{"id":"f38c3655-b66d-43bd-9ac2-26df009c421e","arxiv_id":"2412.06507","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"BATseg adds a class-wise tumor surface distance field branch to nnUNet and reports 10-point Dice gains on a new 653-patient spinal cord tumor MRI dataset.","lead":"The paper introduces a new MRI dataset of 653 spinal cord tumor patients and a segmentation method, BATseg, that learns a per-type distance field to sharpen tumor boundaries. If the results hold, it gives clinicians a rare benchmark and a technique that may transfer to other small, hard-to-segment lesions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The published boundary-aware loss in Eq. 3 has an inverted sign: minimizing it drives predicted distance fields away from ground truth, so the central training objective as written cannot produce the claimed results.","rationale":"The reader's conditional verdict already notes the sign inconsistency in Eq. 3, so my concern aligns with part of the rationale. However, the reader's stated weakest_assumption focuses on whether the distance-field target adds information beyond direct CE/Dice losses, whereas I treat the inverted sign as the single most load-bearing issue: as written, the loss cannot train a distance field at all, regardless of whether such a field would be useful. The sign error is internal and directly falsifiable by a toy gradient check. The normalization formula in Eq. 2 is also garbled—the stated denominator and /2 do not map the field to [0,1] as claimed—but that is secondary once the loss sign is fixed. No significance testing is reported, and the data/code links are not verifiably accessible, but these are standard reproducibility concerns rather than flaws in the argument's logic. A corrected sign and a successful re-run would restore the paper's internal consistency; without that correction, the reported results are not attributable to the proposed method. I therefore keep the reader's conditional verdict unchanged rather than escalating to reject, because the error is plausibly a typo that can be resolved by code release or an erratum.","tokens_in":14682,"tokens_out":4019,"duration_ms":46383,"concrete_test":"Implement Eq. 3 literally on a single-voxel toy problem with target \\bar f = 1 and initialize f = 0.5. Run gradient descent with learning rate 0.1 for 100 steps. If Eq. 3 is used as written, f diverges to large values instead of approaching 1. Then repeat with the sign flipped (+|f - \\bar f|^3) and verify convergence. Additionally, re-run the BATseg versus nnUNet comparison on one fold of the released spinal cord dataset using the corrected sign; if BATseg does not reproduce the Table 3 margin, the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that an auxiliary surface-distance head learns a truncated normalized distance field and thereby improves segmentation—rests on Eq. 3. As printed, ℓba = -(f - \\bar f)^2 |f - \\bar f| = -|f - \\bar f|^3. Its gradient with respect to f is -3|f - \\bar f|(f - \\bar f). Gradient descent on this loss therefore moves f away from \\bar f: a too-high prediction is pushed higher, a too-low prediction is pushed lower, and the unconstrained minima are f → ±∞. The loss is maximized, not minimized, at f = \\bar f. Consequently, the method as published does not learn a distance field; it is repelled from it. This is a reproducibility-critical internal inconsistency, not a disagreement with external consensus. The fact that all ablations and final results were presumably run with some implemented loss means either the equation is missing a plus sign—in which case the paper states the wrong loss—or the experiments used a different objective than reported—in which case the comparisons do not test the described method. The KNIGHT ablations cannot resolve this, because every row of Tables 5–8 inherits whichever sign was actually optimized. The absence of a control auxiliary branch is a secondary concern about isolating the mechanism, but Eq. 3 is the load-bearing point: it must be corrected and re-verified before the reported improvement can be attributed to the proposed boundary-aware loss.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BATseg, a method that augments an nnUNet backbone with an auxiliary head that predicts a truncated, normalized, class-wise tumor surface distance field, supervised by a new \"boundary-aware\" loss. The authors also introduce a new dataset of 653 gadolinium-enhanced T1-weighted 3D MRI scans of spinal cord tumors with four tumor types and five-fold splits. On this dataset and on the public KNIGHT kidney-tumor dataset, BATseg is reported to outperform several end-to-end and two-stage baselines in Dice score and 95th-percentile Hausdorff distance. The paper includes per-fold results, ablations on KNIGHT, and qualitative visualizations.","tokens_in":14948,"tokens_out":6585,"duration_ms":64103,"significance":"If the results hold, the new spinal cord tumor dataset and the simple auxiliary-surface-distance training strategy would be useful contributions to medical image segmentation, particularly for small, irregular tumor structures. The paper provides consistent five-fold cross-validation results and ablation studies on a public dataset, which are strengths. However, the central method as written contains two internal inconsistencies in the definition of the ground-truth distance field and the boundary-aware loss, the most serious being a sign error in Eq. (3) that makes the printed loss non-positive and repulsive from the target. As a result, the reported experiments cannot be reproduced from the described objective, and the attribution of the observed improvements to the proposed loss is not supported as written. These issues are correctable, but they are load-bearing for the manuscript's central claim.","major_comments":[{"comment":"The boundary-aware loss as printed, ℓba = -(f - \\bar f)^2 |f - \\bar f|, is non-positive and unbounded below. Its gradient with respect to f is -3|f-\\bar f|(f-\\bar f), so gradient descent pushes f away from \\bar f: the loss is maximized, not minimized, when f = \\bar f. This cannot be the objective used in the experiments, because Table 8 shows that adding the squared term improves Dice by about 1%, whereas the printed negative sign would make the squared term actively harmful. Please correct the sign in Eq. (3) and explicitly confirm that the reported results were obtained with the corrected objective.","section":"Sec. 4.3, Eq. (3)"},{"comment":"The text states that \"all previously truncated values\" are kept as zeros, but the normalization formula in Eq. (2) does not do this: for a truncated voxel with d'_j = 0, the equation gives \\bar d_j = (0 + 1)/2 = 0.5. Thus truncated background voxels are labeled 0.5, not 0, which contradicts the description in Sec. 4.3 that background voxels \"tend to be easily optimized to be zeros\" and shifts the target distribution of the distance-field head. Please correct Eq. (2) or the accompanying description, and specify the actual normalization used in the experiments.","section":"Sec. 4.2, Eq. (2)"},{"comment":"The ablation studies vary properties of the distance field (truncation, normalization, class-wise vs. class-agnostic, loss form) but do not include a control branch with an auxiliary target that carries no geometric information (e.g., a constant, random, or unrelated head). Without such a control, the improvement over nnUNet could be attributed to the extra parameters or to the presence of an auxiliary supervision signal in general, rather than specifically to the truncated normalized tumor-surface distance field. Adding this control would substantially strengthen the central causal claim.","section":"Sec. 5.3, Tables 5-8"}],"minor_comments":[{"comment":"The word \"emperically\" should be \"empirically\".","section":"Sec. 5.3"},{"comment":"In the author block, \"Insititute\" and \"Acadamy\" should be \"Institute\" and \"Academy\".","section":"Author affiliations"},{"comment":"The GitHub link in Sec. 1 is written as \"https://github.com/vLAR-group/BA T seg\" with a space; please provide a working, correctly spaced URL.","section":"Data availability"},{"comment":"The tables list \"Hausdorff Distance (HD)\" but the metrics paragraph in Sec. 5 specifies the 95th-percentile Hausdorff distance; please make the captions consistent with the metric definition.","section":"Tables 3-8"},{"comment":"The rows for nnUNet+UEnc with mask and nnUNet+UEnc w/o mask are identical (NoAT Dice 50.5, CanAT Dice 0.0); please explain why they coincide or correct the results, since the text describes them as distinct methods.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Eq. (3) appears to be a typographical slip rather than a deliberate design choice, given the ablation results and the surrounding discussion. Nevertheless, the paper in its current form is not reproducible from the stated equations, and the authors must correct the loss and the normalization formula, then confirm that the reported experiments use the corrected definitions. I also recommend asking the authors to provide the control auxiliary-branch experiment and to make the dataset link functional. If these issues are fully addressed, the paper could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe real news here is the dataset: 653 patients, four spinal cord tumor types, first public multiclass spinal cord tumor segmentation dataset. That is a solid contribution and likely to be used. The method is a boundary-aware auxiliary head with class-wise truncated normalized distance fields; the ingredients are established, the combination is new enough. The reported gains over nnUNet are consistent across folds on the in-house data and hold up on KNIGHT.\n\nBut there is a load-bearing error in the published loss. Eq. (3) defines ℓba = -(f - \\bar f)^2 |f - \\bar f| = -|f - \\bar f|^3. Its gradient pushes f away from \\bar f; the loss is minimized at ±∞, not at the target. As written, the method could not learn a distance field, and the ablations in Table 8 inherit whatever sign the authors actually used. Either the equation is missing a plus sign or the experiments used something else. Either way, the paper as printed is not reproducible, and the central claim cannot be checked. The truncation and normalization equations are also garbled, though that is cosmetic next to the sign.\n\nThe other soft spots: no significance testing on the five folds (the per-fold tables show some overlap, e.g., Fold 5 astrocytoma 28.4 vs nnUNet 17.3, but several folds overlap), no control auxiliary branch to show the distance field specifically helps, and the promised code/data link is not verified. The class-aware vs class-agnostic ablation is informative.\n\nIf the sign is fixed and re-run, the method might work—the ablations suggest the distance field helps. But as submitted, the central mechanism as described is mathematically impossible. I would send it to peer review because the dataset needs to be in the literature, and the method is correctable with a re-run and careful rewriting. The authors should be asked to fix Eq. (3), correct the normalization, add statistical summaries, and release the data/code.\n\nWho gets value: anyone working on spinal cord tumor segmentation or boundary-aware losses; the dataset is the main event. A serious referee should engage, but not accept the empirical claims without seeing the corrected loss.","headline":"The spinal cord tumor dataset is a genuine contribution, but the published boundary-aware loss has an inverted sign that makes the method as written impossible; needs major revision.","tokens_in":15506,"tokens_out":2647,"would_cite":true,"duration_ms":26019,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding per-class surface distance maps to a segmentation backbone improves spinal cord tumor segmentation, backed by a new 653-patient MRI dataset","keywords":["spinal cord tumor segmentation","3D MRI","boundary-aware loss","surface distance field","multiclass segmentation","nnUNet","medical image segmentation","KNIGHT kidney tumor dataset"],"falsifier":"Train the same nnUNet backbone with an auxiliary head that regresses a scrambled version of the distance field, or an unrelated per-voxel target, under the same loss weighting; if Dice and Hausdorff gains match BATseg's, the boundary-distance geometry is not the operative cause. A second test is to re-run the five-fold evaluation on masks re-annotated by a different center or scanner: if the truncation and normalization ablations stop favoring the full recipe, the claim that the distance field carries the gain is falsified.","tokens_in":14484,"feed_emoji":"🧠","tokens_out":11270,"duration_ms":97064,"temperature":0.7,"pith_summary":"This paper tries to establish that the hard part of segmenting spinal cord tumors is their boundaries, and that a network can learn those boundaries by regressing a per-class distance-to-surface map rather than only labeling voxels. To test this, the authors built the first large-scale multiclass spinal cord tumor segmentation dataset, 653 patients with gadolinium-enhanced T1-weighted 3D MRI covering meningiomas, ependymomas, astrocytomas, and hemangioblastomas, split into five folds. They augmented the nnUNet backbone with a second head that predicts a truncated, normalized, class-wise tumor surface distance field, supervised by a boundary-aware loss in addition to cross-entropy and Dice. They report that this augmented model beats the backbone and several transformer and two-stage baselines on both their dataset and the public KNIGHT kidney tumor dataset, with mean Dice on the spinal dataset rising from 52.8% to 62.1% and Hausdorff distance falling from 161.4 mm to 108.3 mm. If correct, the result gives clinicians a more accurate automatic way to measure tumor size, location, and type, and gives the field a reusable geometric prior for other small, boundary-dominated lesions.","feed_headline":"Boundary-distance head lifts spinal tumor Dice from 52.8 to 62.1","feed_subtitle":"A new 653-patient MRI dataset and a per-class distance-field loss sharpen small, boundary-variable tumor outlines.","key_machinery":"The load-bearing object is the class-wise tumor surface distance field: a per-voxel map, computed from the ground-truth mask for each tumor type separately, in which voxels inside the tumor store their positive nearest distance to the tumor boundary, voxels outside store negative distances truncated to zero beyond the maximum interior distance, and all retained values are normalized into [0,1]. It is supervised by the multiclass boundary-aware loss, which is the negative of the squared prediction error times the absolute prediction error, averaged over the volume; this focuses optimization on the nonzero boundary band instead of the large zero background. The field is predicted by a single convolutional head running in parallel with the segmentation head of the nnUNet backbone, and the three losses (cross-entropy, Dice, and boundary-aware) are trained jointly with equal weights. The distance field is the mechanism that makes the segmentation head's predicted surfaces contract toward the annotated tumor boundary.","core_discovery":"On its own terms, the paper's claim is that multiclass tumor segmentation is improved when the network is forced, in parallel with ordinary voxel labeling, to reconstruct a surface distance field for each tumor class separately. The field is defined per class from the manual mask: voxels inside the tumor carry their positive nearest distance to the boundary, voxels outside carry negative distances until truncation at the maximum interior distance, and surviving values are normalized to [0,1]. A head on the nnUNet backbone regresses this field under a loss that weights voxels by the product of squared and absolute error, so the optimization concentrates on nonzero (near-boundary) voxels; the authors report this branch is what tightens predicted surfaces. The evidence is experimental: across five folds of the new spinal cord dataset, BATseg reports mean Dice 62.1% versus 52.8% for the nnUNet backbone and mean Hausdorff distance 108.3 mm versus 161.4 mm, and on the KNIGHT kidney dataset it reports mean Dice 40.7% versus 33.8% for nnUNet. Ablations attribute the gains to truncation, normalization, the multiclass (class-wise) definition of the field, and the squared-error weighting in the boundary-aware loss.","pith_inferences":["Editorial inference: the class-wise distance-field recipe is not tested on other small-lesion segmentation tasks, but the KNIGHT transfer result suggests it is a candidate drop-in auxiliary loss for any multiclass segmentation where boundaries rather than bulk volume carry the diagnostic signal.","Editorial inference: because the paper does not include a control branch with an unrelated auxiliary regression target, the reported gains are not yet isolated to the geometric content of the distance field; a scrambled-distance control would settle that.","Editorial inference: a natural next experiment is re-annotation or cross-scanner evaluation; the paper explicitly lists scanner generalization as untested, so a held-out scanner study would show whether the distance-field prior survives realistic clinical variation.","Editorial inference: the dataset's class imbalance and small tumor volumes suggest that reporting per-class precision-recall or volume error, beyond Dice and Hausdorff, would clarify where the boundary head helps most."],"forward_implications":["On the new 653-patient spinal cord tumor dataset, BATseg reports a mean Dice of 62.1% versus 52.8% for its nnUNet backbone, and a mean Hausdorff distance of 108.3 mm versus 161.4 mm across the four tumor types.","On the public KNIGHT kidney tumor dataset, the same recipe reports mean Dice 40.7% versus 33.8% for nnUNet, indicating the boundary-distance loss transfers to a different tumor domain.","Per-class distance fields outperform a class-agnostic field on KNIGHT by 4.3 points in mean Dice, supporting the design choice to treat each tumor type's surface separately.","Ablations show that truncating the outside-distance field, normalizing the retained values, and weighting the loss by squared error each contribute to the reported gains.","Segmentation remains hardest for astrocytomas (35.2% Dice), which the paper attributes to their similar appearance to ependymomas and resulting misclassification."],"supporting_citations":[{"why":"nnU-Net is the backbone network and the strongest baseline; the boundary-aware head is added to it and its cross-entropy and Dice losses are reused.","marker":"[14]"},{"why":"The KNIGHT kidney tumor dataset is the public benchmark used for external validation and for all ablations.","marker":"[2]"},{"why":"The M&Ms challenge supplies the evaluation protocol, Dice coefficient and 95th-percentile Hausdorff distance, used in all comparisons.","marker":"[5]"},{"why":"Lemay et al. is the closest prior work on multiclass intramedullary spinal cord tumor segmentation, the approach the paper positions against.","marker":"[19]"},{"why":"Reza et al. is an earlier spinal tumor segmentation study whose dataset is not public, motivating the new public dataset.","marker":"[25]"},{"why":"nnFormer is one of the transformer-based end-to-end baselines compared on both datasets.","marker":"[38]"},{"why":"3D UX-Net is another end-to-end baseline compared on both datasets.","marker":"[18]"},{"why":"Swin UNETR is the third end-to-end transformer baseline compared on both datasets.","marker":"[10]"},{"why":"Recent signed-distance medical segmentation work that the proposed boundary-aware distance field contrasts with and extends to multiclass 3D cases.","marker":"[15]"}],"fun_headline_variants":["Boundary-aware loss boosts spinal tumor Dice to 62.1%","Tiny spinal tumors segmented better: Dice up from 52.8 to 62.1","Multiclass distance field sharpens spinal tumor boundaries on MRI","New 653-patient spinal tumor dataset and distance-field loss improve Dice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported improvement rests on the assumption that the manual tumor outlines are precise enough that forcing the network to also learn a per-type map of distance-to-boundary adds real geometric information beyond what ordinary voxel-labeling losses already provide.","fun_headline_variants_meta":{"raw":{"variants":["Boundary-aware loss boosts spinal tumor Dice to 62.1%","Tiny spinal tumors segmented better: Dice up from 52.8 to 62.1","Multiclass distance field sharpens spinal tumor boundaries on MRI","New 653-patient spinal tumor dataset and distance-field loss improve Dice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000645,"raw_usage":{"total_tokens":2998,"prompt_tokens":1014,"completion_tokens":1984,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":1903}},"tokens_in":630,"tokens_out":1984,"duration_ms":14570,"temperature":1.0,"reasoning_tokens":1903,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:35:16.791263+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same nnUNet backbone with an auxiliary head that regresses a scrambled version of the distance field, or an unrelated per-voxel target, under the same loss weighting; if Dice and Hausdorff gains match BATseg's, the boundary-distance geometry is not the operative cause. A second test is to re-run the five-fold evaluation on masks re-annotated by a different center or scanner: if the truncation and normalization ablations stop favoring the full recipe, the claim that the distance field carries the gain is falsified.","supporting_citations":[{"cited_title":"Nature Methods (2020)","cited_arxiv_id":null,"evidence_quote":"nnU-Net is the backbone network and the strongest baseline; the boundary-aware head is added to it and its cross-entropy and Dice losses are reused."},{"cited_title":"SSRN (2023)","cited_arxiv_id":null,"evidence_quote":"The KNIGHT kidney tumor dataset is the public benchmark used for external validation and for all ablations."},{"cited_title":"IEEE Transactions on Medical Imaging (2021)","cited_arxiv_id":null,"evidence_quote":"The M&Ms challenge supplies the evaluation protocol, Dice coefficient and 95th-percentile Hausdorff distance, used in all comparisons."},{"cited_title":"NeuroImage: Clinical (2021)","cited_arxiv_id":null,"evidence_quote":"Lemay et al. is the closest prior work on multiclass intramedullary spinal cord tumor segmentation, the approach the paper positions against."},{"cited_title":"In: Medi- calImaging2019:BiomedicalApplicationsinMolecular,Structural,andFunctional Imaging","cited_arxiv_id":null,"evidence_quote":"Reza et al. is an earlier spinal tumor segmentation study whose dataset is not public, motivating the new public dataset."},{"cited_title":"TIP (2023)","cited_arxiv_id":null,"evidence_quote":"nnFormer is one of the transformer-based end-to-end baselines compared on both datasets."},{"cited_title":"ICLR (2023)","cited_arxiv_id":null,"evidence_quote":"3D UX-Net is another end-to-end baseline compared on both datasets."},{"cited_title":"MICCAI Workshop (2022)","cited_arxiv_id":null,"evidence_quote":"Swin UNETR is the third end-to-end transformer baseline compared on both datasets."},{"cited_title":"MIDL (2024)","cited_arxiv_id":null,"evidence_quote":"Recent signed-distance medical segmentation work that the proposed boundary-aware distance field contrasts with and extends to multiclass 3D cases."}],"review_version":1}