{"id":"f0594e5f-bad5-4e3a-bdc5-998212d303e2","arxiv_id":"2507.04008","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PASC-Net's four-direction bendable strip convolutions and hierarchical topology losses improve vessel segmentation by 1.3 to 2.4 Dice points on the ARCADE and FIVES datasets.","lead":"A new plug-in module for medical image segmentation learns four elongated, bendable convolution kernels that follow the shape of blood vessels, plus extra training signals that keep vessel branches connected. Applied to coronary X-ray angiography and retinal fundus photos, the method raises Dice scores by about 1 to 2 points over strong baselines like nnU-Net.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SSL module, the main source of the claimed gains, is not reconstructable from the paper: no offset-prediction mechanism is specified, and the 'convolutional invariance' equalities in Eqs. (1)-(4) are algebraically invalid as written.","rationale":"The reader's weakest-assumption analysis identifies exactly the same load-bearing issue: the SSL module's offset-prediction mechanism is never specified, and Eqs. (1)-(4) contain invalid equalities. I agree that this is the most serious problem because the SSL module is the primary novelty and the largest contributor in the ablations. Even if the experimental results are statistically solid, a paper whose central module cannot be implemented from the text, and whose mathematical derivation is internally inconsistent, cannot support a state-of-the-art claim. The provided code link may resolve the ambiguity, but the manuscript itself must state the mechanism; a code repository without a commit hash or a demonstrated correspondence to the equations is not a substitute. I therefore keep the reader's conditional verdict: the empirical contribution is plausible and worth verifying, but the paper needs a corrected specification of the offset process, fixed equations, a coherent definition of the connectivity matrix, and a reproducible code release before the central claim can be accepted.","tokens_in":16267,"tokens_out":4878,"duration_ms":60863,"concrete_test":"Clone the repository at https://github.com/IPMI-NWU/PASC-Net, instrument the SSL convolution layer to print the offset tensors Δx_m and Δy_m and their gradients during a single ARCADE training step, and verify (a) that the offsets are nonzero, input-dependent, and produced by a learned mechanism described in the code, and (b) that rerunning the nnU-Net SSL-only configuration from Table 2 (square→xyzw) reproduces the reported Dice improvement (81.10%→81.95%) within 0.1 points. If the code has no learned offset branch or the improvement does not reproduce, the central claim is not established as published.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that replacing standard convolutions with the four-direction SSL strip convolution yields consistent Dice gains—depends on the SSL operation being a well-defined, implementable transformation. Section 3.1.1 describes the offset process only qualitatively: kernel points are shifted perpendicularly with an iterative smoothing rule, but no equation or network is given for generating the per-point offsets Δx_m and Δy_m. There is no offset-prediction branch, no input-dependent mechanism, and no parameterization. Section 3.1.2 then asserts a 'convolutional invariance theorem' through Eqs. (1)-(4), but these equalities are not valid: for example, Eq. (1) claims Σ X(i+Δx, j+m)W(i, j+m) = Σ X(i, j+m)W(i+Δx, j+m), which is not an algebraic identity and does not follow from shift invariance; the subscripts on Δ are also absent, leaving the variation along the strip undefined. If the intended implementation is a DSCNet-style learned-offset branch, this is a close variant of a cited method and the novelty is reduced; if it is something else, the reader cannot determine what it is. Equation (10) is also incoherent: it defines connectivity by checking whether a pixel equals one of its eight neighbors, which is false for all k≠5 and trivially self-referential for k=5. The Table 2 checkmarks do not map unambiguously to the components described in the text, and the clDice metric is partially optimized directly by the centerline loss L_cl. Because the method cannot be reconstructed from the manuscript and the equations do not support the derivation, the reported empirical gains cannot currently be attributed to the described SSL mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PASC-Net, a vessel segmentation framework combining a shape self-learning (SSL) convolution module, intended to adapt strip-shaped convolutional kernels to tubular structures, and a hierarchical topology constraint (HTC) module with centerline, mask, and neighbor-connectivity losses. The authors report consistent improvements over U-Net, FCN, U-Mamba, and nnU-Net on the ARCADE and FIVES datasets, achieving state-of-the-art Dice scores of 82.39% and 91.83%, respectively, with claimed gains of about 1–2 Dice points over the strongest baselines.","tokens_in":16472,"tokens_out":3905,"duration_ms":39734,"significance":"If the method is exactly as claimed and reproducible, the plug-and-play nature of the SSL module and the multi-level topology constraints would be a practical contribution to vessel segmentation, and the cross-architecture ablations on two public datasets are a strength. The paper also reports standard deviations and provides a code link, which aids reproducibility. However, the significance is currently weakened by the incomplete formal description of the core SSL convolution and by internal inconsistencies in the connectivity loss and ablation mapping, so the central claims are not yet verifiable from the manuscript alone.","major_comments":[{"comment":"The SSL convolution is not defined precisely enough to be implemented. The offset process (Section 3.1.1) is purely qualitative: no network branch, equation, or parameterization is given for predicting the per-point shifts Δx_m and Δy_m, so it is unclear how the kernel shape is learned. The 'convolutional invariance theorem' in Section 3.1.2 then asserts equalities that are not algebraic identities. For example, Eq. (1) claims Σ_m X(i+Δx, j+m) W(i, j+m) = Σ_m X(i, j+m) W(i+Δx, j+m), which holds only under very special assumptions about X and W, and the m-dependence of the offsets is omitted. Because this is the load-bearing definition of the proposed module, the central claim that replacing standard convolutions with SSL convolutions yields the reported gains is not substantiated by the written description.","section":"3.1.1 and 3.1.2, Eqs. (1)–(4)"},{"comment":"The neighbor connectivity matrix is defined as C_{i,j,k} = 1 if P_{i,j} = N_k, where N_k includes P_{i,j} itself as the 5th element. As written, this makes C_{i,j,5} always 1 and C_{i,j,k} always 0 for k ≠ 5, so it does not encode any meaningful 8-neighbor connectivity. Consequently the ground-truth connectivity matrix used in L_con (Eq. 11) is degenerate, and the claimed topological constraint from the NC unit is not well defined.","section":"3.2.3, Eq. (10)"},{"comment":"The centerline loss L_cl in Eq. (7) is essentially the clDice objective (Dice between predicted and ground-truth centerlines), while clDice is also a headline reported metric in Table 1 and in the ablation Table 2. Therefore the reported clDice improvements are partly a training echo: optimizing L_cl is expected to raise the clDice metric by construction. The paper should either report clDice from models trained without the centerline loss, or discuss this circularity explicitly and justify why the clDice comparison is still informative.","section":"3.2.1 and Tables 1–2"},{"comment":"The ablation table is not aligned with the text. The text describes 'xy-based', 'zw-based', and 'xyzw-based' SSL configurations, while the table has separate checkmarks for x, y, z, w; it is unclear which rows correspond to which configuration. The text also says the fourth row for each backbone shows 'only the mask constraint', but the table rows include checkmarks for SSL components before the NC and CL columns, so the claimed isolation of the HTC components is not actually demonstrated. This makes the central ablation claim—that every added component improves metrics—unverifiable from the reported table.","section":"4.5, Table 2"},{"comment":"The total loss in Eq. (12) is given as L = L_dice + L_cl + L_con, but L_dice is never defined, while the mask constraint defined in Eq. (8) is called L_mask and is not included in the total loss. This leaves the actual training objective ambiguous: it is unclear whether the mask Dice loss is part of the optimization, whether L_dice refers to L_mask, or whether an additional unstated Dice loss is used.","section":"3.2 and Eq. (12)"}],"minor_comments":[{"comment":"Equations (5) and (6) are identical and appear to be a duplication artifact; one should be removed.","section":"3.1.2, Eqs. (5)–(6)"},{"comment":"The caption refers to the 'SSH module', which appears to be a typo for 'SSL module'.","section":"Figure 2 caption"},{"comment":"The phrase 'topological conherence' should be 'topological coherence'.","section":"2.2"},{"comment":"The text says 'zy-base convolutions' where 'zw-base' is clearly intended, and the description of the three SSL configurations does not match the column structure of Table 2.","section":"4.5.1"},{"comment":"The caption states that red indicates 'true negative (over-segmentation)' and blue indicates 'false positive (mis-segmentation)', but over-segmentation is a false positive and under-segmentation is a false negative; the color-to-error mapping appears mislabeled.","section":"Figure 4 caption"},{"comment":"Several entries in Table 2 are malformed, e.g., '68.260.96±' in the U-Mamba ARCADE row and '64.87±0.8 5' in the UNet ARCADE row; these should be corrected.","section":"Table 2"},{"comment":"The experimental setup does not report training epochs, loss-weight hyperparameters, or whether the reported runs use a fixed random seed; these details are needed to assess the stability of the reported standard deviations.","section":"4.2"}],"recommendation":"major_revision","confidential_remarks":"The proposed SSL module, as described, appears very close to the dynamic snake convolution of DSCNet (cite [32]), which uses learned offsets for strip-like kernels. If that is the case, the novelty claim needs to be substantially reframed and the differences made precise. The code link may allow the authors to clarify the implementation, but the written manuscript must be self-contained. Given the several formal errors in the core equations and the ablation mapping, I would not recommend acceptance before a careful revision that addresses these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful empirical paper whose central claim is probably true but whose method is under-specified and whose derivation is wrong as written. If the authors release verified code and fix the math, it's a solid contribution; right now it's not reproducible from the text alone.\n\nWhat's new and good: the combination of four-direction strip convolutions with hierarchical losses (centerline, mask, connectivity) is new, and the evaluation is more thorough than usual: four backbones (U-Net, FCN, U-Mamba, nnU-Net) across two public datasets, with fixed splits, standard deviations, and an ablation table that tracks every component. The headline gains—1.29 Dice on ARCADE and 2.43 on FIVES over nnU-Net—are several standard deviations and consistent across backbones. The related work honestly cites DSCNet, CoANet, and clDice; the novelty is incremental but the empirical claim is concrete.\n\nSoft spots: the offset process in Sec. 3.1.1 never says where Dx_m and Dy_m come from—no prediction branch, no equation, no network. That is the load-bearing component, since the ablation attributes the largest gains to the SSL module. Sec. 3.1.2's 'convolutional invariance theorem' is not a theorem; Eqs. (1)-(4) claim equalities that do not hold, with index and sign errors. Eq. (10) defines the connectivity matrix incoherently (checking whether a pixel equals one of its neighbors is false in general and self-referential at k=5). Also, L_cl is essentially the clDice objective, so the reported clDice gains are partly a training echo, and the ablation table's checkmarks don't cleanly map to the prose. Finally, the SOTA claim would need more recent baselines (e.g., 2024-2025 vessel segmentation methods) to hold.\n\nWhat to do: send to peer review, but as a major-revision. The experiments look honestly run and the central Dice/IoU claim has independent grounding; the problems are presentation, missing detail, and derivational rigor. Require the authors to specify the offset mechanism, correct the equations, release code with a commit hash, and add recent baselines. A serious referee can extract the real value here.\n\nFor a reader: anyone working on vessel or tubular-structure segmentation. It's not a breakthrough, but it is a potentially useful drop-in module. I'd bring it to reading group to discuss the gap between the empirical result and the write-up.","headline":"Plausible empirical gains on two solid datasets, but the SSL module cannot be reconstructed from the paper and the derivation in Eqs. (1)-(4) is mathematically invalid; worth refereeing as a major-revision candidate.","tokens_in":17194,"tokens_out":2500,"would_cite":false,"duration_ms":26328,"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":"PASC-Net claims that replacing standard square convolutions with four self-learning strip kernels and adding hierarchical topology losses consistently improves vessel segmentation on four backbones, reaching state-of-the-art Dice of…","keywords":["vessel segmentation","shape self-learning convolution","strip convolution","hierarchical topology constraints","centerline loss","neighbor connectivity","coronary angiography","fundus image segmentation"],"falsifier":"Run the ARCADE experiment with an SSL module implemented exactly as the text specifies: because Section 3.1.1 does not say how $\\Delta x_m$ and $\\Delta y_m$ are predicted, no independent implementation can be built, and if a DSCNet-style offset branch is supplied instead, the reported 1.29-point Dice gain over nnU-Net is no longer a test of the paper's stated mechanism. A faster check: substitute any random feature map and kernel into Eq. (1); the two sides are not equal, so the kernel-shift/feature-map-shift equivalence the derivation relies on is not the identity the paper claims.","tokens_in":15929,"feed_emoji":"🩸","tokens_out":8635,"duration_ms":92539,"temperature":0.7,"pith_summary":"The paper tries to show that two plug-in components can fix the two things that most often go wrong in blood-vessel segmentation: tiny branches being missed and the vessel tree breaking at bifurcations. The first component, Shape Self-Learning (SSL) convolution, replaces ordinary square kernels with four elongated strip kernels (horizontal, vertical, and the two diagonals) whose points shift perpendicularly while sliding, so the receptive field bends with the vessel. The second, Hierarchical Topology Constraints (HTC), adds losses on centerlines, masks, and eight-neighbour connectivity to push predictions toward connected trees. On two public datasets, coronary angiography and retinal fundus photos, the authors report consistent gains across four segmentation backbones, with the strongest variant exceeding all compared methods. A sympathetic reader would take the claim as: a compact, architecture-agnostic module plus three loss terms raises overlap and topology metrics by roughly one to two Dice points and reduces average surface distance.","feed_headline":"Strip convs lift vessel segmentation to state-of-the-art Dice","feed_subtitle":"A plug-in kernel swap plus topology losses gains 1 to 2 Dice points on X-ray and fundus vessel datasets.","key_machinery":"The central object is the SSL convolution: four length-9 strip kernels (directions $x$, $y$, $z$, $w$) whose points are shifted perpendicularly by offsets $\\Delta x_m$, $\\Delta y_m$ around a fixed midpoint, then applied as ordinary convolutions and summed. The paper invokes a 'convolutional invariance theorem' to claim that shifting the kernel is equivalent to shifting the feature map, so the learnable shape can be implemented as a sum of standard convolutions. The second mechanism is the HTC loss, which combines centerline Dice (line level), mask Dice (surface level), and a BCE on an 8-neighbor connectivity matrix (the neighborhood level). Together these carry the argument: the SSL kernel adapts the receptive field to tubular geometry, and HTC regularizes the output so connectivity is penalized at every scale.","core_discovery":"The authors claim that standard convolutions are a poor inductive bias for tree-like tubular structures, and that making kernels shape-adaptive at inference time is enough to recover lost branches and broken topology. Concretely, they replace every convolutional layer's square kernel with the sum of four strip convolutions oriented horizontally, vertically, diagonally, and anti-diagonally; each strip's non-center points acquire learned perpendicular offsets, and the four outputs are added to form the SSL convolution. Around this they add the HTC loss: centerline Dice to keep the line skeleton aligned, mask Dice to keep the surface overlap, and a binary cross-entropy term on an 8-neighbour connectivity matrix to keep adjacent vessel pixels linked. Integrated into nnU-Net, the method reports Dice 82.39% versus 81.10% for nnU-Net on ARCADE and 91.83% versus 89.40% on FIVES, with IoU, clDice, and ASD improving as well. Ablation tables on U-Net, FCN, U-Mamba, and nnU-Net show every added component improving most metrics on both datasets.","pith_inferences":["Because the offset-generation mechanism is unspecified, the module's novelty may reduce to a four-direction variant of DSCNet-style learned offsets; if so, the distinctive contribution would be the HTC losses, not the shape learning itself.","The precision/recall pattern on FIVES suggests the gains come from detecting more small branches at the cost of a small precision drop; a clinical deployment would need to decide whether the extra positives are acceptable.","The HTC loss has a direct 3D analogue: replace the 8-neighbour connectivity matrix with a 26-neighbour one and apply the same centerline/mask/connectivity triple to CTA or MRA volumes, which the authors name as future work.","On datasets whose annotations are only coarse masks without skeletonized ground truth, the centerline and connectivity losses would have to be computed from the masks themselves, so the reported benefit may depend on annotation quality."],"forward_implications":["Replacing square kernels with the four-direction SSL strip convolution is reported to raise Dice, IoU, and clDice on both datasets for every backbone tested, so the module transfers without architectural redesign.","Adding the HTC losses (centerline, mask, and 8-neighbour connectivity) is reported to reduce topological breakpoints and lower average surface distance compared with the same networks without them.","The full PASC-Net on nnU-Net is reported to outperform all compared methods on ARCADE and FIVES, with Dice gains of 1.29 points and 2.43 points over nnU-Net, respectively.","Because the SSL module is described as plug-and-play, the same gain pattern is what the authors predict if the module is dropped into any convolutional segmentation backbone.","The HTC module's centerline and connectivity losses directly target the bifurcation breakpoints that overlap-only losses miss, so the framework's benefit is largest exactly where topology, not just pixel overlap, matters for clinical use."],"supporting_citations":[{"why":"U-Net serves as the baseline architecture and the first backbone in the ablation studies.","marker":"[35]"},{"why":"FCN provides the second backbone used to test SSL and HTC transferability.","marker":"[24]"},{"why":"U-Mamba is the third backbone and a comparison method representing long-range dependency models.","marker":"[25]"},{"why":"nnU-Net is the main framework the proposed modules are embedded in and the strongest baseline to beat.","marker":"[11]"},{"why":"clDice supplies the centerline-based topological loss that the HTC centerline constraint draws on, and is a comparison baseline.","marker":"[39]"},{"why":"DSCNet is the dynamic snake convolution baseline whose learned-offset idea the SSL module most directly extends.","marker":"[32]"},{"why":"CoANet is the strip-convolution baseline for tubular structures and the second-best method on FIVES.","marker":"[26]"},{"why":"ARCADE supplies the coronary X-ray angiography dataset used for the main evaluation.","marker":"[31]"},{"why":"FIVES supplies the fundus image dataset used for the retinal vessel evaluation.","marker":"[14]"},{"why":"The centerline calculation for the HTC line-level constraint is inspired by this hierarchical topology learning work.","marker":"[45]"}],"fun_headline_variants":["Strip convolutions with topology constraints improve vessel segmentation","Shape-adaptive kernels: plug into U-Net and beat state-of-the-art Dice","Topology-aware strip convolutions recover small vessels and fix breaks","Plug-in shape learning and topology losses push vessel segmentation forward"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the SSL module's per-point perpendicular offsets are produced by a defined learned mechanism; the paper never specifies how $\\Delta x_m$ and $\\Delta y_m$ are generated, and the convolution-invariance identity used to justify the kernel shifts has sign and index errors as written.","fun_headline_variants_meta":{"raw":{"variants":["Strip convolutions with topology constraints improve vessel segmentation","Shape-adaptive kernels: plug into U-Net and beat state-of-the-art Dice","Topology-aware strip convolutions recover small vessels and fix breaks","Plug-in shape learning and topology losses push vessel segmentation forward"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000826,"raw_usage":{"total_tokens":3686,"prompt_tokens":1095,"completion_tokens":2591,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":711,"completion_tokens_details":{"reasoning_tokens":2519}},"tokens_in":711,"tokens_out":2591,"duration_ms":22923,"temperature":1.0,"reasoning_tokens":2519,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:59:15.161494+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the ARCADE experiment with an SSL module implemented exactly as the text specifies: because Section 3.1.1 does not say how $\\Delta x_m$ and $\\Delta y_m$ are predicted, no independent implementation can be built, and if a DSCNet-style offset branch is supplied instead, the reported 1.29-point Dice gain over nnU-Net is no longer a test of the paper's stated mechanism. A faster check: substitute any random feature map and kernel into Eq. (1); the two sides are not equal, so the kernel-shift/feature-map-shift equivalence the derivation relies on is not the identity the paper claims.","supporting_citations":[{"cited_title":"U-Net: Convolutional networks for biomedical image segmentation, in: Medical Image Computing and Computer-Assisted Intervention, Springer","cited_arxiv_id":null,"evidence_quote":"U-Net serves as the baseline architecture and the first backbone in the ablation studies."},{"cited_title":"Fully convolutional networks for semantic segmentation, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"FCN provides the second backbone used to test SSL and HTC transferability."},{"cited_title":"clDice-a novel topology-preservinglossfunctionfortubularstructuresegmentation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"clDice supplies the centerline-based topological loss that the HTC centerline constraint draws on, and is a comparison baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DSCNet is the dynamic snake convolution baseline whose learned-offset idea the SSL module most directly extends."},{"cited_title":"CoANet:Connectivity attention network for road extraction from satellite imagery","cited_arxiv_id":null,"evidence_quote":"CoANet is the strip-convolution baseline for tubular structures and the second-best method on FIVES."},{"cited_title":"Dataset for Automatic Region-based Coronary Artery Disease Diagnostics Using X-Ray Angiography Images","cited_arxiv_id":null,"evidence_quote":"ARCADE supplies the coronary X-ray angiography dataset used for the main evaluation."},{"cited_title":"Fives: A fundus image dataset for artificial Intelligence based vessel segmentation","cited_arxiv_id":null,"evidence_quote":"FIVES supplies the fundus image dataset used for the retinal vessel evaluation."},{"cited_title":"Progressive deep segmentation of coronary arteryviahierarchicaltopologylearning,in:InternationalConference on Medical Image Computing and Computer-Assisted Intervention, Springer","cited_arxiv_id":null,"evidence_quote":"The centerline calculation for the HTC line-level constraint is inspired by this hierarchical topology learning work."}],"review_version":1}