{"id":"cc514eb2-6a73-4133-a360-c6c9ecfd25c0","arxiv_id":"2505.03896","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Adding attention gates and a label-guided pixel contrastive loss to a KAN-based U-Net yields small but consistent F1 improvements across five retinal vessel datasets.","lead":"This paper combines attention gates and a contrastive loss with a Kolmogorov-Arnold network to segment retinal blood vessels in eye images. The authors report top F1 scores on several public datasets, though the gains over the strongest baselines are about one point or less.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 20 defines MIoU as 2TP/(TP+FP+FN), but every reported row has F1 > MIoU; because F1 = 2TP/(2TP+FP+FN) < 2TP/(TP+FP+FN), the printed formulas cannot both hold. The MIoU half of the headline claim is not reproducible as written.","rationale":"The reader's conditional verdict focuses on reproducibility of the evaluation protocol: no code, no seeds, private dataset, small margins. I agree that those are real obstacles to full verification. The specific issue I find more load-bearing is an internal contradiction in the metric definitions. Eqs. 19 and 20, taken together, require MIoU > F1 for any valid confusion matrix, while every reported row has F1 > MIoU. This means the paper as printed does not define the MIoU it reports. The numerical values strongly suggest the authors computed standard foreground IoU (TP/(TP+FP+FN)) and wrote the factor 2 in Eq. 20 by mistake; the F1-to-IoU mapping (IoU = F1/(2−F1)) matches the table entries within about 0.1–0.3 percentage points. I therefore do not treat this as evidence of fabrication or even as a likely error in the experiments, but it is exactly the kind of unverifiable definition that a conditional verdict should require the authors to fix. The same correction would also interact with the missing-code issue: with corrected formulas and released code, a reviewer could recompute all MIoU values and confirm the ranking. Because the reader already assigned CONDITIONAL, my independent concern supports that verdict rather than moving it. I would not reject the paper on this basis alone, since the F1 results are defined consistently and the MIoU ordering is likely to survive correction, but the preprint should not be accepted without the correction and the verification it enables.","tokens_in":23913,"tokens_out":10100,"duration_ms":97070,"concrete_test":"Recompute the DRIVE AttUKAN row from the published SE, SP, ACC, F1 and the public DRIVE ground-truth pixel counts using the equations in Section 4.2. First, solve for TP, FP, FN and TN and evaluate both Eq. 20 and the standard IoU TP/(TP+FP+FN). If Eq. 20 exceeds F1 (or exceeds 100%), the printed definition is invalid. Then check whether the standard IoU reproduces the reported 70.24%; if it does, correct Eq. 20 and re-certify the MIoU rankings in Tables 2–7 to see whether the claimed highest-MIoU result survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim asserts AttUKAN achieves the highest F1 and MIoU among 11 networks on all five datasets. F1 is defined in Eq. 19 as 2TP/(2TP+FP+FN), and MIoU is defined in Eq. 20 as 2TP/(TP+FP+FN). Let D = TP+FP+FN. Then F1 = 2TP/(D+TP) and MIoU = 2TP/D, so the printed definitions require MIoU > F1 for every non-degenerate confusion matrix. Tables 2–6 show the opposite relation in every row, e.g., DRIVE AttUKAN has F1 = 82.50% and MIoU = 70.24%. No TP/FP/FN values can satisfy both printed formulas. The reported MIoU values instead match the standard foreground IoU, TP/(TP+FP+FN); for example F1 = 82.50% implies IoU = 82.50/(200−82.50) = 70.21%, within rounding of 70.24%. Thus Eq. 20 appears to contain a spurious factor of 2, and the manuscript does not actually specify the MIoU metric it reports. Because the SOTA claim is jointly about F1 and MIoU, the MIoU component cannot be verified from the preprint as written. This is an internal formal inconsistency, not merely a missing-seed reproducibility caution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AttUKAN, a U-shaped Kolmogorov-Arnold Network with Attention Gates placed on skip connections, and a Label-guided Pixel-wise Contrastive Loss (LPCL) to improve retinal vessel segmentation. The method is evaluated on DRIVE, STARE, CHASE_DB1, HRF, and a private dataset, with F1 and MIoU reported as the primary metrics. The authors claim state-of-the-art results against 11 baseline networks across all five datasets, and support the claim with ablation studies on the network components, loss terms, the LPCL weighting coefficient, the feature level used for LPCL, and the transferability of LPCL to other architectures.","tokens_in":24202,"tokens_out":9534,"duration_ms":89974,"significance":"If the results are validated, the paper offers a practical and portable improvement: attention-gated KAN blocks plus a supervised contrastive loss that also improves several existing U-shaped networks. The study is unusually broad in its evaluation, covering five datasets and multiple ablations, and the transfer experiment of LPCL across architectures is a genuine strength. However, the central quantitative claim is currently undermined by an internally inconsistent metric definition, a complete absence of statistical uncertainty, and an unclear statement of the spatial resolution at which LPCL operates. The reported F1 gains over the best baselines are small (0.07 to 0.54 percentage points), so without error bars or significance testing the 'highest compared to 11 networks' claim is not yet demonstrated. The paper ships no code and gives no experimental logs, further limiting verification.","major_comments":[{"comment":"The definition of MIoU is internally inconsistent with all reported results. Equation (19) defines F1 = 2TP/(2TP+FP+FN) and Eq. (20) defines MIoU = 2TP/(TP+FP+FN). Since 2TP+FP+FN > TP+FP+FN for any non-degenerate confusion matrix, Eq. (20) forces MIoU > F1 in every row, yet every entry of Tables 2–7 shows the opposite. For example, DRIVE AttUKAN reports F1=82.50% and MIoU=70.24%; combining Eqs. (19) and (20) algebraically gives MIoU = 2F1/(2−F1) = 140.4% for F1=82.50%, not 70.24%. The reported MIoU values are numerically consistent with the standard foreground IoU, TP/(TP+FP+FN): for F1=82.50%, IoU=82.50/(200−82.50)=70.21%, matching 70.24% up to rounding. The formula must be corrected (most plausibly to TP/(TP+FP+FN)) or the reported MIoU values cannot be reproduced. Because the abstract and Section 4.4 base the SOTA claim jointly on F1 and MIoU, this is a load-bearing defect.","section":"Section 4.2, Eq. (20)"},{"comment":"No variance, error bars, or statistical significance tests are reported. The F1 advantages over the best baseline are 0.22 points on DRIVE, 0.15 on STARE, 0.54 on CHASE_DB1, 0.07 on HRF, and 0.11 on the private dataset, while the component ablations in Table 8 improve F1 by 0.10 to 0.75 points. With a single run and no reported seeds, these differences could easily fall within training noise. Please provide mean and standard deviation over at least three random seeds, or an equivalent paired statistical test, for the main comparisons and the ablations, so the claimed superiority can be distinguished from stochastic variation.","section":"Section 4.4, Tables 2–6"},{"comment":"LPCL is computed on the 5th-level bottleneck feature of size S×S. With the stated patch size of 64×64 and the 32× downsampling at the bottleneck shown in Figure 2, S=2, so only four spatial positions per image are available as anchors. This is at odds with the description of a 'pixel-wise' contrastive loss for fine-grained representations. The paper does not specify how the ground-truth label is mapped to the 2×2 feature grid, nor how the positive set Ω+_i is constructed across the S×S positions and across the two random augmentations. Please clarify the actual feature resolution, the label downsampling procedure, and the positive-pair construction; if S=2, justify how such a coarse spatial loss provides fine-grained supervision.","section":"Section 3.4 and Section 4.3"},{"comment":"The paper asserts that 'all experiments across different methods utilize the same dataset settings and partitions,' but it does not state whether the baseline results in Tables 2–6 are re-implemented by the authors or cited from the original papers, and no code is provided. The STARE/CHASE_DB1 split (first 10/14 images for training) is nonstandard, and the private dataset is not publicly available. Without experimental logs or code, the reader cannot verify that the small F1/MIoU gains are not artifacts of differing preprocessing, split, or training protocols. Please specify the provenance of each baseline number and release code or detailed configuration files for all experiments.","section":"Section 4.4, baseline comparisons"}],"minor_comments":[{"comment":"The 3rd-level-feature row for STARE reports MIoU=88.25% and AUC=88.25%, values that are mathematically incompatible with F1=81.04% under either MIoU definition and that deviate sharply from the other rows; these appear to be data-entry errors and should be corrected.","section":"Table 11"},{"comment":"The LPCL loss is not fully specified: the outer factor appears to be printed as '2N' but is likely intended to be '1/(2N)', and the construction of Ω+_i relative to the two augmentations X~2i and X~2i−1 is not defined explicitly.","section":"Eq. (11)"},{"comment":"The sentence 'For every two input images, we can form S² pixel-level contrastive pairs' does not match the loss formula, which sums over anchors s and positives j; the relationship between the number of pairs and the anchor/positive sets should be clarified.","section":"Section 3.4"},{"comment":"The two versions of Figure 3 use inconsistent loss labels ('Dice + BCE Loss' versus 'Dice + Jaccard + BCE Loss'), and Figure 3(c) contains a subscript typo in the activation notation (ReLu(σ1) versus ReLU(σ2)).","section":"Figures 2 and 3"},{"comment":"The Hausdorff distance formula has an unbalanced parenthesis in the quantile expressions, making the definition ambiguous.","section":"Section 4.2, Eq. (15)"},{"comment":"The private dataset is a core part of the five-dataset SOTA claim but is not publicly available; please add a supplementary data sheet or an explicit availability statement describing the acquisition, annotation, and ethical approval details.","section":"Section 4.1, private dataset"}],"recommendation":"major_revision","confidential_remarks":"The metric definition issue in Eq. (20) is likely a typo (the reported numbers match standard IoU), but it affects the central quantitative claim and must be corrected. The paper would be substantially strengthened by adding variance estimates and clarifying the LPCL spatial resolution. The novelty is incremental but may be sufficient for a specialized medical-imaging venue if the empirical claims are made reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: this is a workmanlike combination paper, not a breakthrough. The real new bit is integrating attention gates into UKAN's skip connections and adding a label-guided pixel-wise contrastive loss at the bottleneck. The ingredients are all known—KAN, additive attention, contrastive learning—but the integrated system is new, and the ablation work is broader than usual: five datasets, component ablations, per-loss ablations, a lambda_4 search, and a table showing LPCL also helps other baselines when dropped in. Credit where it's due: the F1 gains are small, 0.10 to 0.75 points, but they are consistent across all five datasets, and the qualitative figures back the quantitative story.\n\nThe biggest soft spot is a formal one that the reader's take missed. Eq. 20 defines MIoU as 2TP/(TP+FP+FN). With F1 = 2TP/(2TP+FP+FN), that printed MIoU is always larger than F1. Every row in Tables 2–6 shows the opposite—for example DRIVE F1 82.50% and MIoU 70.24%. The numbers actually match the standard foreground IoU, TP/(TP+FP+FN): F1 of 82.50 implies IoU of 70.21, within rounding of 70.24. So the reported MIoU column is standard IoU, and Eq. 20 contains a spurious factor of 2. Since the headline SOTA claim is jointly about F1 and MIoU, the MIoU half cannot be verified as written. It is a fixable typo, but it is load-bearing and needs to be corrected.\n\nThe rest is the usual reproducibility package: no code shipped, no seeds, no error bars, private dataset unavailable, and lambda_4 tuned per dataset (0.3 on three, 0.5 on two), which is a mild selection-on-validation issue. Gains of a few tenths of an F1 point could plausibly be training noise. Table 11 also has a typo (STARE MIoU 88.25, AUC column duplicated), which suggests the manuscript was rushed.\n\nBottom line: if the MIoU definition is fixed and code plus multiple-seed statistics are released, this is a solid minor empirical contribution. The method is a combination, not a new capability. The paper is for retinal-vessel segmentation researchers who want a slightly better baseline and a usable contrastive-loss recipe, and it deserves a serious referee. I would send it out, with instructions to the referee to check the metric definition, ask for uncertainty quantification, and require the code release or at least the processed private dataset before acceptance.","headline":"A workmanlike combination of known components with broad but unverified ablations; the MIoU formula as printed contradicts every reported table.","tokens_in":24811,"tokens_out":5377,"would_cite":false,"duration_ms":48254,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that AttUKAN, a U-shaped Kolmogorov-Arnold network with attention-gated skip connections and a label-guided pixel-wise contrastive loss, extracts more discriminative encoder features and reports the highest F1 and MIoU…","keywords":["retinal vessel segmentation","fundus image","Kolmogorov-Arnold Networks","attention mechanism","contrastive learning","fine-grained feature extraction","medical image segmentation"],"falsifier":"Retrain AttUKAN and the 11 baselines under one fixed public protocol (same splits, same preprocessing, several random seeds) on DRIVE, STARE, and CHASE_DB1, then compare each method's F1 mean and standard deviation; if AttUKAN's lead over the second-best method is smaller than the seed-to-seed spread, the reported advantage would fall within training noise.","tokens_in":23669,"feed_emoji":"👁️","tokens_out":8162,"duration_ms":66269,"temperature":0.7,"pith_summary":"Retinal vessel segmentation is a pixel-level classification problem where the hard part is distinguishing vessel from non-vessel within a single image, so the model needs fine-grained, discriminative features rather than just a decoder that matches labels. The paper proposes AttUKAN, a U-shaped Kolmogorov-Arnold network, together with a Label-guided Pixel-wise Contrastive Loss (LPCL) that pulls same-class pixel representations together and pushes different-class pixels apart. Attention gates on the skip connections filter the encoder features that flow to the decoder. The paper reports that AttUKAN achieves the highest F1 and MIoU among 12 compared methods on DRIVE, STARE, CHASE_DB1, HRF, and a private dataset, attributing the gains to the attention gates and the contrastive loss.","feed_headline":"AttUKAN beats 11 networks on five retinal datasets","feed_subtitle":"Attention gates and a pixel-wise contrastive loss lift F1 and MIoU on four public sets and a private one.","key_machinery":"The machinery is two additions to a U-shaped Kolmogorov-Arnold Network (KAN). First, Attention Gates in the skip connections compute a gating coefficient from the encoder feature and a higher-level gating signal, multiply the feature map elementwise, and suppress irrelevant activations before concatenation with the decoder. Second, the Label-guided Pixel-wise Contrastive Loss (LPCL) samples pixel-level feature vectors from the bottleneck feature map; pairs of pixels with the same ground-truth class are positives and different-class pairs are negatives, and a contrastive objective pulls same-class vectors together and pushes different-class vectors apart in representation space. The KAN blocks themselves replace linear layers with learnable nonlinear activation functions, giving the encoder more nonlinear modeling capacity.","core_discovery":"The central claim is that feature-level supervision is the missing ingredient in retinal vessel segmentation. Existing models train primarily by comparing decoder output with ground-truth labels, but AttUKAN applies a contrastive loss directly to the bottleneck feature map, using the label to define positive pairs (same class) and negative pairs (different classes). Attention gates in skip connections suppress irrelevant activations so finer vessel structure is retained. The paper reports F1 scores of 82.50%, 81.14%, 81.34%, 80.21%, and 80.09% and MIoU scores of 70.24%, 68.64%, 68.59%, 67.21%, and 66.94% across the five datasets, each the highest among the 11 baselines, and ablation studies attribute the gains to both the attention gates and LPCL.","pith_inferences":["Because LPCL is applied only at the bottleneck, one could test whether applying it at multiple decoder levels further improves small vessel tips; the paper does not explore this.","The method uses label-guided pixel pairs rather than heavy augmentation or large contrastive batches, which may make it easier to adopt when training data are scarce.","A direct probe of feature-map cluster purity (vessel vs background) could test whether LPCL's mechanism is indeed better feature separability and whether it correlates with F1 gain.","When the private dataset is released, external groups can check whether the advantage persists under an independent evaluation protocol."],"forward_implications":["If AttUKAN's results hold, feature-level contrastive supervision at the bottleneck is an effective complement to decoder-level losses for retinal vessel segmentation.","The ablations show LPCL also improves several existing U-Net style baselines when added to them, suggesting the loss transfers beyond the proposed architecture.","Attention-gated skip connections on a KAN backbone may generalize to other fine-structure segmentation tasks where thin structures dominate.","The reported gains on high-resolution datasets (HRF and the private set) indicate the approach works outside the small-patch regime.","The reported F1 and MIoU numbers, if reproducible, establish new quantitative baselines for these five datasets."],"supporting_citations":[{"why":"Introduces Kolmogorov-Arnold Networks, whose learnable nonlinear activations form the backbone that AttUKAN builds on.","marker":"[16]"},{"why":"Presents UKAN, the U-shaped KAN baseline that AttUKAN extends with attention gates and LPCL.","marker":"[17]"},{"why":"Supplies the attention gate mechanism used in skip connections to suppress irrelevant feature activations.","marker":"[12]"},{"why":"Provides the training/test split protocol for STARE and CHASE_DB1 and the preprocessing pipeline (CLAHE, gamma correction) that the experiments follow.","marker":"[7]"},{"why":"Supplies the contrastive learning formulation that LPCL adapts to pixel-level, label-guided pairs.","marker":"[18]"},{"why":"Provides the DRIVE dataset used for one of the five evaluation benchmarks.","marker":"[31]"},{"why":"Provides the STARE dataset used for evaluation.","marker":"[32]"},{"why":"Provides the CHASE_DB1 dataset used for evaluation.","marker":"[33]"},{"why":"Provides the HRF dataset used for evaluation.","marker":"[34]"},{"why":"Defines the vessel-specific connectivity, overlap, and length metrics used in the comparisons.","marker":"[35]"}],"fun_headline_variants":["Feature-level supervision sharpens retinal vessel segmentation","AttUKAN: attention gates plus contrastive loss for vessels","New loss and attention gates boost vessel segmentation","Retinal vessel segmentation hits new highs with AttUKAN","Contrastive feature learning improves vessel detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes every one of the 12 methods is trained and evaluated under identical splits, preprocessing, and hyperparameter settings, and that the reported single-run differences of about 0.1 to 1.0 F1 points reflect real model differences rather than training noise.","fun_headline_variants_meta":{"raw":{"variants":["Feature-level supervision sharpens retinal vessel segmentation","AttUKAN: attention gates plus contrastive loss for vessels","New loss and attention gates boost vessel segmentation","Retinal vessel segmentation hits new highs with AttUKAN","Contrastive feature learning improves vessel detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1546,"prompt_tokens":1078,"completion_tokens":468,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":395}},"tokens_in":694,"tokens_out":468,"duration_ms":4225,"temperature":1.0,"reasoning_tokens":395,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:42:01.898396+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain AttUKAN and the 11 baselines under one fixed public protocol (same splits, same preprocessing, several random seeds) on DRIVE, STARE, and CHASE_DB1, then compare each method's F1 mean and standard deviation; if AttUKAN's lead over the second-best method is smaller than the seed-to-seed spread, the reported advantage would fall within training noise.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the training/test split protocol for STARE and CHASE_DB1 and the preprocessing pipeline (CLAHE, gamma correction) that the experiments follow."},{"cited_title":"1597–1607","cited_arxiv_id":null,"evidence_quote":"Supplies the contrastive learning formulation that LPCL adapts to pixel-level, label-guided pairs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DRIVE dataset used for one of the five evaluation benchmarks."},{"cited_title":"Shuang ZengET AL 15","cited_arxiv_id":null,"evidence_quote":"Provides the STARE dataset used for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CHASE_DB1 dataset used for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the HRF dataset used for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the vessel-specific connectivity, overlap, and length metrics used in the comparisons."}],"review_version":1}