{"id":"5a1d8908-efe8-40b0-aa4e-b31afe2c155e","arxiv_id":"2506.11134","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A context-aware critical pixel mask loss improves topology preservation and gap closing for 2D and 3D segmentation.","lead":"This paper introduces ContextLoss, a training loss that helps image segmentation networks keep thin structures such as blood vessels and road networks connected. It also adds two new metrics and a 3D bone-imaging dataset, and reports closing more missed connections than previous topology-aware losses.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main worry is evaluation fairness: CLoss is reported at fixed γ while compound clDice is selected by best clDice, and the headline 44% uses the authors' own e0-Gt on one dataset; standard e0 is not consistently better, so the context-mask claim is not yet solidly established.","rationale":"CLoss is a clearly specified, fast loss and the CLD dataset is a useful contribution; the comparison to compound clDice is the right design and the CLD ablation (Tab. 4) gives internal support. My concern is not that the claim is false, but that the main cross-dataset evidence is ambiguous: the baseline is optimized for a metric on which CLoss is not reported, and the metrics that carry the conclusion are proposed by the same paper and may reward the thickness increase the authors observe. The 44% figure is an extreme point (γ=0.2, HRF-Retina, e0-Gt) rather than a robust summary. A matched selection and significance check, ideally with standard e0/e1 as well as the proposed metrics, would either confirm the mask advantage or show it shrinks to noise. Since the reader already conditioned on a fairer evaluation, the verdict stays CONDITIONAL.","tokens_in":15447,"tokens_out":7529,"duration_ms":78217,"concrete_test":"Run all methods (at least CLoss, clDice, compound clDice) with hyperparameter selection on a validation split using the same criterion, e.g. highest e0-Gt, and also a separate run selecting by highest clDice. Report the selected γ per fold, mean±std over the five folds, and paired significance for e0, e1, clDice, e0-Gt, AGS. Then recompute the '44% more missed connections' claim using standard e0 instead of e0-Gt and with matched γ; if CLoss is not significantly better on standard e0 while showing the same Dice drop, the central mask-superiority claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 4.1 concludes that CLoss's context mask is superior to the full-skeleton mask of compound clDice, but the supporting table does not enforce a matched comparison. In Tab. 1, compound clDice's γ is chosen as 'the result with the highest clDice metric' (Sec. A.6), whereas CLoss is shown at fixed γ∈{0.08,0.1,0.2} with no stated selection rule. On HRF-Retina the headline 44% gap-closing result is e0-Gt=1.378 at γ=0.2 versus clDice=2.475, yet at γ=0.08 CLoss's e0-Gt is 2.347, only 5% better than compound clDice's 2.429. The same γ=0.2 also lowers Dice from 82.33 to 80.57 and leaves standard e0 at 0.172, worse than compound clDice's 0.155. Across datasets CLoss is not consistently better on standard topology metrics: on Vessap its clDice score at γ=0.2 is 93.13 versus compound clDice's 95.80, and on Roads its e0 at γ=0.2 is 0.205 versus 0.199. Thus the design claim rests mainly on the two self-proposed metrics, e0-Gt and AGS, both of which reward coverage of the GT skeleton and can benefit from the reported thickening of predictions. No per-fold error bars or paired tests are given, so the differences, some of which are small, could be within cross-validation noise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ContextLoss (CLoss), a topology-preserving segmentation loss whose critical pixel mask includes the context around topological errors (missed and false-positive connections), obtained by combining skeletonization with distance transforms. The authors also introduce two new metrics (e0-Gt and AGS) intended to measure missed-connection reparation, and present a new 3D bone cement-line dataset (CLD). CLoss is evaluated on Roads, HRF-Retina, Vessap, and CLD against nnU-Net, Dice+CE, clDice, and a newly introduced compound clDice that shares CLoss's pixel-wise loss. The central design claim is that the context mask, not the pixel-wise loss, drives the reported topological improvements.","tokens_in":15824,"tokens_out":3829,"duration_ms":38802,"significance":"The context-mask idea is a reasonable and potentially useful contribution to topology-preserving segmentation, and the new CLD dataset plus the two proposed metrics will likely interest the community. The paper ships public code, uses a standard nnU-Net protocol, and includes an ablation (Table 4) that isolates the critical pixel mask, which are strengths. However, the current evidence for the central claim is weakened by an asymmetric hyperparameter-selection protocol, by reliance on the authors' own metrics for the headline result, and by the absence of variance or significance information. If the evaluation is made symmetric and statistically grounded, the contribution would be solid; in its present form the main design claim is not yet established beyond reasonable doubt.","major_comments":[{"comment":"The comparison between CLoss and compound clDice is asymmetric. Compound clDice is reported after a weight search over 0.1 to 0.5 in 0.1 steps, selecting the result with the highest clDice score, while CLoss is reported at fixed gamma values 0.08, 0.1, and 0.2 with no stated selection rule. On HRF-Retina the headline e0-Gt improvement (CLoss 1.378 at gamma=0.2 versus compound clDice 2.429) shrinks to about 5% at gamma=0.08 (2.347 versus 2.429), and at gamma=0.2 CLoss has a worse standard e0 (0.172 versus 0.155) and a lower Dice (80.57 versus 82.33). On Vessap the clDice score at gamma=0.2 is 93.13 versus 95.80 for compound clDice, and on Roads e0 at gamma=0.2 is 0.205 versus 0.199. The conclusion in Sec. 4.1 that CLoss's mask is superior to the full-skeleton mask therefore depends on which gamma is chosen and which metric is used to select it. The authors should report CLoss under the same selection rule as compound clDice, or compare all methods over a shared gamma grid without per-method selection.","section":"Sec. A.6 and Table 1"},{"comment":"The paper reports five-fold cross-validation for all methods but gives no per-fold variances, confidence intervals, or paired statistical tests. Several reported differences are small enough to be plausibly within cross-validation noise (for example, Roads e0 0.185 for CLoss gamma=0.08 versus 0.195 for compound clDice; CLD clDice 86.22 for CLoss gamma=0.08 versus 86.22 for clDice). Without a measure of variability, the claims of consistent superiority in Sec. 4.1 and the conclusion are not verifiable. At minimum, per-fold metric distributions or standard deviations should be reported, and a paired test (e.g., Wilcoxon signed-rank over folds) should be provided for the main comparisons.","section":"Sec. 4 and Sec. A.5"},{"comment":"The headline claim of repairing up to 44% more missed connections rests on e0-Gt, one of the two metrics proposed by the authors, evaluated on a single dataset (HRF-Retina). The standard topology metrics do not consistently support the claim: on HRF-Retina CLoss at gamma=0.2 has e0=0.172, which is worse than compound clDice's 0.155, and on Vessap CLoss at gamma=0.2 has a clDice score 93.13 versus 95.80 for compound clDice. Since e0-Gt and AGS both measure accuracy on or relative to the ground-truth skeleton, they can be improved by thicker predictions, and the reported Dice drop at higher gamma is consistent with such thickening. The paper should either present the 44% claim alongside the standard e0, e1, and clDice results at the same operating point, or explicitly justify why the self-proposed metrics are the primary evidence for the claim; the current presentation overstates the strength of the evidence.","section":"Sec. 2.2 and Sec. 4.1"},{"comment":"The sentence 'CLoss consistently outperforms compound clDice on the topology metrics' is not supported by Table 1. On Vessap the clDice score is worse for CLoss at gamma=0.2 (93.13 versus 95.80) and on Roads e0 at gamma=0.2 is worse (0.205 versus 0.199). The discussion in Sec. A.9 acknowledges the Vessap clDice discrepancy by attributing it to skeletonization artifacts, but this attribution is not empirically demonstrated. The claim of consistency should be qualified to specific metrics, datasets, and gamma values, or the conclusion should be rewritten to match the actual table.","section":"Sec. 4.1"}],"minor_comments":[{"comment":"The phrase 'recently proposed solutions are loss functions based on critical pixel masks that consider the whole skeleton of the segmented structures in the critical pixel mask' is hard to parse; consider simplifying to 'recent loss functions use critical pixel masks that include the entire skeleton of the segmented structures'.","section":"Abstract"},{"comment":"The notation beta_0^{Xbin⊙Y} is not defined before use in Equation (3); please state that it denotes the number of connected components of the binary prediction masked by the ground truth.","section":"Sec. 2.2, Eq. (3)"},{"comment":"The definition of AGS as sum(Xbin ⊙ S_Y)/sum(S_Y) is clear, but the text should explicitly note that AGS is not a skeleton-based metric and therefore is expected to be insensitive to prediction-skeleton artifacts; this is currently only implied by Figure 4.","section":"Sec. 2.2, Eq. (4)"},{"comment":"The new CLD dataset is a significant asset, but the main text gives only its shape and voxel size. A brief statement of the inter-annotator agreement or annotation QA procedure appears only in the appendix; consider moving a one-sentence version into the main text.","section":"Sec. 3"},{"comment":"The sentence 'CLoss repairs 44% more missed connections than clDice (e0-Gt)' should state the gamma value and the comparison method (clDice vs compound clDice) in the same sentence, since Table 1 shows different values for these two baselines.","section":"Sec. 4.1"},{"comment":"The pseudo-code for the critical pixel mask uses the soft-skeleton notation S_X and S_Y, but the text of Sec. 2.1 describes the mask on binary predictions; please clarify how the soft skeleton is binarized or used in the distance-transform formulas.","section":"Sec. A.1, Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be of interest to the topology-preserving segmentation community, and the new dataset and code release are genuine contributions. The main technical concern is evaluative rather than conceptual: the comparison against compound clDice must be made symmetric, and the statistical and metric-selection limitations must be addressed before the context-mask claim is accepted. I would not recommend rejection, but the current evidence is not yet at the level required to support the paper's strongest statements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes CLoss, a topology-preserving loss that builds a critical pixel mask by combining skeletonization with a distance transform, thereby including context around topological errors rather than just skeleton pixels. That is a real, modest new idea, and the comparison against compound clDice (which shares the same pixel-wise loss and uses the full-skeleton mask) is the right way to isolate the mask's contribution. The new 3D bone cement line dataset (CLD) and the two gap-focused metrics (e0-Gt and AGS) are useful additions, and the code is released. The Thin CLoss ablation also supports the claim that context, not just the pixel-wise loss, matters.\n\nWhere it gets shaky is the evaluation protocol. Compound clDice's weight is selected as the one maximizing clDice (Supp. A.6), while CLoss is reported at three fixed gammas with no explicit selection rule. That asymmetry matters. On HRF-Retina, the headline 44% repair figure uses e0-Gt at gamma=0.2; at gamma=0.08 the margin over compound clDice shrinks to about 5%. On standard metrics, CLoss is not consistently better: on Vessap its best clDice is 93.13 versus compound's 95.80; on Roads e0 at gamma=0.2 is 0.205 versus 0.199; on HRF e0 at gamma=0.2 is 0.172 versus 0.155. No error bars or paired significance tests are reported despite five-fold CV. The authors do honestly discuss the thickness/artifact trade-offs, and the proposed metrics are reasonable, but using self-proposed metrics for the headline is more persuasive when standard metrics move in the same direction, which they do not always.\n\nOverall, the central design claim is plausible but not firmly established. The paper is clearly written, the method is easy to adopt, and the new dataset is likely to be useful to others. I would send this to peer review rather than desk reject, but I would ask the authors to run a matched hyperparameter selection (same criterion for both methods, or full sweeps for both), report per-fold variability, and clarify which gamma they recommend and why. With that, the paper could be solid.","headline":"ContextLoss is a genuinely useful incremental loss with a new dataset and metrics, but the evaluation protocol is asymmetric and lacks error bars, so the headline claim about the context mask isn't yet solid.","tokens_in":16359,"tokens_out":6673,"would_cite":true,"duration_ms":65100,"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":"ContextLoss improves topological correctness by putting the whole context of each missed connection into the loss, repairing more gaps than skeleton-only masks.","keywords":["topology-preserving segmentation","loss function","critical pixel mask","skeletonization","distance transform","missed connections","e0-Gt","AGS"],"falsifier":"Run multiple training seeds for CLoss and compound clDice on HRF-Retina with all hyperparameters selected by the same criterion, for instance best e0-Gt or best combined Betti error, and compute e0-Gt and AGS with confidence intervals; if CLoss no longer closes more missed connections than compound clDice, the paper's attribution of the improvement to the context mask is refuted.","tokens_in":15262,"feed_emoji":"🩻","tokens_out":5000,"duration_ms":47837,"temperature":0.7,"pith_summary":"The paper proposes ContextLoss (CLoss), a training loss that helps segmentation networks repair topological errors such as broken vessels, roads, or bone cement lines. Its central claim is that the critical pixel mask should include the entire context of a missed or false connection, not just single critical pixels or the whole skeleton. To verify improved gap closing, the authors introduce two metrics, e0-Gt and AGS, that isolate connectivity and resist evaluation artifacts. On four datasets, including a new 3D nano-imaging bone cement line dataset, CLoss improves topology-aware metrics and closes up to 44% more missed connections than clDice on the HRF-Retina dataset.","feed_headline":"Context mask closes 44% more missed connections","feed_subtitle":"A loss that weights the full context of broken connections improves topology in vessel, road, and bone-line segmentation.","key_machinery":"The load-bearing mechanism is the context-extended critical pixel mask. Skeletonization identifies topologically critical locations, and a distance transform extends the error part of the skeleton to all nearby pixels belonging to the label (for gaps) or prediction (for false positives), so the loss weighs the whole neighborhood of each topological error. Two complementary metrics, $e_0$-$Gt$ and AGS, are also part of the machinery: $e_0$-$Gt$ masks predictions with the ground truth before counting connected components, and AGS measures prediction accuracy on the ground-truth skeleton, making both less sensitive to cropping and contour artifacts than standard $e_0$ and clDice.","core_discovery":"The central claim is that context around topological errors, extracted by combining skeletonization with distance transforms, is what makes a topology-preserving loss effective. CLoss splits the label skeleton into the part at the topological error and the correctly predicted part, then uses the distance transform to keep only pixels closest to the error skeleton, cropped to label size; false-positive connections are handled symmetrically. This context-rich critical pixel mask $M$ enters the loss $L_{\\text{context}} = (1-\\gamma)L_{\\text{pixel}} + \\gamma L_{\\text{pixel}} \\odot M$. The paper argues that this mask, not the choice of pixel-wise loss, drives the improvement, supported by comparisons against compound clDice, which differs only in the mask. Training is staged: pretrain with a standard pixel-wise loss, fine-tune with CLoss, and post-process by keeping only fine-tuned foreground components that overlap pretraining predictions.","pith_inferences":["The context-mask design is independent of the pixel-wise loss and skeletonization routine, so it could be combined with adaptive or learned skeleton extraction to handle noisy 3D surfaces.","The reported comparison assumes a particular selection rule for baselines and fixed gamma values for CLoss; readers should check whether CLoss retains its margin when all methods, including CLoss, are tuned by the same criterion.","The proposed metrics, particularly e0-Gt, could serve as inexpensive connectivity checks for any segmentation pipeline, not only for training.","The thickening effect of CLoss suggests that combining it with a boundary- or thickness-aware term might recover Dice while keeping the topology gains."],"forward_implications":["Training with CLoss closes more gaps in elongated structures such as vessels, roads, and bone cement lines than training with clDice or Dice & Cross-Entropy.","CLoss works on 2D and 3D data and can be plugged into any segmentation network because it changes only the loss function.","The proposed topological post-processing removes small separate noise components without introducing new false-positive connections.","The metrics e0-Gt and AGS give a less artifact-prone reading of connectivity than e0 and clDice, which may change how topology-aware methods are compared.","CLoss runtime per epoch is comparable to clDice, so the topological gains do not require the high runtime of persistent-homology-based masks."],"supporting_citations":[{"why":"Defines clDice, the main baseline loss, and supplies the soft-skeletonization routine used throughout CLoss.","marker":"[3]"},{"why":"Provides the nnU-Net framework and training protocol used for all experiments.","marker":"[10]"},{"why":"Represents the prior critical-pixel approach that masks only singular critical pixels, the gap CLoss aims to address.","marker":"[6]"},{"why":"Supports the combined Dice and Cross-Entropy pixel-wise loss used inside CLoss.","marker":"[1]"},{"why":"Supports the claim that including more pixels in the critical mask helps network learning.","marker":"[9]"},{"why":"Motivates the need for evaluation metrics robust to artifacts in topology-aware segmentation.","marker":"[12]"},{"why":"Provides the runtime comparison showing that persistent-homology-based critical masks are much slower than skeleton-based ones.","marker":"[13]"}],"fun_headline_variants":["ContextLoss fixes 44% more broken vessel and road links","Topology loss with context repairs 44% more missed connections","Context mask boosts topology preservation by 44%","CLoss: context-aware loss for topological segmentation","ContextLoss: 44% better at reconnecting segmented structures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation protocol is assumed to be reliable: baselines are tuned by their best clDice score while CLoss is reported at fixed gamma values, no error bars are given, and the headline 44% improvement is measured with the authors' own e0-Gt metric on a single dataset.","fun_headline_variants_meta":{"raw":{"variants":["ContextLoss fixes 44% more broken vessel and road links","Topology loss with context repairs 44% more missed connections","Context mask boosts topology preservation by 44%","CLoss: context-aware loss for topological segmentation","ContextLoss: 44% better at reconnecting segmented structures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000532,"raw_usage":{"total_tokens":2532,"prompt_tokens":889,"completion_tokens":1643,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":1563}},"tokens_in":505,"tokens_out":1643,"duration_ms":11128,"temperature":1.0,"reasoning_tokens":1563,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:55:26.032629+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run multiple training seeds for CLoss and compound clDice on HRF-Retina with all hyperparameters selected by the same criterion, for instance best e0-Gt or best combined Betti error, and compute e0-Gt and AGS with confidence intervals; if CLoss no longer closes more missed connections than compound clDice, the paper's attribution of the improvement to the context mask is refuted.","supporting_citations":[{"cited_title":"4: Evaluation artifacts","cited_arxiv_id":null,"evidence_quote":"Defines clDice, the main baseline loss, and supplies the soft-skeletonization routine used throughout CLoss."},{"cited_title":"clDice - a Novel Topology-Preserving Loss Function for Tubular Structure Segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the nnU-Net framework and training protocol used for all experiments."},{"cited_title":"We apply our proposed topological post-processing (Sec","cited_arxiv_id":null,"evidence_quote":"Represents the prior critical-pixel approach that masks only singular critical pixels, the gap CLoss aims to address."},{"cited_title":"ContextLoss: Context Information for Topology-Preserving Segmentation","cited_arxiv_id":"2506.11134","evidence_quote":"Supports the combined Dice and Cross-Entropy pixel-wise loss used inside CLoss."},{"cited_title":"Do we really need dice? the hidden region-size biases of segmentation losses,","cited_arxiv_id":null,"evidence_quote":"Supports the claim that including more pixels in the critical mask helps network learning."},{"cited_title":"Skeleton recall loss for connectivity conserving and resource efficient seg- mentation of thin tubular structures,","cited_arxiv_id":null,"evidence_quote":"Motivates the need for evaluation metrics robust to artifacts in topology-aware segmentation."},{"cited_title":"Topology- preserving deep image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the runtime comparison showing that persistent-homology-based critical masks are much slower than skeleton-based ones."}],"review_version":1}