{"id":"43723da6-3443-4473-a46b-3ae7080d7b19","arxiv_id":"2507.15777","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Tree-aware loss functions, especially a Wasserstein compound loss, improve whole brain parcellation and sparse hyperspectral surgical segmentation over standard CE+Dice baselines.","lead":"Two segmentation losses that use a hierarchy of label classes improve medical image segmentation on whole brain MRI and surgical hyperspectral imaging, with a Wasserstein-based loss giving the strongest gains on brain parcellation. The paper shows that penalizing errors by semantic tree distance and supervising aggregated class probabilities can outperform the standard Dice-plus-cross-entropy loss in these settings.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GIF pseudo-ground-truth validity is the load-bearing weak point: the AOMIC/IXI gains in Sec. 5.1 may reflect alignment with GIF's systematic label errors rather than anatomical accuracy, and the MB42 manual evidence alone does not establish the cross-dataset claim.","rationale":"I selected the GIF pseudo-ground-truth issue because it is the only concern that, if it lands, would weaken the paper's strongest empirical claim rather than a secondary application. The paper's main quantitative support for WBP is Table 1, and two of the three test datasets in that table are evaluated exclusively against GIF-derived masks. The paper acknowledges the limitation in Sec. 4.1.1, but acknowledgement does not remove the dependency: the reported gains in NSD_small and Dice_small on AOMIC/IXI could be inflated if the semantic loss selectively learns GIF's error patterns. The MB42 manual columns are genuine evidence in the paper's favour, and AOMIC/IXI-trained models also improve on MB42 manual labels, which partially mitigates the concern. However, the specific claim of consistent gains in pseudo-ground-truth agreement cannot be separated from the reference generation without the dual-reference check I propose. I therefore keep the reader's CONDITIONAL verdict unchanged rather than rejecting the paper, because the manual MB42 results and the statistical corrections in Table A1 provide real support for the loss.","tokens_in":26894,"tokens_out":11547,"duration_ms":123833,"concrete_test":"Use the existing MB42 data, which has both manual labels and GIF pseudo-ground-truth masks, to recompute the MB59-trained L_wass+seg and L_seg predictions per subject for Dice_small and NSD_small under both references. Form per-subject paired differences (L_wass+seg minus L_seg) for the GIF reference and for the manual reference, and check whether positive GIF-reference differences co-occur with positive manual-reference differences on matched subjects. If they do, the pseudo-ground-truth concern is mitigated; if GIF-reference gains appear without manual-reference gains, the AOMIC/IXI cross-dataset claim in Sec. 5.1 should be downgraded to a tool-agreement result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The cross-dataset WBP claim in Sec. 5.1 depends on AOMIC and IXI metrics computed against GIF-derived pseudo-ground-truth masks, as Sec. 4.1.1 explicitly acknowledges. Because the same GIF pipeline supplies the reference masks for these test sets and, for the AOMIC/IXI training rows, the training labels, any tendency of L_wass+seg to reproduce GIF's systematic labelling errors will be measured as an apparent improvement. This is not a purely theoretical risk: the paper's own control on MB42 reports GIF-vs-manual label-wise Dice of only 78.6±15.4, and the claimed gains are concentrated in Dice_small and NSD_small, where atlas-fusion label noise is largest. The MB42 manual-reference columns provide real evidence that L_wass+seg improves segmentation quality, but they do not show that GIF-agreement improvements track manual-reference improvements on matched subjects. Until that link is demonstrated, the AOMIC/IXI columns support only the weaker statement that the loss increases agreement with an automated tool, not that it improves anatomical accuracy on those datasets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two tree-based semantic loss functions for multi-class medical image segmentation: a Wasserstein distance-based loss combined with a standard segmentation loss (L_wass+seg, Eq. 3) and a tree-weighted semantic cross-entropy loss (L_twce+seg, Eq. 6). Both losses encode a label hierarchy derived from anatomical knowledge or expert consensus. The authors integrate these losses into a positive-only sparse-annotation framework with OOD-based background detection, and evaluate on whole-brain parcellation (MB42, AOMIC, IXI) and surgical hyperspectral imaging (HSI) with sparse annotations. The main reported findings are that L_wass+seg consistently improves over the nnU-Net CE+Dice baseline on WBP, especially for NSD and small-structure metrics, and that hierarchy-aware losses improve top-level and, to a lesser extent, leaf-level HSI segmentation.","tokens_in":27139,"tokens_out":5064,"duration_ms":55704,"significance":"If the results are fully supported, the paper would provide a practical way to inject label semantics into high-cardinality segmentation with no architectural change, and would extend sparse positive-only training to hierarchical settings. Strengths of the manuscript include the use of established baselines (nnU-Net, EfficientNet-U-Net), paired subject-level statistical tests with Holm-Bonferroni correction on the WBP task, public code, and the MB42 manual-reference evidence, which is the most trustworthy part of the evaluation. However, the cross-dataset WBP claims rely on GIF pseudo-ground-truth agreement, the HSI leaf-level F1 scores are near zero and lack a reported baseline, and key hierarchy hyperparameters are chosen outside the target evaluation or in unreported preliminary experiments. These issues currently limit the strength of the central claims.","major_comments":[{"comment":"The cross-dataset WBP claim for AOMIC and IXI rests entirely on GIF-derived pseudo-ground-truth masks, as Section 4.1.1 explicitly acknowledges. The manuscript's own MB42 control reports GIF-vs-manual Dice of 78.6±15.4, so GIF label error is substantial. Because the same GIF pipeline supplies the reference masks for the AOMIC/IXI test sets, the reported L_wass+seg gains on those datasets could reflect better reproduction of GIF's systematic labelling errors rather than improved anatomical accuracy. The MB42 manual-reference columns support the claim for MB42, but they do not establish that GIF-agreement gains track manual-reference gains on matched subjects. The authors should either provide such matched-subject evidence on MB42 or re-evaluate at least a subset of AOMIC/IXI with manual WBP labels; without this, the non-MB42 WBP results should be described only as improved agreement with an automated tool.","section":"§5.1, §4.1.1"},{"comment":"The HSI leaf-level F1 scores are reported as approximately 0.069 and 0.073 for L_wass+seg (Mℓ and Mh) and 0.068 and 0.037 for L_twce (Mℓ and Mh), with the statement that both losses outperform the baseline, but no baseline leaf-level F1 value is given. These numbers are near zero, which indicates that leaf-level segmentation is essentially not functioning on this task. Since the paper's central motivation is \"rich multi-class\" segmentation, the authors must report the baseline leaf-level F1, explain why leaf-level performance is so low, and reconcile the abstract's claim of strong support for hierarchy-weighted top-level supervision with the leaf-level results.","section":"§5.2, leaf-node results"},{"comment":"The hierarchy scaling parameter κ that defines the M_h ground-distance configuration is not validated on the target tasks. For WBP, the text states that the authors \"reuse the best configuration identified in (Wang et al., 2025a)\" and fix κ=10, but Wang et al. 2025a is an HSI paper, so this is a cross-task hyperparameter transfer without evidence. For HSI, κ=10 for L_wass+seg and κ=2 for L_twce were chosen in unreported preliminary experiments, while Section 5.5 later reports that L_wass+ce peaks at κ=50. Since κ controls the semantic ground distance that defines the proposed loss, the reported gains could be sensitive to favourable hyperparameter choices. A sensitivity analysis on the actual evaluation datasets, or a clearly described validation-based selection procedure, is needed to support the main results.","section":"§4.2.3, §5.5"},{"comment":"The abstract and Section 5.1 claim \"consistent improvements\" over baselines, but the paper's own paired statistics show that L_twce+seg is sometimes significantly worse than L_seg after Holm-Bonferroni correction. For example, in Table A1, the IXI-train / MB42-test comparison for NSD_small has MeanΔ=-0.0080 with Holm significance marked *, meaning the proposed loss is significantly worse on that metric. The claim should be restricted to the losses and metrics where the evidence is positive, and the text should acknowledge the negative paired comparisons. In addition, Table A1 reports only significance markers rather than numeric p-values, which is insufficient for readers to verify the Holm-Bonferroni correction; numeric p-values (or a supplementary table with them) should be provided.","section":"§5.1, Table A1"}],"minor_comments":[{"comment":"The aggregated ground-truth vector g† used in CE_T is never defined; the authors should define it analogously to p† in Eq. (4) so the reader can see how internal-node supervision is constructed from leaf labels.","section":"§3.2, Eq. (5)"},{"comment":"The notation is inconsistent: Section 4.2.3 and Table 2 use L_wass+seg, while Section 5.5 and Figure 7 refer to L_wass+ce, and the reported optimal κ values differ (κ=10 vs κ=50). These should be unified and the discrepancy explained.","section":"§4.2.3, §5.5"},{"comment":"The table rows labelled Lwass Mt and Lwass+seg Mt are easy to confuse; the caption should explicitly state that Lwass denotes the pure Wasserstein loss (Eq. 2) and Lwass+seg denotes the compound loss (Eq. 3), and should clarify which rows are baselines.","section":"Table 2 caption"},{"comment":"The sentence \"Because manual WBP annotation is often difficult due to single manual annotations may contain local inconsistencies\" is grammatically incomplete, and the following claim that GIF-derived masks \"can therefore be qualitatively more anatomically coherent\" is not supported by any evaluation in the manuscript.","section":"§4.1.1"},{"comment":"The DOI for Wang et al. 2025a appears malformed (\"10.1007/978-3-032-04984-1 56\"); please correct it.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The WBP evaluation is the strongest part of the paper, and the MB42 manual-reference results give real evidence for L_wass+seg. My main concern is that the cross-dataset AOMIC/IXI results are presented as if they supported anatomical accuracy when they only measure agreement with GIF pseudo-ground truth; this should be reframed or supplemented. The HSI leaf-level results need a baseline and discussion before the 'rich multi-class' claim can be accepted. The novelty relative to the authors' own prior MICCAI 2025 paper and OOD-SEG work should be clarified in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful, honest paper, but the novelty is thinner than the framing suggests and the strongest claims need qualification. The two losses are not new: the Wasserstein label-tree loss uses the Fidon et al. (2018) closed form, and tree-weighted CE is a standard hierarchical CE variant. What is new is the application to 108-class whole brain parcellation and sparse neurosurgical HSI, plus the hierarchy-level OOD threshold. That is a legitimate extension, and the paper works hard to evaluate it properly: paired tests on matched subjects, Holm-Bonferroni correction across 72 comparisons, an nnU-Net baseline, a Generalised Wasserstein Dice baseline, and an explicit statement of the GIF pseudo-ground-truth limitation in the methods.\n\nThe MB42 manual-reference results are the core evidence, and they are decent: L_wass+seg improves NSD and small-structure metrics over L_seg, and the paired tests hold after correction. I would not fight the claim that a hierarchy-aware Wasserstein loss helps high-cardinality WBP on small structures.\n\nThe soft spots are real but not fatal. First, AOMIC/IXI results are agreement with GIF, not accuracy against manual labels. The paper says so in the methods, but the abstract and conclusion say 'consistent improvements' without that caveat. The stress-test worry is legitimate: GIF's label-wise Dice is 78.6 on MB42, and the gains are concentrated in small structures where atlas-fusion noise is highest. Until someone shows GIF-agreement gains track manual-reference gains on the same subjects, those columns support only 'increases agreement with an automated tool.' Second, the HSI leaf-level F1 scores are around 0.07. That means the 107-subclass problem is essentially unsolved at leaf level, and the paper's 'both losses outperform baseline' claim for leaf nodes is not backed by a baseline number in the table. Third, kappa is chosen from the authors' prior paper or unpublished preliminary experiments; this is mild, but it means the reported configuration is not fully validation-blind. Fourth, they cite Graham et al., Deng et al., and Li et al. but do not compare against any of them. Fidon et al. is included, so this is not fatal, but it limits the claim that these losses are 'better' than existing hierarchy-aware methods.\n\nWho is this for? People working on high-cardinality segmentation or sparse positive-only medical segmentation. It deserves a serious referee; conditional accept after revision would be reasonable. The authors should either de-emphasize the AOMIC/IXI claims or add a matched-subject validity check, and they need to report leaf-level baselines for HSI.","headline":"Solid, honest empirical evaluation of two borrowed loss ideas in a new setting; the manual-reference WBP gains are credible, but the cross-dataset claims rest on pseudo-ground-truth and the HSI leaf-level numbers are too weak to ignore.","tokens_in":27678,"tokens_out":2552,"would_cite":true,"duration_ms":26941,"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 claims that encoding the semantic hierarchy of medical labels as a tree, and pricing each prediction error by the tree distance between true and predicted classes, consistently improves segmentation over standard per-pixel…","keywords":["semantic segmentation","label hierarchy","Wasserstein distance","tree-weighted cross-entropy","whole brain parcellation","hyperspectral imaging","sparse annotation","positive-only learning"],"falsifier":"Take a held-out subset of AOMIC or IXI and have a neuroanatomist manually label the 108 DKT classes; then compare $L_{\\mathrm{wass+seg}}$ against the CE+Dice baseline on those manual labels. If the NSD and small-structure gains disappear or reverse under manual reference, the paper's cross-dataset claim is falsified. Alternatively, train with the same loss but a scrambled or random label tree: if the gains persist, the hierarchy is not the responsible mechanism.","tokens_in":26692,"feed_emoji":"🧠","tokens_out":4578,"duration_ms":41082,"temperature":0.7,"pith_summary":"The paper proposes that a segmentation model should be told how wrong it is, not just that it is wrong. It encodes the labels' anatomic or surgical hierarchy as a tree, then trains with two losses that use that tree: a Wasserstein distance that prices each pixel's mistake by the path length between the true and predicted leaf classes, and a tree-weighted cross-entropy that supervises aggregated probabilities at every node of the hierarchy. On whole-brain parcellation across three MRI datasets, replacing the nnU-Net CE+Dice baseline with the Wasserstein compound loss produced consistent gains, with the strongest improvements on Normalised Surface Dice and on the smallest structures; the tree-weighted loss helped more in sparse hyperspectral surgical scene segmentation. If these results hold, hierarchies of labels are a near-free source of supervision for rich medical segmentation tasks.","feed_headline":"Label-tree Wasserstein loss lifts whole-brain parcellation","feed_subtitle":"Pricing errors by anatomy path length beats the CE+Dice baseline, especially on surface and small-structure metrics.","key_machinery":"The core objects are two tree-based loss functions. The first is the Wasserstein distance in label space, $W_M(p,g) = p^\\top M g$ when $g$ is a one-hot ground truth, with $M$ the tree-path-length ground distance between leaves; with $L_{\\mathrm{seg}}$ set to Dice+CE (WBP) or CE (HSI), the compound loss is $L_{\\mathrm{wass+seg}} = \\alpha W_M + \\beta L_{\\mathrm{seg}}$. The second is tree-weighted semantic cross-entropy, which aggregates leaf probabilities up the tree via $p^{\\dagger} = (I-A)^{-1} \\tilde{p}$ and computes a weighted CE over all nodes: $\\mathrm{CE}_{\\mathcal{T}}(p,g) = -\\sum_v w_v g^{\\dagger}_v \\log p^{\\dagger}_v$. Both are evaluated with an edge-weighting scheme $M_h$ that scales parent edge weights by a factor $\\kappa$, and both plug into the sparse positive-only framework by thresholding aggregated probabilities at a coarse hierarchy level instead of at the leaves.","core_discovery":"On its own terms, the paper's central claim is that semantic structure in the label space is a trainable signal that standard per-pixel losses throw away. For whole-brain parcellation with 108 classes under the DKT protocol, the compound Wasserstein loss $L_{\\mathrm{wass+seg}} = \\alpha W_M + \\beta L_{\\mathrm{seg}}$, where the ground distance $M_{l,l'}$ is the path length between leaves in a neuroanatomical label tree, outperforms the nnU-Net default CE+Dice on manual-reference (MB42) and pseudo-ground-truth (AOMIC, IXI) evaluation; after Holm-Bonferroni correction the gains are significant for NSD and small-structure metrics in nearly all train-test settings. In the sparse positive-only hyperspectral setting, supervising at coarse hierarchy levels and weighting hierarchical edges improves F1 over leaf-only CE training. The paper also claims these losses integrate into a positive-only OOD-detection framework so background is flagged at inference by thresholding coarse-level aggregated probabilities.","pith_inferences":["A testable extension the authors do not run: obtaining manual WBP annotations on a subset of AOMIC or IXI would distinguish genuine anatomical gains from improved matching to GIF's own errors, since GIF's label noise may share the same hierarchical inductive bias.","The $M_h$ edge-weighting scheme reduces the hierarchy to a single global scaling factor $\\kappa$; optimizing $\\kappa$ per subtree or per hierarchy level could push the small-structure gains further than the paper reports.","If the mechanism is truly the tree metric, then constructing hierarchies automatically, for example from label co-occurrence or atlas geometry, could extend these gains to tasks where no expert hierarchy exists, a direction the paper leaves implicit.","The coarse-level OOD thresholding implies that background detection can be made robust by choosing the granularity at which 'unknown' is defined, which could generalize the sparse-supervision framework to applications where the background itself is hierarchical."],"forward_implications":["If the WBP result is correct, model-based brain parcellation can improve boundary and small-structure accuracy without any architectural change, simply by swapping the loss for one that knows the DKT hierarchy.","Semantic losses make 'better mistakes': errors that stay within the same parent class cost less, so confusion matrices shift from anatomically meaningless to anatomically plausible errors.","The sparse HSI results imply that hierarchy-aware supervision can be extracted from cheap coarse labels even when fine labels are too costly, which could lower annotation burden for surgical scene understanding.","Because the losses are drop-in replacements, they can be combined with any segmentation backbone trained with CE/Dice-type objectives, including future nnU-Net variants."],"supporting_citations":[{"why":"Provides the nnU-Net framework and its default CE+Dice loss, which serve as the WBP baseline and the base architecture the proposed losses modify.","marker":"[Isensee et al., 2021]"},{"why":"Supplies the closed-form Wasserstein distance with crisp ground truth and the Generalized Wasserstein Dice loss used as a semantically-aware baseline and as the starting point for the proposed Wasserstein compound loss.","marker":"[Fidon et al., 2018]"},{"why":"Introduces the positive-only sparse supervision framework with OOD detection that the paper extends with tree-based losses and coarse-level thresholding.","marker":"[Wang et al., 2026]"},{"why":"Preliminary tree-based semantic losses from which the paper takes the compound-loss formulation, the M_h edge-weight configuration, and the alpha=beta=0.5 setting.","marker":"[Wang et al., 2025a]"},{"why":"Defines the Mindboggle101 dataset and the DKT protocol that fix the 108-class WBP label space and provide the manual references for MB42.","marker":"[Klein and Tourville, 2012]"},{"why":"The GIF algorithm generates the pseudo-ground-truth masks on AOMIC and IXI, which are the evaluation references for the cross-dataset WBP claims.","marker":"[Cardoso et al., 2015b]"},{"why":"Supplies the neuroanatomical label hierarchy that the paper manually adapts to the DKT label set, forming the tree structure used by both proposed losses.","marker":"[Graham et al., 2020]"}],"fun_headline_variants":["Label-tree Wasserstein loss beats CE+Dice for brain parcellation","Hierarchy-weighted top-level loss improves sparse HSI segmentation","Label-tree losses lift both brain MRI and sparse HSI tasks","Pricing errors by label-tree path length improves two segmentation tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole-brain-parcellation gains on AOMIC and IXI are measured against automated GIF pseudo-ground-truth masks, not manual annotations; if GIF's own errors share the hierarchy the losses encode, the reported gains could partly reflect fitting the tool rather than better anatomy.","fun_headline_variants_meta":{"raw":{"variants":["Label-tree Wasserstein loss beats CE+Dice for brain parcellation","Hierarchy-weighted top-level loss improves sparse HSI segmentation","Label-tree losses lift both brain MRI and sparse HSI tasks","Pricing errors by label-tree path length improves two segmentation tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001607,"raw_usage":{"total_tokens":6405,"prompt_tokens":953,"completion_tokens":5452,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":5380}},"tokens_in":569,"tokens_out":5452,"duration_ms":42086,"temperature":1.0,"reasoning_tokens":5380,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:23:41.604550+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out subset of AOMIC or IXI and have a neuroanatomist manually label the 108 DKT classes; then compare $L_{\\mathrm{wass+seg}}$ against the CE+Dice baseline on those manual labels. If the NSD and small-structure gains disappear or reverse under manual reference, the paper's cross-dataset claim is falsified. Alternatively, train with the same loss but a scrambled or random label tree: if the gains persist, the hierarchy is not the responsible mechanism.","supporting_citations":[{"cited_title":"C., Ekanayake, J., Kitchen, N., Ourselin, S., et al","cited_arxiv_id":null,"evidence_quote":"Supplies the closed-form Wasserstein distance with crisp ground truth and the Generalized Wasserstein Dice loss used as a semantically-aware baseline and as the starting point for the proposed Wasserstein compound loss."},{"cited_title":"S., Sudre, C","cited_arxiv_id":null,"evidence_quote":"Supplies the neuroanatomical label hierarchy that the paper manually adapts to the DKT label set, forming the tree structure used by both proposed losses."}],"review_version":2}