{"id":"a56dea9c-e7b1-488c-98dc-139c2e18b858","arxiv_id":"2607.07123","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":7,"one_line_summary":"A differentiable Max-Min reachability loss redirects segmentation gradients to connectivity bottleneck pixels, improving topological correctness across nine architectures and six datasets.","lead":"This paper introduces a training loss for image segmentation that concentrates gradient updates on the pixels most critical for keeping thin structures connected. It matters because broken connections in vessel or road networks can invalidate downstream medical or mapping analysis, and the method is plug-and-play with no inference cost.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Max-Min operator is never isolated from alternative reachability formulations on the same support graph, so the central novelty claim (that Max-Min algebra specifically is responsible for gradient routing) is empirically unverified.","rationale":"The paper's empirical results are genuinely strong: 47/54 improvements, up to +7.2 pp clDice, verified gradient redistribution (Fig. 7), and zero inference overhead. The TGS diagnosis is well-motivated and the Max-Min gradient routing argument is mathematically sound. However, the central novelty claim—that Max-Min algebra specifically is the key mechanism—rests on theoretical argument alone without the critical control experiment of swapping the operator on the same graph framework. This is not a minor gap: the paper explicitly frames its contribution against sum-based and product-based paradigms (Section II), so the absence of a head-to-head comparison on identical infrastructure is a real hole. The reader's verdict of CONDITIONAL is appropriate, but the specific condition should be 'verify that Max-Min outperforms sum/product reachability on the same support graph' rather than 'verify skeletonization quality,' since the latter is lower-risk and the former strikes at the core novelty. If the operator-swap experiment shows Max-Min is decisively superior, the paper merits acceptance; if not, the contribution is still valuable but the framing needs revision. Confidence remains HIGH because the concern is about attribution of the mechanism, not about whether the method works—the empirical evidence for the latter is convincing.","tokens_in":26452,"tokens_out":4170,"duration_ms":266952,"concrete_test":"On DRIVE and Massachusetts Roads with Mask2Former (matching Table IX's ablation base), replace the Max-Min DP in Eq. (9) with (a) a sum-based shortest-path reachability: r_sum^(k)(u,v) = max over paths of (1/k)·Σ w_uv, and (b) a product-based reachability: r_prod^(k)(u,v) = max over paths of Π w_uv, keeping V*, L_edge, L_seg (BA), multi-scale K, and all hyperparameters identical. If Max-Min's clDice exceeds both alternatives by >1 pp, the operator-specific claim is confirmed. If the gap is <0.5 pp, the contribution is the graph framework, not the algebra.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central novelty is that Max-Min algebra induces bottleneck-dominated gradient routing (Section III-C, Fig. 3), and this is what distinguishes WPRF from prior sum-based and product-based methods (Section II). However, the ablation in Table IX only compares against pixel-level baselines (BCE, Focal, BCE+clDice). It never replaces the Max-Min DP in Eq. (9) with a sum-based path cost (e.g., additive geodesic distance) or product-based affinity chain on the same domain-restricted support graph V* with the same multi-scale sampling, edge loss, and BA segmentation term. Without this comparison, the clDice gains cannot be attributed to the Max-Min operator specifically rather than to the general framework of graph-based reachability supervision + domain restriction + multi-scale propagation. The theoretical argument for why Max-Min should outperform sum/product formulations (Section II, Section III-C) is plausible but not self-evidently sufficient: sum-based path costs also concentrate gradients on short paths, and the domain restriction already prevents background shortcuts regardless of the algebraic operator. If a sum-based reachability loss on V* achieves comparable clDice gains, the paper's core contribution narrows from 'Max-Min gradient routing' to 'graph-based reachability supervision with domain restriction,' which is a weaker claim. The reader's concern about skeletonization quality (Phi_px artifacts) is valid but lower-risk because skeletonization operates on GT masks with standard algorithms and results are robust across datasets with varying morphology.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper introduces Widest-Path Reachability Fields (WPRF), a plug-and-play training module for connectivity-preserving segmentation of slender curvilinear structures (vessels, cracks, roads). The authors identify Topological Gradient Starvation (TGS)—the phenomenon where sparse connectivity-critical bottleneck pixels receive insufficient gradients under standard sum-based pixel-wise losses—and propose a differentiable Max-Min reachability objective on a domain-restricted support graph to redirect gradient flow to these bottlenecks. The method is evaluated across nine architectures and six datasets (including a newly introduced oral microvessel dataset, OMVIS) with fixed hyperparameters, showing clDice improvements in 47 of 54 method-dataset pairs. The experimental design is thorough: multiple backbone families, ablations isolating each loss component, gradient routing verification (Fig. 7), and hyperparameter sensitivity analysis.","tokens_in":27218,"tokens_out":1534,"duration_ms":230819,"significance":"The paper addresses a well-known and practically important problem: topological breaks in curvilinear structure segmentation. The TGS formalization provides a clear diagnostic framing, and the Max-Min gradient routing mechanism is a principled solution grounded in classical graph theory (widest-path problem). Key strengths include: (1) the gradient routing verification in Fig. 7, which directly visualizes the shift of gradient energy from thick to thin structures; (2) the breadth of the experimental validation—nine architectures, six datasets, fixed hyperparameters, 9 runs per configuration; (3) the ablation in Table IX isolating edge supervision, reachability supervision, domain restriction, and multi-scale propagation; (4) the plug-and-play, backbone-agnostic design with no inference overhead (Table X). The newly introduced OMVIS dataset with expert annotation and inter-observer agreement reporting is a useful community contribution. Code is stated to be available.","major_comments":[{"comment":"The central novelty claim is that Max-Min algebra specifically induces bottleneck-dominated gradient routing (Section III-C, Fig. 3), distinguishing WPRF from prior sum-based and product-based methods (Section II). However, the ablation in Table IX only compares against pixel-level baselines (BCE, Focal, BCE+clDice). It never replaces the Max-Min DP in Eq. (9) with an alternative reachability formulation—e.g., a sum-based additive path cost or a product-based affinity chain—on the same domain-restricted support graph V* with the same multi-scale sampling, edge loss, and BA segmentation term. Without this comparison, the clDice gains cannot be attributed to the Max-Min operator specifically rather than to the general framework of graph-based reachability supervision + domain restriction + multi-scale propagation. The theoretical argument in Section III-C is plausible but not self-evident:","section":null},{"comment":"Dice drops are observed in several configurations (e.g., Swin-UNet on OCTA-500 6mm: 0.801→0.775 in Table V; Swin-UNet on OCTA-500 3mm: 0.828→0.798 in Table IV; CS-Net on OCTA-500 3mm: 0.796→0.788 in Table IV). The paper acknowledges a 'slight Dice drop' in the ablation (Section IV-E) but does not systematically analyze when or why Dice degrades. For a method claiming to preserve connectivity 'without sacrificing region overlap' (Section IV-C), the conditions under which pixel-level overlap significantly decreases should be characterized. Is this a trade-off inherent to the bottleneck-aware weighting (Eq. 12), or does the reachability loss actively suppress certain foreground regions? A brief analysis of the Dice-clDice trade-off space would strengthen the paper.","section":null}],"minor_comments":[{"comment":"Section III-B, Eq. (2): the deterministic skeletonization operator Phi_px is described as 'closing (3x3) + Zhang-Suen thinning' in Table I but the text mentions 'lightweight morphological preprocessing to fill small gaps.' The sensitivity of the support graph V* to the choice of skeletonization algorithm is not discussed. A brief note on robustness to skeletonization quality would strengthen the paper.","section":null},{"comment":"Table I states 'Augmentation: None.' For a method evaluated across six datasets with varying characteristics, the absence of any data augmentation is unusual and may disadvantage baselines. A brief justification for this choice would be helpful.","section":null},{"comment":"Section IV-F.3, Table X: training time increases by 2-3x with WPRF (e.g., UNet on DRIVE: 9.7→79.7 ms/iter). While the paper notes this is training-only overhead, the practical implications for large-scale experiments are not discussed. The paper mentions 'sparse Max-Min propagation' as future work but does not quantify how much overhead could be reduced.","section":null},{"comment":"Fig. 4: the y-axis label 'ΔclDice (pp)' and the markers are clear, but the 'dataset min-max' whiskers overlap with individual backbone markers in some cases, making it difficult to distinguish per-backbone values from the range. Consider jittering or using a different visual encoding.","section":null},{"comment":"Section III-D, Eq. (10): the positive pair distance criterion dist ∈ [ceil(k/2), k] is motivated as excluding 'trivially reachable pairs,' but the choice of ceil(k/2) as the lower bound is not justified. Is this sensitive to the choice of lower bound?","section":null},{"comment":"The paper introduces the term 'Bottleneck-Aware Balanced Hard-Negative BCE' for L_seg (Section III-D) but the 'bottleneck-aware' aspect refers to the weight W-bar(x) based on skeleton radius (Eq. 12), which is conceptually distinct from the Max-Min bottleneck routing in L_reach. The shared use of 'bottleneck' terminology for both the pixel-level reweighting and the graph-level Max-Min routing may cause confusion. Clarifying the relationship between the two mechanisms would help.","section":null},{"comment":"Reference formatting: some entries have inconsistent capitalization and venue abbreviations (e.g., 'arXiv preprint' vs. full conference names). A pass through the reference list for consistency would improve presentation.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The stress-test concern about isolating the Max-Min operator from alternative reachability formulations is the most substantive issue. It is a fair point: the paper's central novelty claim is specifically about Max-Min algebra, but the ablation does not test this in isolation. However, I assess this as a minor revision rather than major because: (1) the theoretical argument for Max-Min gradient routing is sound and is directly supported by the gradient routing verification in Fig. 7; (2) the comparison against BCE+clDice in Table IX does include a topology-aware baseline; (3) the breadth of the experimental validation (9 architectures, 6 datasets) provides strong evidence for the general framework. The missing ablation (Max-Min vs. sum-based reachability on the same graph) would strengthen the specific novelty claim but is not strictly necessary to validate the method's effectiveness. I recommend requesting this ablation but not making it a condition of acceptance if the authors provide a convincing theoretical argument for why sum/product formulations would not achieve the same gradient concentration. The Dice drop issue is worth flagging but is not load-bearing for the central claim about connectivity improvement."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"Here's the short version: this paper proposes WPRF, a differentiable Max-Min reachability objective for connectivity-preserving segmentation, and it works well across nine architectures and six datasets with fixed hyperparameters. The core idea — that pixel-wise losses starve connectivity-critical bottleneck pixels of gradient (TGS), and that Max-Min algebra on a domain-restricted graph fixes this — is clean and well-motivated. The experimental breadth is genuinely impressive: 47/54 method-dataset pairs improve in clDice, the gradient routing mechanism is directly visualized (Fig. 7), and there's zero inference overhead. Code and a new dataset (OMVIS) are released. This is real, reproducible work that practitioners will use. Credit earned on all of that. The soft spot that matters most: the ablation in Table IX never isolates the Max-Min operator from alternative reachability formulations on the same support graph. The paper's central novelty claim is that Max-Min algebra specifically induces bottleneck-dominated gradient routing, distinguishing it from sum-based and product-based methods (Section II). But the ablation only compares against pixel-level baselines (BCE, Focal, BCE+clDice). It never swaps the Max-Min DP in Eq. (9) for a sum-based path cost or product-based affinity chain on the same domain-restricted graph V* with the same multi-scale sampling and edge loss. Without that comparison, the clDice gains could be driven by the general framework of graph-based reachability supervision + domain restriction + multi-scale propagation rather than by Max-Min specifically. If a sum-based reachability loss on V* achieves comparable gains, the contribution narrows from 'Max-Min gradient routing' to 'graph-based reachability supervision with domain restriction,' which is a weaker but still useful claim. The theoretical argument for why Max-Min should outperform sum/product is plausible but not self-evidently sufficient — sum-based path costs also concentrate gradients on short paths, and domain restriction already prevents background shortcuts regardless of the operator. Secondary concerns are minor by comparison. Dice degrades in some configs (e.g., Swin-UNet on OCTA-500 3mm: 0.828→0.798), but the connectivity-overlap tradeoff is expected and usually favorable. The skeletonization dependency (Phi_px on GT masks) is not ablated, but it uses standard algorithms on clean annotations, so risk is low. The 2-3x training overhead is non-trivial but acceptable given zero inference cost. The 2D tree-like structure limitation is honestly acknowledged. This paper is for medical imaging and remote sensing researchers who need topologically correct segmentation without instance labels. It deserves a serious referee. The missing operator-isolation ablation is the one thing I'd push the authors to add — it's a clean experiment and would either confirm the central claim or honestly narrow it. Either outcome is publishable.","headline":"Solid practical method with a real gap in the ablation logic — the Max-Min operator itself is never isolated from alternatives.","tokens_in":27267,"tokens_out":685,"would_cite":true,"duration_ms":106229,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Max-Min algebra fixes broken vessel segmentation by targeting","keywords":["topological gradient starvation","Max-Min algebra","widest path problem","connectivity-preserving segmentation","gradient routing","differentiable dynamic programming","slender structure segmentation","bottleneck pixels"],"falsifier":"If replacing the Max-Min operator with a sum-based or product-based reachability formulation on the same domain-restricted support graph produced equivalent clDice improvements, the bottleneck-dominated gradient routing mechanism would not be the operative cause—instead, domain restriction and multi-scale supervision alone would explain the gains.","tokens_in":26660,"feed_emoji":"🔬","tokens_out":1250,"duration_ms":169263,"temperature":0.7,"pith_summary":"The paper identifies a systematic failure mode in segmenting thin curvilinear structures—retinal vessels, cracks, roads—where standard pixel-wise losses like Dice and BCE distribute gradient updates uniformly across all pixels, yet actual connectivity depends on a vanishingly small number of bottleneck pixels. Because these bottleneck pixels are so sparse, their aggregate gradient contribution is negligible compared to thick structures and background, so the network never learns to repair the breaks that fragment the network. The authors call this Topological Gradient Starvation (TGS). Their solution, Widest-Path Reachability Fields (WPRF), replaces sum-based gradient aggregation with Max-Min algebra: a path's strength equals its weakest edge, so backpropagation through the min operator concentrates gradients exclusively on the current bottleneck edge, and once that edge strengthens, the max operator shifts focus to the next weakest link. This produces a seek-and-repair dynamic that progressively fixes connectivity breaks without any inference-time overhead. The method is implemented as a plug-and-play training loss on top of any segmentation backbone, using differentiable dynamic programming over a skeleton-derived support graph, and improves connectivity metrics in 47 of 54 architecture–dataset combinations tested.","feed_headline":"Max-Min algebra fixes broken vessel segmentation by targeting","feed_subtitle":"Pixel-wise losses starve the sparse bottleneck pixels that hold thin structures together; a differentiable widest-path objective redirects","key_machinery":"Differentiable Max-Min dynamic programming over a skeleton-derived support graph. For each source node, the algorithm iteratively propagates reachability values using the recurrence r_{t+1}(v) = max(r_t(v), max_{p} min(r_t(p), w_{pv})), where w_{pv} is the learned edge weight between adjacent nodes. The min operator retains only the bottleneck edge's gradient per candidate path; the max operator selects the strongest path. Ties return a valid subgradient. The support graph is restricted to the ground-truth skeleton domain V* to prevent background shortcuts. Three loss terms are combined: a bottleneck-aware segmentation loss that upweights thin structures via inverse skeleton radius, a local ","core_discovery":"The central claim is that connectivity failures in slender-structure segmentation stem not from inadequate model capacity or training data, but from an algebraic property of standard loss functions: sum-based operators spread gradients uniformly, starving the sparse bottleneck pixels that actually determine topological continuity. By replacing the sum operator with a Max-Min operator—where a path's reachability is defined by its weakest edge and the gradient flows only through that bottleneck—training automatically redirects learning signal to the pixels that matter most for connectivity. The differentiable k-step dynamic programming implementation on a domain-restricted support graph makes ","pith_inferences":["The Max-Min gradient routing principle may extend to other domains where a sparse subset of elements determines global property correctness—e.g., chain-of-thought reasoning where one weak logical link breaks an argument, or program synthesis where a single incorrect operation invalidates output—suggesting that bottleneck-dominated gradient flow is a general optimization principle beyond pixel segm","The 2–3× training time overhead from Max-Min propagation could potentially be reduced by sparse or approximate variants that only propagate through candidate bottleneck regions identified by low-confidence edges, rather than full k-step DP over all nodes, which would make the approach practical for 3D volumetric data where the current formulation may be prohibitively expensive.","If the skeletonization operator Phi_px introduces systematic biases—e.g., preferentially retaining certain vessel calibers or missing sub-pixel connections—the Max-Min objective would faithfully optimize the wrong connectivity targets, suggesting that sensitivity to skeletonization quality is the critical failure mode that would distinguish between TGS as a fundamental phenomenon versus an artifac"],"forward_implications":["Any segmentation task where topology matters more than pixel overlap—neuron tracing, road network extraction, crack detection—could benefit from replacing or augmenting sum-based losses with Max-Min reachability objectives, since the gradient routing mechanism is architecture-agnostic and requires no skeleton labels beyond what is derivable from binary masks.","The seek-and-repair gradient pattern suggests a natural curriculum: as bottlenecks are progressively strengthened, the network automatically shifts focus, which could reduce or eliminate the need for multi-stage training pipelines or hard-example mining heuristics currently used in vessel and road segmentation.","The domain-restriction strategy—constraining graph propagation to skeleton-derived support rather than full foreground occupancy—demonstrates that preventing background shortcuts is as important as the Max-Min operator itself, which has implications for any graph-based learning signal where spurious paths could create shortcuts.","The identification of TGS as a structural property of sum-based losses, rather than a data or capacity problem, reframes connectivity preservation as an optimization geometry question that could be addressed at the loss-function level for many beyond-segmentation tasks involving graph-structured outputs."],"fun_headline_variants":["Max-Min reachability fields route gradients to connectivity bottlenecks in vessel segmenta","Differentiable widest-path objective fixes topological breaks in slender-structure segment","Bottleneck-focused gradient routing improves connectivity in retinal vessel and crack segm","Widest-path reachability fields counter topological gradient starvation in curvilinear seg","Max-Min algebra redirects training signal to sparse bottleneck pixels holding thin structu"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The method assumes that a deterministic skeletonization operator applied to the ground-truth binary mask faithfully captures the connectivity structure the network should learn, and that supervising reachability on this coarsened skeleton graph transfers to the full-resolution pixel mask used at inference. If the skeletonization introduces artifacts or the stride-based coarsening loses critical bottleneck topology, the Max-Min gradient routing targets may not align with real","fun_headline_variants_meta":{"raw":{"variants":["Max-Min reachability fields route gradients to connectivity bottlenecks in vessel segmentation","Differentiable widest-path objective fixes topological breaks in slender-structure segmentation","Bottleneck-focused gradient routing improves connectivity in retinal vessel and crack segmentation","Widest-path reachability fields counter topological gradient starvation in curvilinear segmentation","Max-Min algebra redirects training signal to sparse bottleneck pixels holding thin structures together"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":720,"prompt_tokens":634,"completion_tokens":86,"prompt_tokens_details":null},"tokens_in":634,"tokens_out":86,"duration_ms":55943,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T19:34:52.819031+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If replacing the Max-Min operator with a sum-based or product-based reachability formulation on the same domain-restricted support graph produced equivalent clDice improvements, the bottleneck-dominated gradient routing mechanism would not be the operative cause—instead, domain restriction and multi-scale supervision alone would explain the gains.","supporting_citations":[],"review_version":1}