{"id":"005cdf74-e055-4847-9ba6-f793bd39617e","arxiv_id":"2501.04361","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A nnU-Net-based toolkit segments body foreground and anonymized regions in 3D CT/MRI with high Dice scores.","lead":"This paper releases an open-source toolkit with two neural networks that segment anatomical foreground and defaced regions in CT and MRI scans. The toolkit aims to cut self-supervised learning time and prevent models from learning from anonymized data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Foreground CT generalization is untested: training and external MAAL labels share a TotalSegmentator-derived protocol, so the 98.57 'external' Dice may reflect label-process replication rather than anatomical robustness.","rationale":"The reader's CONDITIONAL verdict is appropriate. The central claim requires that the segmentation models are accurate and robust enough for SSL preprocessing; the paper reports strong Dice and HD95 numbers and releases code and weights, which supports reproducibility. The weakest link is the foreground CT evaluation: the only external CT test set, MAAL, is labeled with the same TotalSegmentator-based pipeline used to create training labels. This is not a formal circularity, but it means the external score cannot rule out label-protocol bias. The paper's own failure cases in Fig. 3b and 3c show that mask definition matters (e.g., artificial constant-value boundaries causing false foreground), so the concern is not merely hypothetical. The deface segmentation evaluation is more independent, with AFNI-generated masks on OASIS3 and external OpenNeuro masks, and is less affected by this issue. The absence of direct SSL downstream experiments would strengthen the application claim, but the segmentation accuracy claim is the primary load-bearing assertion, and the concern about label-protocol dependence is the most decisive. Therefore, keeping the verdict at CONDITIONAL is justified, with independent re-labeling of a small external CT set as the key check that could turn it into ACCEPT or REJECT.","tokens_in":5462,"tokens_out":8289,"duration_ms":84685,"concrete_test":"Independently re-annotate the 65 MAAL CT volumes (or at least a random subset of 20) with a foreground mask protocol that does not use TotalSegmentator, e.g., two radiologists manually delineate body/foreground from scratch, then compute Dice/HD95 between the published model's predictions and these independent masks. If the mean Dice falls materially below the reported 98.57 (e.g., below 95), the external generalization claim is inflated by the shared labeling protocol. A cheaper preliminary check is to compute Dice between the raw TotalSegmentator body masks and the final ground-truth labels: agreement above 99% would indicate the manual refinement barely changed the automatic outputs, making the evaluation effectively a test of reproducing TotalSegmentator.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that the CT foreground evaluation demonstrates generalization. In the Materials and methods section, all CT foreground masks are generated with the same multi-step protocol: TotalSegmentator 'body mode' followed by manual refinement with MITK/ITK-SNAP. The MAAL dataset, described as the external CT test set, is therefore external only in image origin, not in label generation. If TotalSegmentator's body-mode definition encodes specific conventions about where air pockets, table edges, or field-of-view boundaries lie, a nnU-Net trained on those masks will reproduce those conventions, and evaluating on masks from the same protocol cannot reveal systematic disagreement with independently labeled foreground. The in-distribution test splits are also drawn 50:50 from the same datasets, so they share scanner and acquisition distributions. Consequently, the reported foreground Dice values (99.56 in-distribution, 98.57 on MAAL) support internal consistency and reproduction of the labeling pipeline, but not the stronger claim that the model matches a foreground definition independent of TotalSegmentator. The paper's own failure cases in Fig. 3b and 3c show sensitivity to boundary definition, making this concern concrete. The deface component is less vulnerable because OASIS3 masks are artificially generated by AFNI and OpenNeuro provides externally derived masks; however, the foreground model is load-bearing for the claimed CT preprocessing utility of the toolkit.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an open-source toolkit consisting of two nnU-Net models for medical imaging preprocessing: a foreground segmentation model that delineates anatomical foreground in CT and MRI scans, and an anonymization-area segmentation model that identifies defaced, refaced, and blurred regions. The foreground model is trained on a multi-dataset collection with labels produced by TotalSegmentator 'body mode' plus manual refinement, and the deface model is trained on OASIS3 images artificially anonymized with AFNI and tested on OpenNeuro data. Results report high Dice coefficients (mean 99.56 on the in-distribution test split, 98.57 on the external CT/MRI set; mean 92.90-99.50 across deface types) and HD95 values. The authors argue these models enable efficient patch sampling and prevent erroneous supervision in reconstruction-based self-supervised learning (SSL) for 3D medical imaging.","tokens_in":5691,"tokens_out":6247,"duration_ms":62614,"significance":"If the robustness claims hold, the toolkit could be a practical resource for the medical imaging SSL community, particularly because the code and weights are made publicly available and the deface evaluation includes a reasonably independent OpenNeuro test set with externally provided masks. The paper also merits credit for honest reporting of failure cases and for using standard, reproducible nnU-Net methodology. However, the central generalization claim for CT foreground is weakened by label-protocol circularity, and the claimed SSL benefits are not demonstrated by any experiment. The significance is therefore conditional: the contribution is a potentially useful engineering resource, but the scientific validation as presented is incomplete.","major_comments":[{"comment":"The external CT foreground evaluation is compromised by label-protocol overlap. The text states that 'a multi-step approach was applied to obtain anatomical foreground masks for both CT and MRI images' immediately after listing all datasets, including MAAL, and no independent label source is described for MAAL. Since the same TotalSegmentator 'body mode' plus manual-refinement pipeline produced both the training labels and the MAAL test labels, the mean Dice of 98.57 on the 'external' test set (Fig. 1) largely measures the model's ability to replicate a labeling protocol rather than its generalization to an independent foreground definition. The same issue applies to the reserved TotalSegmentator MRI test set, whose labels were generated by the same semi-automatic workflow. The boundary-sensitive failure cases in Fig. 3b and 3c make this concern concrete. Please add an independently labeled CT foreground test set (e.g., manual body-region delineations from another source), report per-dataset results, or explicitly re-frame the MAAL and TotalSegmentator MRI results as protocol-replication checks rather than evidence of external generalization.","section":"§2 Materials and methods, Anatomical foreground segmentation; §3 Results"},{"comment":"The paper's motivating claims about SSL utility are not tested. The abstract and discussion state that the foreground model 'reduces training time' by enabling selective patch sampling and that the deface model prevents 'erroneous supervision in reconstruction-based SSL methods', but no SSL method is trained or evaluated, and no wall-clock training-time or throughput measurements are reported. The reported Dice and HD95 scores, while impressive, do not by themselves establish that the toolkit improves SSL preprocessing in practice. Please add at least one experiment that integrates both models into a representative SSL pipeline, comparing training time and downstream task performance with and without the toolkit, or substantially soften the claims to describe the models as segmentation tools whose SSL benefit remains to be demonstrated.","section":"§3 Results; §4 Discussion; Abstract"}],"minor_comments":[{"comment":"The 'OpenNeuro - external' row in Table 2 is described in the text as 14 manually annotated images, but the reported standard deviation of 0.05 percentage points for both Dice and HD95 is implausibly small for 14 images and suggests either a misreport or a much smaller sample; please clarify the exact number of images and the annotation protocol.","section":"§2 Deface area segmentation; Table 2"},{"comment":"The boxplots aggregate results across datasets and anonymization types; please report the number of cases per group, per-dataset medians, and the whisker/IQR definitions so readers can assess variability and outliers.","section":"Fig. 1 and Fig. 2; §3 Results"},{"comment":"The sentence 'with some datasets providing original deface masks originating from theDefaced method' appears to contain a typo ('theDefaced') and should also specify which OpenNeuro sub-datasets and which anonymization methods are included.","section":"§2 Deface area segmentation"},{"comment":"The abstract contains the grammatical error 'The weights and code is available'; consider adding a dedicated Code/Data Availability section and using 'are available'.","section":"Abstract and elsewhere"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads more like a technical report than a full research paper: the core novelty is two nnU-Nets with semi-automatically generated labels, and the central SSL-benefit claim is unverified. The foreground evaluation circularity should be addressed head-on; an independent CT foreground test set would make the contribution credible. If the authors add the missing SSL experiments and independent evaluation, the paper could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a toolkit paper, and judged as one it is mostly fine. The authors train nnU-Net models for two tasks the SSL community actually needs: foreground segmentation to sample patches efficiently in 3D, and anonymized-region segmentation to mask out defaced areas from reconstruction losses. They report Dice above 98.5 across the board, ship code and weights, and document the label-generation pipeline. That is real value. The MRI foreground model is genuinely new, since TotalSegmentator does not work on MRI, and the paper shows a sensible way to bootstrap labels with thresholding, manual correction, and iterative refinement. The deface model covers three AFNI anonymization modes and is validated externally on OpenNeuro, which is a reasonable setup.\n\nThe soft spot is the CT foreground evaluation, and it is load-bearing. Every CT foreground mask in the paper, training and the MAAL \"external\" test, comes from the same pipeline: TotalSegmentator body mode plus manual refinement. So the 98.57 mean Dice on MAAL does not demonstrate generalization to an independent labeling protocol; it demonstrates that the model reproduces TotalSegmentator's conventions about where the body ends, where table edges lie, and how field-of-view boundaries are handled. The paper even shows failure cases where those conventions break down. The deface segmentation is on firmer ground because OASIS3 masks are generated by AFNI and OpenNeuro provides independently derived masks, though the blurring test set involves only 14 manual annotations.\n\nA second gap: the abstract and intro promise SSL benefits, reduced training time and avoiding erroneous supervision, but no SSL experiment is run. That is an omission, not a deal-breaker, but it leaves the motivating claim unverified.\n\nI agree with the conditional verdict. The central segmentation results are likely reproducible and useful, but the overstatement about external robustness should be fixed. The right remedy is an independent label audit on a small CT set, say 20-30 scans manually labeled by a different rater, and ideally a simple SSL proof-of-concept showing time savings or a downstream gain. That would turn a plausible toolkit into a solid one.\n\nI would bring this to reading group only if someone is about to use the toolkit. It deserves peer review, as a competent, well-documented engineering contribution with public assets, but it needs honest revision before acceptance. Recommend a major revision addressing the label circularity and either adding the SSL validation or toning down the claims.","headline":"Useful, well-packaged preprocessing toolkit for SSL, but the CT foreground generalization claim is weaker than it looks because the external test shares the labeling protocol.","tokens_in":6225,"tokens_out":2148,"would_cite":true,"duration_ms":21866,"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 claims that two nnU-Net segmenters—one for anatomical foreground, one for anonymized facial regions—can be trained once on mixed CT and MRI data and then preprocess unannotated 3D scans for self-supervised learning, with mean…","keywords":["foreground segmentation","anonymization segmentation","self-supervised learning","3D medical imaging","CT","MRI","deface","nnU-Net"],"falsifier":"Have independent experts manually trace foreground and anonymized regions on a sample of CT and MRI volumes that were never processed by the paper's labeling pipeline, then compare network predictions to those manual masks with Dice and HD95; if the scores fall far below 99.56 and 98.57 for foreground or below 98.5 for deface, the generalization claim is refuted.","tokens_in":5263,"feed_emoji":"🩻","tokens_out":8432,"duration_ms":77420,"temperature":0.7,"pith_summary":"The paper aims to remove two obstacles to self-supervised learning on 3D medical images: much of a scan is empty air, wasting computation, and anonymized facial regions inject distorted signals into reconstruction-based SSL losses. It builds an open-source toolkit with two segmentation networks, one that marks all anatomical foreground voxels in CT and MRI and one that marks defaced, refaced, or blurred regions. On internal test splits the foreground network reaches a mean Dice of 99.56, and on external CT data 98.57; the anonymization network stays at or above 98.5 mean Dice across deface, reface, reface-plus, and an external blurred-face set. The authors conclude the models are accurate and dependable enough to serve as preprocessing tools for 3D medical SSL.","feed_headline":"Toolkit segments body and anonymized face regions at Dice above 98","feed_subtitle":"Two open-source networks help self-supervised models avoid empty air and masked facial areas in CT and MRI.","key_machinery":"The machinery is a pair of nnU-Nets, a self-configuring deep-learning segmentation architecture, one trained to label anatomical foreground and one trained to label anonymized regions. Both use a 192 by 192 by 192 patch size and z-score normalization so that a single set of weights covers CT and MRI. The foreground labels were produced by taking an existing automatic body-region segmentation tool's output, manually refining it, and then using an initial model to speed up labeling the remaining MRI volumes; the anonymization labels come from applying three common defacing algorithms to non-anonymized scans. The network outputs are intended as masks: the foreground map guides patch sampling, and the anonymization map tells a reconstruction loss which voxels to ignore.","core_discovery":"The central claim is that a single nnU-Net, trained on a mixed CT and MRI dataset with a 192 by 192 by 192 patch size and z-score normalization, can segment the anatomical foreground accurately enough for both modalities, and a second nnU-Net can localize anonymized facial regions across three anonymization schemes plus real-world blurred faces. On the test portion of the training data, the foreground model reaches a mean Dice of 99.56; on an external CT dataset it reaches 98.57. For anonymization, the reported mean Dice values are 99.05 for deface, 99.50 for reface, 99.01 for reface plus, 92.90 on the external OpenNeuro-style collection, and 98.67 on the external blurred-face subset. The authors argue these results make it possible to sample only informative patches and to mask altered regions out of SSL reconstruction losses without needing dataset-specific adapters.","pith_inferences":["The external CT evaluation is weaker than it looks: the external CT test set was labeled with the same segmentation pipeline that produced the CT training masks, so the reported external Dice may reflect labeling-protocol overlap rather than generalization to independently drawn ground truth.","The two masks compose naturally: an SSL loss mask of 'foreground AND not anonymized' would combine both networks and is the direct practical use the paper leaves implicit.","The foreground network's stated failure mode on scans with large empty fields and constant-value reconstruction boundaries suggests a cheap pre-filter that flags such slabs would remove most remaining errors; this extension is not tested in the paper."],"forward_implications":["Foreground masks let an SSL trainer sample patches only from anatomical content, so training runs avoid the large air-filled parts of 3D scans.","Anonymization masks can be fed into reconstruction-based SSL losses, so defaced, refaced, or blurred voxels do not produce erroneous supervision signals.","Because one foreground model handles both CT and MRI with the same weights and z-score normalization, a single preprocessing step replaces modality-specific pipelines.","The deface model's coverage of all three common anonymization schemes and an external blurred-face dataset means it should apply to the main ways public medical data are currently altered."],"supporting_citations":[{"why":"Supplies the automatic body-region segmentation tool whose outputs are manually refined to label most CT foreground masks.","marker":"[17]"},{"why":"Supplies the three anonymization algorithms (deface, reface, reface plus) used to generate training images and masks.","marker":"[3]"},{"why":"Supplies the large set of non-anonymized scans that are artificially anonymized to create the deface training set.","marker":"[8]"},{"why":"Supplies the external collection of publicly shared, already anonymized MRI scans with masks, used as the external test set.","marker":"[10]"},{"why":"Provides the self-configuring segmentation method on which both trained networks are built.","marker":"[6]"},{"why":"Supplies the external CT dataset used to test foreground segmentation on data outside the training distribution.","marker":"[9]"},{"why":"Supplies 600 CT and MRI scans that contribute foreground training data.","marker":"[7]"},{"why":"Give the external MRI datasets whose original deface masks were used as test labels.","marker":"[4,11,14,5]"}],"fun_headline_variants":["nnU-Net pair hits Dice 98+ for body and anonymized face in CT/MRI","Two nnU-Nets segment body and anonymized areas with Dice 98+","Open-source toolkit for body/face mask segmentation with Dice >98 in CT/MRI","Segment body and anonymized face regions at Dice >98 for SSL data prep","Dice 98+ for body and anonymized face masks from a unified CT/MRI toolkit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on the ground-truth masks being correct, especially the CT foreground labels that come mostly from an automatic segmentation tool with manual refinements; if that tool has systematic biases, Dice scores measured against its masks overstate real accuracy on independently labeled scans.","fun_headline_variants_meta":{"raw":{"variants":["nnU-Net pair hits Dice 98+ for body and anonymized face in CT/MRI","Two nnU-Nets segment body and anonymized areas with Dice 98+","Open-source toolkit for body/face mask segmentation with Dice >98 in CT/MRI","Segment body and anonymized face regions at Dice >98 for SSL data prep","Dice 98+ for body and anonymized face masks from a unified CT/MRI toolkit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001191,"raw_usage":{"total_tokens":4880,"prompt_tokens":876,"completion_tokens":4004,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":3891}},"tokens_in":492,"tokens_out":4004,"duration_ms":26623,"temperature":1.0,"reasoning_tokens":3891,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:33:56.391360+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have independent experts manually trace foreground and anonymized regions on a sample of CT and MRI volumes that were never processed by the paper's labeling pipeline, then compare network predictions to those manual masks with Dice and HD95; if the scores fall far below 99.56 and 98.57 for foreground or below 98.5 for deface, the generalization claim is refuted.","supporting_citations":[{"cited_title":"Radiology: Artificial Intelligence5(5) (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the automatic body-region segmentation tool whose outputs are manually refined to label most CT foreground masks."},{"cited_title":"Computers and Biomedical Research29(3), 162–173 (1996)","cited_arxiv_id":null,"evidence_quote":"Supplies the three anonymization algorithms (deface, reface, reface plus) used to generate training images and masks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the large set of non-anonymized scans that are artificially anonymized to create the deface training set."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the external collection of publicly shared, already anonymized MRI scans with masks, used as the external test set."},{"cited_title":"Nature Methods (2021)","cited_arxiv_id":null,"evidence_quote":"Provides the self-configuring segmentation method on which both trained networks are built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the external CT dataset used to test foreground segmentation on data outside the training distribution."}],"review_version":1}