without registration or ground truth (Sect. 3.2, 'Quantitative analysis'); surface distances may reflect misalignment, not segmentation error, so the central claim is not established.","rationale":"This is the load-bearing point because the abstract's specific claim is about which method is better on MRI. The quantitative section is the only evidence offered beyond a few qualitative figures. The flaw is not a minor statistical issue; it invalidates the comparison as a measure of accuracy. The paper's other observations (e.g., TotalSegmentator removing the bed) may be true, but the main comparative conclusion is unsupported. I agree with the reader's weakest assumption; the missing registration and ground truth are the same core problem. Since the reader already rejects the paper on these grounds, my verdict is unchanged.","tokens_in":10465,"tokens_out":4934,"duration_ms":43703,"concrete_test":"Run the quantitative analysis on a subset of the Bauer dataset using the provided ground-truth transformations to register CT and MR of each subject. Have an expert manually delineate the skin surface for a subset of slices (or use an independent reference segmentation), then compute symmetric Hausdorff distance, mean surface distance, and Dice between each method's surface and the manual reference in each modality. Also report both directed distances. If the graphics-based method's MRI advantage does not persist after alignment and against ground truth, the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that the graphics-based method is better on MRI rests on the quantitative analysis in Sect. 3.2. There, the authors segment skin from CT and MR of the same subject with each method and treat the distance between the resulting surfaces as a measure of accuracy, writing that 'the best result corresponds to two coincident surfaces.' This assumes the CT and MR acquisitions are already aligned. No registration step is described, and the Bauer dataset—which is a registration benchmark with ground-truth transformations—is used only for the images. Any patient positioning difference, breathing-phase mismatch, or field-of-view difference between CT and MR will produce large surface distances even if both segmentations are perfect, while a method that aggressively stops at internal structures (e.g., the skull, as noted for head MRI) can produce artificially low distances. The Hausdorff distance is also defined one-sided (dX1(X2) = max_{x∈X1} min_{y∈X2} ||x−y||2), so it is not symmetric and does not measure what is stated. Consequently, Fig. 4 cannot support the conclusion that the graphics-based method is more accurate on MRI; it only shows that its CT and MR surfaces are closer to each other, which may be due to systematic bias or surface smoothness rather than segmentation correctness. A comparison against manual segmentations with standard metrics (Dice, symmetric surface distance) after registration is required.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares a graphics-based iterative region-growing skin segmentation method (Paccini et al., 2024) with TotalSegmentator, a deep-learning tool, across CT and MRI acquisitions. The authors report qualitative and quantitative comparisons using the Hausdorff distance and visual inspection, and conclude that TotalSegmentator excels at automation but struggles with MRI because of its CT-based training, while the graphics-based method performs better on MRI but introduces more noise. The quantitative analysis in Section 3.2 compares CT and MR skin segmentations of the same subject without registration and without a manual ground truth, treating surface coincidence as a proxy for accuracy.","tokens_in":10748,"tokens_out":4270,"duration_ms":39122,"significance":"If the comparison were valid, the paper would provide practically useful guidance for choosing between algorithmic and AI-based skin segmentation in multimodal imaging. The paper correctly identifies several real trade-offs, such as TotalSegmentator's automatic bed removal versus the graphics-based method's need for manual intervention, and the interpretability advantages of rule-based methods. However, the central quantitative claim is not established because the evaluation design does not measure segmentation accuracy in a sound way. The paper's positive aspects include a clear description of the graphics-based algorithm and a concrete dataset choice; yet the load-bearing experimental evidence is currently inadequate.","major_comments":[{"comment":"The accuracy metric is invalid for the stated purpose. The authors compare segmentations of CT and MR images of the same subject and state that 'the best result corresponds to two coincident surfaces,' which assumes that the two acquisitions are already aligned. No registration step is reported, and the Bauer dataset is a registration benchmark with ground-truth transformations that are not used. Patient positioning, breathing-phase differences, and field-of-view differences (e.g., arms present only in MR) can produce large CT-MR surface distances even for perfect segmentations. Therefore Fig. 4 and the conclusion that the graphics-based method 'achieves better results' on MRI are not supported.","section":"Section 3.2, Quantitative analysis"},{"comment":"The distance measure is defined one-sided as dX1(X2) = max_{x in X1} min_{y in X2} ||x-y||_2, which is a directed distance, not the symmetric Hausdorff distance used in medical image evaluation. The surrounding text also mismatches the formula: it says the minimum distance of each vertex of X2 from vertices of X1 is computed, but the formula takes the maximum over X1. Because the measure is not symmetric, the reported values depend on which surface is labeled X1 and can be misleading. Standard symmetric metrics such as Dice similarity coefficient or symmetric surface distance should be used.","section":"Section 3.2, Hausdorff distance definition"},{"comment":"The comparison is not like-for-like. The graphics-based method is tuned by the authors for this dataset: a skin isovalue of 0.1 and a gradient isovalue of 0.01 are described as yielding 'optimal results' in Section 3.2. TotalSegmentator, by contrast, is used off-the-shelf with no MRI-specific adaptation or fine-tuning. The conclusion that the graphics-based method 'performs better for MRIs' is therefore partly forced by the evaluation design, because the authors' method has been manually calibrated to the test data while the competing method has not.","section":"Section 3.2, parameter tuning"},{"comment":"The Hausdorff distances reported in Table 1 and Figures 1-3 are distances between the two methods' surfaces, not distances to any reference standard. A low distance can occur when both methods make the same error (e.g., the head MRI case where TotalSegmentator's output 'seems to be related more to the skull than the actual skin'), so these numbers cannot be interpreted as evidence that one method is more accurate. No manual ground truth or standard accuracy metrics appear anywhere in the paper, so the abstract's claim that one method is 'better' on MRI is not backed by a valid quantitative comparison.","section":"Table 1 and Figures 1-3"}],"minor_comments":[{"comment":"There is a typo in 'Totasegmentator' in the sentence about Figure 1; it should be 'TotalSegmentator'.","section":"Section 3.2"},{"comment":"The notation for the distance is inconsistent: the text and the equation label the surfaces differently, and the equation appears to use a squared norm (||x-y||_2) without clarifying whether the distance is in millimeters or squared millimeters. The units in Table 1 and Figures 2-4 are stated as mm, so the norm should be explicitly defined.","section":"Section 3.2"},{"comment":"Figure 2(a) is said to show an abdominal MRI, but the caption does not describe the color coding or the meaning of the distance distribution in detail; a reader must infer from the text which surface is red and which is green.","section":"Figure 2"},{"comment":"The conclusion that the graphics-based approach is 'versatile and adaptable' is stated strongly, but the only direct evidence is the flawed quantitative analysis and a few qualitative examples. A more measured conclusion would acknowledge the limitations of the comparison.","section":"Section 4"}],"recommendation":"reject","confidential_remarks":"The paper's central quantitative claim is undermined by the lack of registration and ground truth in the CT-MR comparison, and by the asymmetric tuning of the two methods. The issues are not merely presentational; they affect the main conclusion. If the authors redo the evaluation with registered data, standard metrics, and fair tuning, the paper could be reconsidered, but in its current form it does not meet the standard for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Chiara,\n\nQuick take: the survey half is a decent overview and the qualitative observations are plausible, but the only quantitative evidence for the headline claim doesn't measure what it claims. I'm with the skeptic here.\n\nWhat's actually new: the head-to-head comparison of the authors' region-growing method with TotalSegmentator on the Bauer dataset (52 subjects) is the only original quantitative content. The survey in Sect. 2 is a useful, if not exhaustive, map of skin segmentation methods — manual, thresholding, region-growing, active contours, atlas, and the current deep-learning tools. And the qualitative points — TotalSegmentator removes the bed, struggles on MRI because it was trained on CT, and on head MRI often captures the skull rather than skin — are consistent with what others have observed and are worth recording.\n\nThe soft spot is load-bearing. The quantitative analysis (Sect. 3.2) treats the distance between CT and MR skin surfaces of the same subject as a measure of accuracy: 'the best result corresponds to two coincident surfaces.' That is only valid if the two acquisitions are aligned. The paper describes no registration step. The Bauer dataset is a registration benchmark with ground-truth transformations, but those aren't used. So the distances in Fig. 4 mostly reflect misalignment, breathing-phase differences, and field-of-view mismatch. The authors even acknowledge that the arm appears in the MRIs but not in the CTs, which alone will inflate distances for both methods without saying anything about segmentation error. The Hausdorff distance is also defined one-sided, which isn't the standard symmetric metric.\n\nThe comparison is also not apples-to-apples: the graphics-based thresholds (isovalue 0.1, gradient 0.01) are tuned on this type of data, while TotalSegmentator runs off-the-shelf. So the MRI result is partly designed into the authors' method.\n\nWhat the paper does well: the computational cost discussion is sensible, and someone new to skin segmentation could use the survey as an entry point. But the central claim — 'graphics-based performs better on MRI' — is not established by the experiments as presented.\n\nMy recommendation: let a referee look at it, because the question is legitimate and the Bauer dataset makes a proper evaluation feasible. But the current version should not be accepted; the metric needs to be redesigned, not patched.\n\nBest,\n[Name]","headline":"The survey is fine and the qualitative observations are plausible, but the quantitative comparison lacks registration and uses cross-modality surface distance as an accuracy proxy, so the headline claim is not established.","tokens_in":11274,"tokens_out":6014,"would_cite":false,"duration_ms":52304,"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":"This paper reports that an iterative region-growing skin segmentation method outperforms a CT-trained deep learning model on MRI, while the AI model provides superior automation and bed removal on CT.","keywords":["3D skin segmentation","medical image segmentation","TotalSegmentator","region growing","MRI segmentation","CT segmentation","Hausdorff distance","multimodal imaging"],"falsifier":"Segment both modalities in a dataset with expert-drawn skin contours in co-registered CT and MRI, and compare each method's surface to those contours; if TotalSegmentator is as close or closer to the expert MRI skin than the graphics-based method, the central claim fails. Alternatively, recompute the 52-subject Hausdorff and mean distances after explicit rigid registration and check whether the graphics-based advantage persists.","tokens_in":10216,"feed_emoji":"🩻","tokens_out":7506,"duration_ms":68873,"temperature":0.7,"pith_summary":"Skin segmentation creates the outer body surface used for multimodal image registration, surgical planning, and patient-specific simulation. This paper compares two representative approaches: a graphics-based iterative region-growing algorithm and TotalSegmentator, a deep-learning model trained on CT. The authors claim that when MRI is involved, the graphics-based method is the more reliable choice because the AI model struggles with MR sequences it was not trained on, while the AI method wins on automation and cleanly removes the patient bed in CT. The quantitative support comes from comparing CT and MR skin surfaces of the same subjects and measuring their distances. The practical conclusion is that method choice should follow the imaging modality and whether full automation or cross-modality accuracy matters more.","feed_headline":"Graphics method beats CT-trained AI on MRI skin","feed_subtitle":"Region-growing works across modalities; the deep-learning model wins on automated CT and bed removal","key_machinery":"The comparison rests on three components. The graphics-based method is a single-seed iterative region-growing algorithm: it floods the volume from a background pixel, marks voxels above an automatically computed skin isovalue as boundary, labels the interior, and applies this per slice with cost linear in the voxel count. For normalized intensities the paper reports that an isovalue of 0.1 works for T2 MRI and CT, while fully automated T2 MRI segmentation uses a gradient image with an isovalue of 0.01. TotalSegmentator is a deep-learning network trained on CT with 104 anatomical structures, including the body and skin; it is fully automatic but tied to its CT training distribution. The quantitative evaluation extracts each skin surface as a mesh with Marching Cubes and compares surfaces using the Hausdorff distance and the mean distance distribution.","core_discovery":"The paper's central claim is that neither approach dominates across all modalities. On CT, TotalSegmentator produces clean skin surfaces, removes the patient bed automatically, and needs no manual tuning; on MRI, it often fails to follow body boundaries because its training data are CT-only. The graphics-based region-growing method segments both CT and MRI without retraining, tracks the MRI skin boundary more closely, and runs in time linear in voxel count on standard hardware, but it leaves the patient bed in CT and yields noisier surfaces. On the 52-subject multimodal dataset, accuracy is defined as agreement between the CT-derived and MR-derived skin surfaces of the same subject, and the paper reports lower Hausdorff and mean distances for the graphics-based method, which it interprets as better cross-modality segmentation.","pith_inferences":["If the MRI gap is caused by training distribution, then fine-tuning TotalSegmentator on MRI or applying unsupervised domain adaptation should shrink or reverse the reported advantage; the paper itself points to domain adaptation as future work.","The paper's cross-modality coincidence metric likely mixes segmentation error with CT-MR misregistration and differences in patient positioning, since no registration or ground-truth contours are reported.","The complementary failure modes (noise from the graphics method, holes and bed artifacts handled by the AI) suggest a hybrid pipeline could be tested: graphics-based surface extraction plus AI-based bed removal and hole filling."],"forward_implications":["Multimodal workflows that must fuse CT with MRI should use the graphics-based method for the MRI skin surface rather than retraining or fine-tuning a CT-only network.","CT-only automated pipelines, especially where removing the patient bed matters, can rely on TotalSegmentator with little or no manual intervention.","The graphics-based method stays accurate on sub-sampled images and runs without a GPU, making it a candidate for resource-constrained settings.","Skin surfaces from either method can support surgical planning tasks such as localizing spinal landmarks for trajectory guidance."],"supporting_citations":[{"why":"Supplies the graphics-based iterative region-growing method that is one of the two compared approaches.","marker":"Paccini et al., 2024"},{"why":"Supplies TotalSegmentator, the deep-learning CT-trained model that is the other compared approach.","marker":"Wasserthal et al., 2023"},{"why":"Provides the annotated multimodal ground-truth dataset used for the quantitative 52-subject comparison.","marker":"Bauer et al., 2021"},{"why":"Releases the CT/MR inhale-exhale dataset on which the quantitative evaluation is run.","marker":"Zöllner, 2022"},{"why":"Defines the Marching Cubes algorithm used to extract and compare the skin surface meshes.","marker":"Lorensen and Cline, 1998"}],"fun_headline_variants":["Graphics method edges out AI for MRI skin","CT-trained AI lags on MRI skin segmentation","Region-growing beats deep learning on MRI skin","Skin segmentation tradeoff: AI for CT, graphics for MRI","AI's CT bias hurts MRI skin accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quantitative comparison defines segmentation quality as agreement between the CT and MR skin surfaces of the same subject, yet the paper reports no registration or ground truth to establish that the two acquisitions are aligned; if they are not, the measured distances reflect alignment error rather than segmentation error.","fun_headline_variants_meta":{"raw":{"variants":["Graphics method edges out AI for MRI skin","CT-trained AI lags on MRI skin segmentation","Region-growing beats deep learning on MRI skin","Skin segmentation tradeoff: AI for CT, graphics for MRI","AI's CT bias hurts MRI skin accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000894,"raw_usage":{"total_tokens":3804,"prompt_tokens":846,"completion_tokens":2958,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":462,"completion_tokens_details":{"reasoning_tokens":2887}},"tokens_in":462,"tokens_out":2958,"duration_ms":22479,"temperature":1.0,"reasoning_tokens":2887,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:03:12.695424+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Segment both modalities in a dataset with expert-drawn skin contours in co-registered CT and MRI, and compare each method's surface to those contours; if TotalSegmentator is as close or closer to the expert MRI skin than the graphics-based method, the central claim fails. Alternatively, recompute the 52-subject Hausdorff and mean distances after explicit rigid registration and check whether the graphics-based advantage persists.","supporting_citations":[{"cited_title":", author Paschina, G","cited_arxiv_id":null,"evidence_quote":"Supplies the graphics-based iterative region-growing method that is one of the two compared approaches."},{"cited_title":", author Breit, H.C","cited_arxiv_id":null,"evidence_quote":"Supplies TotalSegmentator, the deep-learning CT-trained model that is the other compared approach."},{"cited_title":"o nnes, C. , author Segars, W.P. , author Schad, L.R. , author Z \\","cited_arxiv_id":null,"evidence_quote":"Provides the annotated multimodal ground-truth dataset used for the quantitative 52-subject comparison."},{"cited_title":", author Cline, H.E","cited_arxiv_id":null,"evidence_quote":"Defines the Marching Cubes algorithm used to extract and compare the skin surface meshes."}],"review_version":1}