{"id":"b81d6d55-4622-4f74-81fc-9df32f348867","arxiv_id":"2509.03154","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Two connectivity-aware losses reduce instance discontinuities and improve AIS length distribution estimates in 3D segmentation, with the best Wasserstein distance dropping from 55.5 to 42.9.","lead":"The paper proposes two loss functions that penalize segmentation models when they split long, thin structures into disconnected pieces. On 3D light-sheet images of mouse axon initial segments, these losses reduce breaks and improve length distribution estimates compared with standard and existing topology-aware losses.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Direct discontinuity metric contradicts headline: Negative Centerline worsens overlapping-instances (1.042 vs 1.038 baseline) in the main 3x-downscaled run; precision/Wasserstein gains may reflect false-positive reduction, not fewer splits.","rationale":"The reader's conditional verdict already captures the general lack of statistical robustness and proxy concerns. My stress test identifies a more specific, internally verifiable inconsistency: the paper's own direct metric for discontinuity (overlapping instances) does not improve for the headline method in the main experiment. This strengthens the case for conditional acceptance but does not warrant rejection, since the proposed losses are clearly specified, code is released, and some secondary metrics (small-dataset Wasserstein) show large improvements. However, the central claim as stated in the abstract needs revision or additional evidence. Thus the verdict remains CONDITIONAL, unchanged from the reader.","tokens_in":11670,"tokens_out":2288,"duration_ms":27757,"concrete_test":"Report the per-fold values of the 'overlapping instances' metric for the 3x-downscaled runs (baseline vs Negative Centerline Loss) and run a paired significance test (e.g., Wilcoxon signed-rank over the 5 folds). If the metric is not significantly less than 1.038 (i.e., p<0.05 and negative direction), the headline claim of reduced segmentation discontinuities per instance is unsupported by the direct measure. Additionally, decompose precision gains into split-driven vs false-positive-driven components by counting matched/unmatched instances.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the proposed losses 'reduce the number of segmentation discontinuities per instance.' The direct metric for this claim is 'overlapping instances' (Section 2.4), defined as the average number of prediction instances each label instance overlaps with, where 1.0 is ideal. In the main 3x-downscaled experiment (Table 2), Negative Centerline Loss—the method touted as best—achieves 1.042, worse than the baseline nnU-Net's 1.038. This is a direct contradiction of the headline claim. The reported improvements in Wasserstein distance and precision are indirect: precision can improve from fewer false-positive instances (instances overlapping no label) rather than fewer splits, and Wasserstein distance measures length distributions, which can shift for reasons unrelated to within-instance continuity. Moreover, no error bars or significance tests are reported for any metric; the 5-fold means could obscure large fold-to-fold variance. Since the central claim is about discontinuity reduction, and the one metric that directly measures it fails to improve for the primary method in the flagship experiment, the evidence does not currently support the abstract's strongest assertion. The reader flagged the length proxy and label reliability as the weakest assumptions; those are relevant, but the internal metric contradiction is more immediately damaging to the stated claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses a practical problem in biomedical segmentation of elongated structures: preserving instance continuity and length when signal dropout causes disconnected predictions. It proposes two new loss functions, Negative Centerline Loss (penalizing uncovered label centerline) and Simplified Topology Loss (reapplying BCE in regions causing discontinuities), and evaluates them on a 3D light-sheet fluorescence expansion microscopy dataset of axon initial segments using an nnU-Net architecture. Comparisons are made against a baseline nnU-Net, clDice, and a persistence-diagram topology loss, under both original and 3x-downscaled conditions, plus a small-dataset setting. The paper reports improved Wasserstein distance of length distributions and precision, and claims the losses reduce segmentation discontinuities per instance. The source code is released as an nnU-Net fork.","tokens_in":12021,"tokens_out":3666,"duration_ms":47845,"significance":"If the central claim were established, the paper would make a useful, computationally modest contribution: two simple losses with clear algorithmic descriptions, an open implementation, and a real biological dataset where continuity matters. The paper also raises an interesting subsidiary point about downscaling improving connectivity. However, the evidence as presented is not yet convincing for the headline claim. The direct metric for per-instance discontinuity does not improve for the main method, no uncertainty quantification is provided, and important hyperparameters and experimental choices are tuned on the validation signal. The length-proxy and label-quality issues further complicate the downstream interpretation. The work is potentially valuable, but it needs a substantial revision to align claims with evidence.","major_comments":[{"comment":"The abstract claims the losses 'reduce the number of segmentation discontinuities per instance.' The metric that most directly measures this is 'overlapping instances' (average number of prediction instances per label instance, ideally 1). In the main 3x-downscaled run, Negative Centerline Loss gives 1.042, which is worse than the baseline's 1.038; Simplified Topology Loss is 1.054. The reported precision and Wasserstein improvements are indirect: precision can improve by eliminating false-positive instances rather than fewer splits, and Wasserstein distance concerns the aggregate length distribution, which can shift for reasons other than within-instance continuity. A direct per-instance split count, or a re-framing of the claim away from 'fewer discontinuities,' is needed.","section":"Abstract and Table 2, Section 2.4"},{"comment":"The auxiliary loss weights w_eval are selected 'empirically to make w_eval as large as possible while still maintaining good segmentation quality improvement as measured by the Dice validation metric during the first epochs.' This means each method is compared after tuning its weighting on the validation signal. No sensitivity analysis or fixed-weight protocol is reported, so part of the observed Wasserstein/precision gain could reflect this tuning rather than the loss design itself. The authors should report results for a range of weights or justify that the chosen weights are not the source of the differences.","section":"Section 2.3 and Table 1"},{"comment":"The length-based claims rest on a proxy: skeleton graph diameter of each instance after discarding border-intersecting instances. The paper acknowledges there is no standardized AIS length measurement and does not validate this proxy against manual tracing or an independent length estimate. In addition, Figure 3's caption states that in some regions models 'outperform the ground truth label' because annotators missed dim AIS. Since precision, recall, and length metrics all use these labels as ground truth, and since Negative Centerline Loss is defined against the label centerline, incomplete labels can penalize correct bridging behavior. The authors should quantify label uncertainty or demonstrate robustness to it.","section":"Section 2.5 and Figure 3"},{"comment":"All performance metrics are reported as 5-fold means without error bars, confidence intervals, or significance tests. Several comparisons, e.g., overlapping instances 1.038 vs. 1.042, are small relative to likely fold-to-fold variance. The claim that one loss is preferable to another requires at least per-fold results or a paired statistical test. This is especially important because the experimental setup fixes many choices (downscaling, spacing correction, loss weights) after exploratory analysis on the same data.","section":"General experimental reporting (all tables)"}],"minor_comments":[{"comment":"The 'Strictly standardized mean difference' (SSMD) metric is never defined or referenced. Please provide its formula and explain how it is computed from the length distributions.","section":"Section 2.4"},{"comment":"The length computation uses voxel counts with z-stack spacing considered, but it is not stated whether the skeletonization is performed on downsampled-then-upsampled masks or at original resolution. Clarify the exact pipeline used for the downscaled experiments.","section":"Section 2.5"},{"comment":"The soft-skeleton routine is a key component, but the pooling window sizes and the stopping criterion are not fully specified. The text says 'until convergence,' while the pseudocode loops while |I| > 0; please clarify how this behaves for 3D inputs and how it differs from clDice's fixed-iteration variant.","section":"Section 2.1 / Algorithm 2"},{"comment":"The AISuite URL is given in the text but not in the references. Also, the reference formatting for [3] is incomplete ('Neurophotonics, 6:1').","section":"Introduction, Section 1.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is not fundamentally unsound; the algorithms are clearly described and the code is released. However, the abstract's strongest claim is directly contradicted by the overlapping-instances metric in Table 2, and the comparative conclusions are weakened by validation-based weight selection and missing uncertainty quantification. These are fixable with additional analysis, re-framing, or both, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the two loss functions are clearly specified and shipped as code, and the paper is worth a referee's time, but the headline claim about reducing discontinuities is not supported by the paper's own direct metric in the main experiment. The 'overlapping instances' metric for Negative Centerline Loss (1.042) is worse than baseline (1.038) in the 3x-downscaled run. That metric directly measures per-instance splits, so the abstract overstates what the evidence shows.\n\nWhat is actually new: Negative Centerline Loss is a one-sided skeleton-recall term, distinct from clDice and CP-loss, and Simplified Topology Loss reimplements the Topograph idea with connected-component analysis. Both are simple, properly described algorithms (Algorithms 1 and 3), and the authors release the nnU-Net fork with clDice and topology baselines. The discussion of downscaling and spacing correction is a useful subsidiary result. The comparison against clDice and a bottleneck-distance topology loss is the right baseline set.\n\nSoft spots in proportion: the central empirical claim is shaky. No error bars or significance tests anywhere, so 5-fold means could hide large variance. The overlapping-instances point estimate for the main run contradicts the claim; even if noise is responsible, the paper needed to address that. Loss weights were hand-tuned on validation performance, which inflates the reported gains. The dataset is private, so the numbers are not independently reproducible. The length computation, graph diameter of skeletonized instances, is a plausible but unvalidated proxy; the authors acknowledge there is no standard. The label quality issue (annotators missing dim AIS) is acknowledged in a figure caption but not factored into the analysis. The reader's concern about the length proxy is fair, but the internal metric contradiction is more immediately damaging.\n\nThe methodological idea is sound enough to be worth a referee. If the authors add error bars, report all four metric tables with variance, and either remove or substantially soften the discontinuity claim, this could be a useful contribution. As is, it is a conditional pass: the losses are worth considering, but the evidence for the continuity claim is not there.\n\nRecommendation: send to peer review, but expect careful scrutiny of the overlapping-instances metric and the lack of uncertainty quantification.","headline":"Clearly specified loss variants and honest baselines, but the paper's own overlapping-instances metric contradicts its headline claim in the main experiment; needs error bars and a revised claim before the evidence is convincing.","tokens_in":12447,"tokens_out":3960,"would_cite":false,"duration_ms":39671,"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 two connectivity-aware loss functions reduce segmentation discontinuities in elongated biological structures and thereby improve downstream length estimation.","keywords":["axon initial segment","connectivity-aware loss","Negative Centerline Loss","Simplified Topology Loss","tubular structure segmentation","instance continuity","light-sheet fluorescence microscopy","length distribution"],"falsifier":"Segment synthetic tubular objects with known true lengths and controlled signal dropout using the proposed losses, then compare the graph-diameter length estimates against the known lengths; if the Wasserstein-distance improvement disappears or reverses when true length is known, the downstream-length claim is an artifact of the proxy. As a second check, re-measure real AIS length by manually tracing from soma to tip on the same instances and see whether the ranking of losses survives.","tokens_in":11605,"feed_emoji":"🧠","tokens_out":7631,"duration_ms":77670,"temperature":0.7,"pith_summary":"Segmentation models optimized for voxel-wise accuracy tend to split elongated biological structures wherever the input signal fades, and every split corrupts downstream measurements such as axon length. This paper proposes two loss functions that make continuity itself the optimization target: Negative Centerline Loss penalizes the fraction of the label's centerline that the prediction fails to cover, and Simplified Topology Loss re-applies binary cross-entropy inside the specific regions that would connect or disconnect instances. On 3D light-sheet fluorescence microscopy of axon initial segments, both losses reduce the number of prediction fragments per labeled instance and improve the Wasserstein distance between predicted and labeled length distributions, with the centerline loss best in the full-data setting and the topology loss best when training data are scarce. The paper additionally reports that downscaling the in-plane resolution by three improves connectivity, a result that runs against the usual assumption that higher resolution is better. The central claim is that structural priors embedded in the loss, rather than more parameters or more data, are the effective lever for continuity in elongated-structure segmentation.","feed_headline":"Centerline-aware loss cuts axon length error by 23%","feed_subtitle":"By penalizing gaps in the centerline, the loss reduces split instances and improves downstream length distributions.","key_machinery":"Negative Centerline Loss is a differentiable loss whose support is the label's soft skeleton; it computes |(1−P)∘L_CL|/|L_CL|, so the gradient is nonzero only on the centerline and largest precisely at gaps in the prediction. Simplified Topology Loss is a non-differentiable region-selection procedure—threshold, dilation, connected-component labeling—that finds label regions bordering at least two distinct prediction regions, as well as prediction regions overlapping no label, and then applies binary cross-entropy only inside those regions. The skeletonization is the soft-skeleton pooling routine from clDice run to convergence, with O(nd) worst-case time; the region finder runs in amortized n","core_discovery":"On the paper's own terms, the central discovery is that two simple losses can make a U-Net segment elongated structures as connected instances rather than as fragments, and that this translates into better length statistics downstream. Negative Centerline Loss computes the soft skeleton of the label and measures the fraction of that skeleton not intersected by the prediction; its gradient is therefore concentrated exactly where a gap breaks the instance. Simplified Topology Loss binarizes the prediction, dilates it, labels connected components, and identifies label regions that touch at least two distinct prediction regions—the bridges that, if added, would join fragments—then retrains the n","pith_inferences":["A natural extension the authors leave implicit: Simplified Topology Loss's region finder already identifies exactly where the model is fragmented, so it could double as an active-learning acquisition function that tells annotators which locations to correct.","The downscaling result hints that continuity errors in 3D patch-based training are driven by limited context or receptive field rather than by resolution; an explicit test would be to train at full resolution with larger patches or dilated convolutions and see whether the gap to downscaled runs closes.","The same 'connectivity as the objective' logic could apply across time rather than space—e.g., video instance segmentation under occlusion—where the analogue of signal dropout is a missing frame and the analogue of length is track duration.","A direct validation of the length proxy—comparing graph diameter against manually traced soma-to-tip paths on the same instances—would show whether the reported length improvements reflect true biological length or an artifact of the measurement."],"forward_implications":["If the central claim holds, segmentation models trained with these losses produce fewer split instances per labeled AIS, so automated length distributions match manual annotations more closely than baseline training does.","In the full-data 3x-downscaled setting, Negative Centerline Loss lowers the Wasserstein distance of length distributions from 55.5 to 42.9 and raises precision from 0.822 to 0.862 without meaningfully changing Dice.","Simplified Topology Loss retains most of the continuity benefit and gives the best results in the halved-dataset run (Wasserstein 167.4 vs 244.3 for the baseline), making it the recommended choice when labels are scarce.","Both losses add roughly 13% to per-epoch training time relative to the baseline, so the continuity gain comes at modest compute cost.","Because the losses are agnostic to imaging modality and structure shape, the authors expect them to transfer to other elongated-structure tasks, such as vascular or road-network segmentation, where continuity is important."],"supporting_citations":[{"why":"Supplies the soft-skeleton subroutine used by Negative Centerline Loss and is the main topology-aware baseline it is compared against.","marker":"[36]"},{"why":"Connectivity-preserving skeleton-alignment loss that motivates penalizing gaps on the skeleton rather than the full mask.","marker":"[42]"},{"why":"Non-differentiable graph-based topology-preserving method whose idea of re-applying loss in topology-critical regions inspired Simplified Topology Loss.","marker":"[26]"},{"why":"Provides the self-configuring U-Net pipeline used to train and evaluate every loss function.","marker":"[14]"},{"why":"Defines the base U-Net architecture shared by all experiments.","marker":"[33]"},{"why":"Documents the light-sheet fluorescence expansion microscopy protocol that produces the AIS volumes and motivates the high-throughput setting.","marker":"[3]"}],"fun_headline_variants":["Centerline loss reduces axon segmentation gaps","Two losses keep neural structures connected in 3D","Penalize skeleton gaps to improve length accuracy","Connectivity-aware losses fix split axon instances","Simplify topology to preserve instance continuity"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The length-based evaluation assumes that the longest path through a skeletonization of the segmentation equals the object's true AIS length, and that the manual labels are complete enough to define the target centerline; if either assumption fails, the reported length improvements may not reflect true biological length.","fun_headline_variants_meta":{"raw":{"variants":["Centerline loss reduces axon segmentation gaps","Two losses keep neural structures connected in 3D","Penalize skeleton gaps to improve length accuracy","Connectivity-aware losses fix split axon instances","Simplify topology to preserve instance continuity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000148,"raw_usage":{"total_tokens":987,"prompt_tokens":664,"completion_tokens":323,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":408,"completion_tokens_details":{"reasoning_tokens":257}},"tokens_in":408,"tokens_out":323,"duration_ms":3956,"temperature":1.0,"reasoning_tokens":257,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:06:04.334629+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Segment synthetic tubular objects with known true lengths and controlled signal dropout using the proposed losses, then compare the graph-diameter length estimates against the known lengths; if the Wasserstein-distance improvement disappears or reverses when true length is known, the downstream-length claim is an artifact of the proxy. As a second check, re-measure real AIS length by manually tracing from soma to tip on the same instances and see whether the ranking of losses survives.","supporting_citations":[{"cited_title":"cldice-a novel topology-preserving loss function for tubular structure seg- mentation","cited_arxiv_id":null,"evidence_quote":"Supplies the soft-skeleton subroutine used by Negative Centerline Loss and is the main topology-aware baseline it is compared against."},{"cited_title":"CP-loss: Connectivity-preserving Loss for Road Curb Detection in Autonomous Driving with Aerial Images","cited_arxiv_id":"2107.11920","evidence_quote":"Connectivity-preserving skeleton-alignment loss that motivates penalizing gaps on the skeleton rather than the full mask."},{"cited_title":"Berger, Alexander Weers, Nico Stucki, Daniel Rueckert, Ulrich Bauer, and Johannes C","cited_arxiv_id":null,"evidence_quote":"Non-differentiable graph-based topology-preserving method whose idea of re-applying loss in topology-critical regions inspired Simplified Topology Loss."},{"cited_title":"Jaeger, Simon Kohl, Jakob Wasserthal, Gregor Koehler, Tobias Norajitra, Sebastian Wirkert, and Klaus H","cited_arxiv_id":null,"evidence_quote":"Provides the self-configuring U-Net pipeline used to train and evaluate every loss function."},{"cited_title":"Light-sheet fluo- rescence expansion microscopy: Fast mapping of neural cir- cuits at super resolution","cited_arxiv_id":null,"evidence_quote":"Documents the light-sheet fluorescence expansion microscopy protocol that produces the AIS volumes and motivates the high-throughput setting."}],"review_version":1}