{"id":"db38cbb1-3de7-4e13-ad6d-abf311ea18aa","arxiv_id":"2501.02559","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"KM-UNet reports the best average IoU and F1 among compared methods on five medical segmentation benchmarks with only 7.35M parameters.","lead":"This paper describes KM-UNet, a U-shaped neural network that mixes two recent ideas, Kolmogorov-Arnold networks and Mamba-style state-space models, for segmenting medical images such as skin lesions and polyps. The authors report higher average accuracy than several U-Net variants while keeping the model small enough to run efficiently.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline comparability, not architecture, is the load-bearing uncertainty: Table 1 has no matched training protocols or error bars, so the claimed accuracy margins could disappear under a fair re-run.","rationale":"The reader's weakest assumption correctly identifies baseline comparability and missing variance estimates as the load-bearing issue. My reading of Tables 1 and 2 confirms that the reported averages are internally consistent, so the numerical claim is not arithmetically wrong; the risk is experimental fairness. The absence of per-baseline hyperparameters, seeds, and error bars makes it impossible to determine whether the observed improvements are due to the KM-UNet architecture or to the training protocol. The conclusion's references to ablation studies that are not present further weaken the causal attribution of the gains to specific design choices. A controlled re-run with official baseline implementations and identical protocol is the decisive check. Because this is exactly the condition the reader attached to their CONDITIONAL verdict, no new adjustment is needed; the verdict remains conditional pending those experiments.","tokens_in":9153,"tokens_out":3414,"duration_ms":34590,"concrete_test":"Retrain all six baselines under the exact protocol stated for KM-UNet in Section 4: Adam, initial learning rate 1e-4 with cosine decay to 1e-5, batch size 8, 300 epochs, random rotation and flipping, same 80/20 data split per dataset, and three seeds. Use official public implementations of U-Net, Att-UNet, UNet++, U-NeXt, Rolling-UNet, and U-Mamba. Report per-dataset mean and standard deviation in the format of Table 1, then recompute the Table 2 averages and FLOPs with the same input resolutions. If the mean IoU margin between KM-UNet and the best baseline falls below about 1 point on key datasets, or if the margins overlap within one standard deviation, the claim that KM-UNet 'achieves the highest IoU and F1 scores' must be weakened to a claim of competitive performance under a specific protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that KM-UNet achieves the highest IoU/F1 on most datasets with a favorable efficiency trade-off. For this claim to hold, the comparisons in Table 1 must reflect architecture quality rather than differences in training setup. The manuscript's Section 4 reports KM-UNet's own batch size (8), learning rate (1e-4, cosine to 1e-5), optimizer (Adam), loss (BCE+Dice), and 300-epoch schedule, but gives no comparable protocol for U-Net, Att-UNet, UNet++, U-NeXt, Rolling-UNet, or U-Mamba. No seeds or standard deviations are reported, even though the text says results were averaged over three runs. The margins that matter are small on several datasets (e.g., BUSI IoU 65.42 vs. 61.81 for U-Mamba; ISIC17 IoU 84.05 vs. 82.14 for Rolling-UNet), so run-to-run variance or protocol mismatch could plausibly change the ranking. The conclusion also cites ablation studies showing 2-3% gains, but no ablation table appears in the manuscript, so the attribution of the improvement to specific components (SEM, KAN, cosine schedule) is not currently supported. This is not an objection to the architecture itself; it is a statement that the paper's evidence does not yet rule out a much weaker conclusion such as 'KM-UNet is competitive under its own training protocol.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes KM-UNet, a U-shaped segmentation network that integrates Kolmogorov-Arnold Networks (KANs), state-space models (SSMs) via a Mamba-inspired selective-scan module, and a multi-scale attention module. The approach is evaluated on five medical segmentation benchmarks (BUSI, GlaS, CVC-ClinicDB, ISIC17, ISIC18) against six baselines, reporting an average IoU of 81.17, average F1 of 89.20, 7.35M parameters, and 17.66 Gflops. The authors claim that KM-UNet achieves the highest IoU and F1 on most datasets and strikes an excellent accuracy-efficiency balance, and that this is the first medical image segmentation framework integrating KANs and SSMs.","tokens_in":9425,"tokens_out":4645,"duration_ms":42668,"significance":"If the reported results are reproducible, the architecture is a potentially valuable contribution: it combines KAN-based tokenized blocks with SSM-style selective scanning in a U-Net, and the claimed efficiency (17.66 Gflops, 7.35M parameters) is attractive. The paper also provides open-source code, and the interpretability discussion is a useful direction. However, the current experimental evidence does not yet separate architectural gains from training-protocol differences, because baseline protocols, error bars, and ablation results are missing.","major_comments":[{"comment":"The implementation details state that 'all results were averaged over three independent runs,' but Table 1 reports only point estimates with no standard deviations, confidence intervals, or per-run values. The margins on several datasets are small (e.g., BUSI IoU 65.42 vs. 61.81 for U-Mamba; ISIC17 IoU 84.05 vs. 82.14 for Rolling-UNet), so without variance information it is impossible to assess whether the reported ranking is statistically meaningful. Please provide error bars or per-run results for all entries.","section":"Section 4, Table 1"},{"comment":"The manuscript gives the full training protocol for KM-UNet (Adam, learning rate 1e-4 with cosine schedule, batch size 8, 300 epochs, BCE+Dice loss, random rotation and flipping) but gives no comparable protocol for the baselines U-Net, Att-UNet, UNet++, U-NeXt, Rolling-UNet, and U-Mamba. Without evidence that baselines were trained under matched data splits, augmentation, and hyperparameters, the advantage in Table 1 could reflect training choices rather than architecture quality. Please describe the baseline implementations and, ideally, retrain all methods under a common protocol.","section":"Section 4, Table 1"},{"comment":"The conclusions state that 'ablation studies show' the SEM module improves IoU and F1 by 2%-3% and that the cosine annealing schedule outperforms a fixed learning rate, but no ablation table or corresponding experiments appear in Section 4. These claims are currently unsupported. Please include the ablation experiments with concrete numerical results, or remove the claims.","section":"Section 5, Conclusions"},{"comment":"There is a substantive numerical inconsistency: the main-results text reports an average IoU of 80.45% and an average F1 of 88.63%, while Table 2 reports 81.17% and 89.20%. Please correct the discrepancy and ensure all reported numbers are consistent. Additionally, the Experiments section opens by saying 'three distinct and heterogeneous datasets' but then lists and evaluates five datasets; this should be corrected.","section":"Section 4, Main results and Table 2"},{"comment":"The architecture description is not sufficiently detailed for reproducibility. The channel hyperparameters C1-C5 and D1-D5 are named but not given, the KAN grid size and spline order are not specified, and Eq. (6) uses Φ(Z) without defining Φ as the composition of KAN layers from Eqs. (7)-(9). Please provide a complete specification in the text or explicitly point to the released code with the exact configuration used for the reported results.","section":"Section 3, Equations (6)-(9)"}],"minor_comments":[{"comment":"The keyword 'Manba' appears to be a typo for 'Mamba'.","section":"Keywords"},{"comment":"Several citations appear to point to the wrong papers: [16] is cited as the source of Mamba/S6 but is 'Hungry Hungry Hippos'; [17] is cited as S4 but is U-Mamba; [20] is cited as U-NeXt but is 'White-Box Transformers via Sparse Rate Reduction'. Please recheck all references against the text.","section":"References"},{"comment":"The GlaS dataset description states it 'comprises 612 standard-definition frames derived from 31 sequences' and then says 'we used 165 images following established protocols'; this is confusing because the standard GlaS benchmark contains 165 images. Please clarify the dataset split and the source of the 612-frame count.","section":"Section 4, Datasets"},{"comment":"The explainability experiment reports qualitative heatmap comparisons and refers to changes in IoU, but it does not provide the actual IoU values or a quantitative protocol. Please report the numbers or clarify that the comparison is qualitative.","section":"Section 4, Explainability"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The architectural idea is interesting and the paper is positioned in a timely area, but the experimental reporting in the current version is below the standard needed to support the central claims. The missing baseline protocols, missing error bars, inconsistent averages, and unsupported ablation claims are load-bearing issues. I recommend major revision rather than rejection because the deficiencies are fixable and the architecture may be valuable if re-evaluated rigorously."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a straightforward combination of KAN and Mamba-style SSM blocks into a U-Net, with a four-direction scan and a small attention module. The numbers in Table 1 are plausible but the paper does not yet show that they come from the architecture rather than from training-protocol differences. I would send it to review, but with clear instructions to add the missing experimental support.\n\nWhat is actually new is the specific combination. U-KAN and VM-UNet each use one of the two ingredients, and I have not seen this particular fusion plus the SEM module. The efficiency claim (7.35M parameters, 17.66 Gflops, 81.17 average IoU) is a useful data point if it holds. Releasing the code is credit too, even though I did not run it.\n\nThe paper also does some things right. It tests on five standard datasets, reports both IoU and F1, and the baseline list includes recent relevant work like U-Mamba and Rolling-UNet. The explainability section is thin but at least it engages with a mechanism.\n\nThe soft spots are mostly in experimental reporting, and they are load-bearing. The text says results are averaged over three runs, but Table 1 has no error bars. The conclusion cites ablation studies for the SEM module, cosine schedule, and KAN, but there is no ablation table anywhere in the manuscript. Baseline protocols for U-Net, Att-UNet, UNet++, U-NeXt, Rolling-UNet, and U-Mamba are not described; only KM-UNet's own settings appear. Several margins are small (BUSI IoU 65.42 vs 61.81 for U-Mamba; ISIC17 84.05 vs 82.14 for Rolling-UNet), so run-to-run variance could easily flip the ranking. There are also citation problems: reference [31] is supposed to be U-Net but points to Recurrent Residual U-Net, and [33] is called Att-UNet but points to a MICCAI labeling challenge. The paper says 'three heterogeneous datasets' when there are five.\n\nOn the architecture side, the 'rotation from the outer layers toward the center' scan is described vaguely; equation (1) just lists standard directions. If part of the novelty rests on that scan, it needs a proper definition.\n\nWho this is for: people building efficient U-Net variants for medical segmentation. It is a useful benchmark point, not a breakthrough. With the missing ablations and error bars added, and the citations fixed, this could become a solid contribution. As it stands, I would treat the rankings as suggestive rather than proven.","headline":"Plausible incremental architecture with promising efficiency numbers, but the empirical claims currently outrun the experimental reporting.","tokens_in":9989,"tokens_out":2260,"would_cite":false,"duration_ms":22239,"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":"KM-UNet claims that fusing KAN layers with Mamba-style selective scanning in a U-Net beats established U-Net variants on five medical segmentation benchmarks while staying small and fast.","keywords":["Kolmogorov-Arnold Networks","Mamba","state-space models","U-Net","medical image segmentation","selective-scan attention","efficient deep learning","interpretability"],"falsifier":"Retrain U-Net, Att-UNet, UNet++, U-NeXt, Rolling-UNet, and U-Mamba using exactly KM-UNet's protocol (Adam, initial learning rate 1e-4 with cosine decay to 1e-5, batch size 8, 300 epochs, basic rotation and flipping augmentation, identical 80/20 splits) and average over at least three seeds; if the average IoU margins on ISIC17, ISIC18, CVC, BUSI, and GLAS shrink to within noise or reverse, the paper's central accuracy claim fails.","tokens_in":8930,"feed_emoji":"🩺","tokens_out":8130,"duration_ms":69663,"temperature":0.7,"pith_summary":"KM-UNet proposes a U-shaped segmentation network that fuses Kolmogorov-Arnold Networks (KANs) with Mamba-style state-space models, aiming to get long-range context and interpretable feature learning without Transformer-scale compute. The paper reports that this hybrid reaches the highest IoU and F1 scores on most of five medical benchmarks — ISIC17, ISIC18, CVC, BUSI, and GLAS — with an average IoU of 81.17, average F1 of 89.20, 7.35M parameters, and 17.66 Gflops. The author's point is that KAN-plus-SSM is a viable and efficient alternative to CNN and Transformer backbones for medical image segmentation. A sympathetic reader would care because it stakes out a concrete design point combining interpretability, long-range dependency, and low computational cost in one architecture.","feed_headline":"KAN-Mamba U-Net tops five medical segmentation benchmarks","feed_subtitle":"It reports 81.17 average IoU and 89.20 F1 with 7.35M parameters and 17.66 Gflops — accuracy without heavy compute.","key_machinery":"The load-bearing mechanism is the SEM (Selective-Scan Efficient Multi-scale) attention module paired with tokenized KAN blocks. SEM unfolds the feature map into sequences along four scan directions, passes each through an Mamba-derived S6 block, merges them back, then runs multi-scale attention through parallel 1×1 and 3×3 convolutions. The Tok-KAN phase replaces MLP linear transforms with learnable activation functions based on the Kolmogorov-Arnold representation theorem, using a residual connection of the form $Z' = \\text{LN}(Z + \\text{DwConv}(\\Phi(Z)))$. This carries the argument because the four-direction selective scan supplies long-range dependency, while KAN supplies parameter-efficient nonlinear feature representation and the claimed interpretability.","core_discovery":"The central claim is that KM-UNet, by replacing MLP layers in the U-Net bottleneck with tokenized KAN blocks and inserting the Selective-Scan Efficient Multi-scale (SEM) attention module in encoder and decoder, outperforms established U-Net variants (U-Net, Att-UNet, UNet++, U-NeXt, Rolling-UNet, U-Mamba) on most datasets while keeping cost low at 7.35M parameters and 17.66 Gflops. The paper also claims that the KAN layer improves explainability, with channel-activation heatmaps aligning better to ground-truth masks than without it, and that the rotation-style scanning strategy helps on single-object or simpler-shape tasks. The author frames this as the first medical-image segmentation framework integrating KANs and SSMs, providing a baseline and insights for efficient, interpretable segmentation.","pith_inferences":["If the reported margins are real, a natural next test is ablating the four-direction scan down to one direction on BUSI and ISIC to quantify exactly how much of the gain comes from the rotation strategy versus the KAN bottleneck.","The same KAN-SSM design could plausibly be lifted to 3D medical segmentation, where Mamba has already shown strength and where the 17.66 Gflops budget would matter even more.","Follow-up work will likely check whether the components are swappable — for instance, different KAN basis functions or different SSM scan orders — and whether the interpretability advantage survives outside the reported heatmaps."],"forward_implications":["If the reported numbers hold, KAN-SSM hybrids become a credible middle ground between CNNs and Transformers for medical image segmentation.","The architecture's 7.35M parameters and 17.66 Gflops imply it could run in resource-constrained or real-time clinical settings where heavier SSM baselines such as U-Mamba (86.3M parameters, 2087 Gflops in Table 2) are impractical.","The KAN layer's claimed interpretability, if reproducible, would give clinicians a feature-activation map that tracks target boundaries more faithfully than an MLP bottleneck.","The rotation-from-outer-layers scanning strategy suggests a cheap accuracy lever for single-object and simple-shape segmentation tasks."],"supporting_citations":[{"why":"Supplies the Mamba S6 selective-scan block that the SEM feature extraction adapts.","marker":"[23]"},{"why":"Provides the SSM+CNN U-Net baseline (U-Mamba) that KM-UNet compares against and improves on.","marker":"[17]"},{"why":"Establishes KAN as a medical-image-segmentation backbone, which KM-UNet extends by fusing KAN with SSM.","marker":"[35]"},{"why":"Demonstrates a Vision-Mamba U-Net, defining the prior art for Mamba-in-U-Net that KM-UNet builds on.","marker":"[36]"},{"why":"The U-Net baseline whose Table 1 accuracy numbers KM-UNet's advantage is measured against.","marker":"[31]"},{"why":"The Rolling-UNet MLP baseline that anchors the efficiency comparison (params/Gflops) in Table 2.","marker":"[34]"}],"fun_headline_variants":["KAN and Mamba join forces in new U-Net for medical segmentation","KM-UNet fuses KAN and state-space models for efficient segmentation","First KAN-SSM segmentation model matches SOTA on five benchmarks","Lightweight KAN-Mamba U-Net rivals heavier medical segmenters","KAN-Mamba U-Net: accurate medical segmentation at low cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ranking in Table 1 is taken to reflect architecture quality, but the paper reports KM-UNet's own optimizer, learning rate, batch size, augmentation, and 300-epoch schedule without stating matching protocols for the six baselines, and reports no seeds or error bars; if the baselines were trained under different conditions, the accuracy advantage could vanish.","fun_headline_variants_meta":{"raw":{"variants":["KAN and Mamba join forces in new U-Net for medical segmentation","KM-UNet fuses KAN and state-space models for efficient segmentation","First KAN-SSM segmentation model matches SOTA on five benchmarks","Lightweight KAN-Mamba U-Net rivals heavier medical segmenters","KAN-Mamba U-Net: accurate medical segmentation at low cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1495,"prompt_tokens":969,"completion_tokens":526,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":429}},"tokens_in":585,"tokens_out":526,"duration_ms":5067,"temperature":1.0,"reasoning_tokens":429,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:09:40.081775+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain U-Net, Att-UNet, UNet++, U-NeXt, Rolling-UNet, and U-Mamba using exactly KM-UNet's protocol (Adam, initial learning rate 1e-4 with cosine decay to 1e-5, batch size 8, 300 epochs, basic rotation and flipping augmentation, identical 80/20 splits) and average over at least three seeds; if the average IoU margins on ISIC17, ISIC18, CVC, BUSI, and GLAS shrink to within noise or reverse, the paper's central accuracy claim fails.","supporting_citations":[{"cited_title":"Mamba: Linear-Time Sequence Modeling with Selective State Spaces","cited_arxiv_id":null,"evidence_quote":"Supplies the Mamba S6 selective-scan block that the SEM feature extraction adapts."},{"cited_title":"U-Mamba: Enhancing Long-Range Dependency for Biomedical Image Segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the SSM+CNN U-Net baseline (U-Mamba) that KM-UNet compares against and improves on."},{"cited_title":"Journal of Medical Imaging 6(1), 014006– 014006 (2019)","cited_arxiv_id":null,"evidence_quote":"The U-Net baseline whose Table 1 accuracy numbers KM-UNet's advantage is measured against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Rolling-UNet MLP baseline that anchors the efficiency comparison (params/Gflops) in Table 2."}],"review_version":1}