{"id":"2a7a456d-2a9f-4d19-8934-0ac37e1c92d9","arxiv_id":"2411.11926","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"KAN-Mamba FusionNet, a U-Net variant with a KAMBA block, reports higher IoU and F1 than five segmentation baselines on BUSI, Kvasir-Seg, and GlaS, with gains up to 2.2 IoU points.","lead":"A new medical image segmentation network combines KAN layers, Mamba-style state space models, and a learned mix of activation functions. The authors report consistent but modest IoU and F1 improvements over five baselines on three public datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Superiority claim rests on a single unseeded 4:1 validation split with no significance testing; reported IoU gaps on Kvasir-Seg and GlaS are within one standard deviation of Seg. U-KAN.","rationale":"The reader's weakest_assumption identifies the same load-bearing issue: the empirical superiority claim depends entirely on validation scores from an unspecified 4:1 split. My stress-test confirms this is the most important weakness. The paper's internal logic is coherent; I found no mathematical inconsistency in the KAMBA block formulation or the loss definition. However, the central claim is not supported at the required level because the evaluation is uncontrolled: no fixed seed, no separate test set, no significance testing, and no evidence that baselines were tuned comparably. The small margins on two of three datasets, relative to the reported standard deviations, make the 'consistently outperforms' statement fragile. The ablations are directionally sensible but do not isolate spatial attention and skip connections, which is a secondary gap. Overall, the CONDITIONAL verdict is appropriate: the architecture is plausible and the results may be reproducible, but the evidence as presented does not establish a robust improvement. I recommend no change to the reader's verdict.","tokens_in":8512,"tokens_out":2897,"duration_ms":30017,"concrete_test":"Fix five random 4:1 splits with recorded seeds. Train KAN-Mamba FusionNet and Seg. U-KAN (plus U-Mamba) with the same per-dataset hyperparameters from Section 4.1. On each split, evaluate the held-out 20%, compute per-image IoU and F1, and form paired differences between methods. Report bootstrap 95% confidence intervals and paired Wilcoxon p-values for each dataset. If the Kvasir-Seg or GlaS intervals include zero, the 'outperforms' claim should be softened to 'comparable on these datasets'; if intervals exclude zero on all three, the concern is resolved. Publishing code and seeds would allow this check to be reproduced.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1 is the sole basis for the central superiority claim, but the evaluation protocol in Sections 4.1-4.2 leaves the comparison uncontrolled. Each dataset is split once, 4:1 train/validation, with no seed, no held-out test set, and no statement that the validation set was untouched during hyperparameter or architecture selection. The three 'independent runs' appear to repeat training on the same split, so the reported plus-or-minus values capture initialization noise, not split variability. The largest IoU gaps over the strongest baseline, Seg. U-KAN, are 2.2 (BUSI), 1.1 (Kvasir-Seg), and 0.28 (GlaS); on Kvasir-Seg and GlaS, the gap is smaller than the pooled standard deviation, so without a paired significance test the claim 'consistently outperforms' is not established. The baselines are also not shown to be tuned on the same validation split, so part of the margin could be baseline under-tuning. This is load-bearing because every quantitative conclusion in the paper rests on these validation scores.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes KAN-Mamba FusionNet, a U-Net-like segmentation architecture that inserts a newly designed KAMBA block—comprising a KAN layer, a bag-of-activations (BoA) module, a state-space model (SSM), a spatial attention module, and a skip connection—into the U-KAN pipeline. The model is evaluated on three public medical segmentation datasets (BUSI, Kvasir-Seg, GlaS) and compared against U-Net, U-NeXt, Rolling-UNet, U-Mamba, and Seg. U-KAN using IoU and F1 scores (Table 1). The paper also reports model complexity (Table 2) and ablation studies on the BUSI dataset (Tables 3-4). The central claim is that the proposed architecture consistently outperforms state-of-the-art methods on these datasets.","tokens_in":8713,"tokens_out":5460,"duration_ms":47975,"significance":"If the reported gains hold under a statistically controlled evaluation, the architecture would be a useful incremental contribution to the line of work combining Kolmogorov-Arnold Networks and Mamba for medical image segmentation. The paper's strengths are the use of three public datasets, a clear description of the proposed KAMBA block, and the inclusion of complexity comparisons and ablations. However, the empirical claim currently rests entirely on validation scores from a single unseeded split per dataset, with no significance testing and no evidence that the baselines were tuned with the same care. The paper does not provide code or pretrained models, which limits reproducibility. These issues are load-bearing because every quantitative conclusion in the abstract and conclusion depends on the validity of the Table 1 comparison.","major_comments":[{"comment":"The central claim that KAN-Mamba FusionNet 'consistently outperforms' the baselines is not established by the reported protocol. Each dataset is split once into a 4:1 train/validation ratio with no fixed seed and no separate held-out test set, and the three 'independent runs' repeat the same split, so the reported plus-or-minus values capture only initialization noise, not split variability. On Kvasir-Seg and GlaS, the IoU advantage over Seg. U-KAN is 1.10 and 0.28 points, respectively, while the reported standard deviations for the baseline are 0.53 and 0.34; without a paired significance test, these margins are not sufficient to support the claim of consistent superiority. I recommend reporting test-set metrics from multiple independent splits (e.g., five random seeds or k-fold cross-validation) with paired bootstrap or Wilcoxon signed-rank tests, and clearly distinguishing validation from test performance.","section":"Sections 4.1-4.2, Table 1"},{"comment":"The comparison against baselines is uncontrolled with respect to training protocol. The paper does not state whether U-Net, U-NeXt, Rolling-UNet, U-Mamba, and Seg. U-KAN were trained with the same optimizer, learning rate schedule, loss function, data augmentation, and number of epochs, nor whether their hyperparameters were tuned on the same validation splits. Without this information, part of the reported margin may be attributable to under-tuned baselines rather than to the proposed architecture. The authors should provide the exact training configuration used for every baseline, ideally with the same hyperparameter search budget, so that the comparison is fair.","section":"Section 4.1, Table 1"},{"comment":"The ablation studies are performed only on the BUSI dataset and report single-run values without standard deviations. For example, the improvement of BoA over a single ReLU is +1.23 IoU (from 64.61 to 65.84), which is comparable to the run-to-run noise reported for the full model in Table 1 (±0.54 IoU on BUSI). Additionally, the 'Classical Mamba' baseline in Table 3 is not defined (which Mamba variant, what backbone), making the ablation irreproducible. Please repeat the ablations over multiple seeds, report means and standard deviations, and define the classical Mamba reference precisely.","section":"Section 4.4, Tables 3-4"}],"minor_comments":[{"comment":"The BUSI dataset description is ambiguous: 'We utilized the images representing breast cancer' could mean excluding the 133 normal images, but this is not stated explicitly; please specify the exact number of images used per class and how the split was applied.","section":"Section 3"},{"comment":"The symbol ⊕ in Equation (6) is not defined; please state that it denotes concatenation and clarify how the channel dimensions of Z''_out, Ψ(Zin), and Zin are aligned.","section":"Section 2.2, Eq. (6)"},{"comment":"The summation notation in Equation (3) appears garbled in the typeset version ('P ∑ p=1'); please correct the notation and state how the constraint Σαp=1 is enforced during optimization (e.g., softmax normalization) and how αp is initialized.","section":"Section 2.2, Eq. (3)"},{"comment":"The term 'Classical Mamba' in Table 3 is not defined; specify which Mamba architecture and backbone are used so the ablation can be reproduced.","section":"Section 4.4"},{"comment":"Reference [18] points to Kiu-Net rather than the original GlaS dataset paper (Sirinukunwattana et al., MICCAI 2015); please cite the original dataset source.","section":"References"},{"comment":"The GFLOPs numbers depend on the input resolution and batch size; please specify these settings so that the complexity comparison is meaningful.","section":"Table 2"},{"comment":"The terms 'consistently outperforms' (abstract and conclusion) and 'significantly fewer false positive regions' (Section 4.3 and Fig. 2 caption) are stronger than the statistical evidence supports; please soften these statements or provide quantitative false-positive analysis.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a medical image segmentation venue, but the evaluation falls short of current standards: validation-only results from a single unseeded split, no significance testing, and no evidence of baseline tuning. The margins over the strongest baseline are small on two of the three datasets, so the claimed superiority is not yet convincing. I would encourage the editor to request the code and full experimental details upon resubmission, as the architecture itself is plausible and the paper is clearly written."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the architecture is a legitimate new assembly, and the ablations are internally consistent. The evaluation, however, does not support the headline claim of consistent SOTA gains.\n\nWhat's new: the KAMBA block chains a KAN block, a bag-of-activations layer, an SSM, spatial attention, and a skip connection, then slots it into the U-KAN encoder. That specific combination isn't in the cited prior work, and the ablation on BUSI shows each addition (KAN, then BoA, then the full block) improves IoU and F1. The complexity table is also honest: the model costs essentially the same as Seg. U-KAN (13.96 GFLOPs, 6.37M params vs 13.87/6.36). That is a legitimate engineering contribution.\n\nWhere it gets soft: Table 1 is the whole evidence for \"consistently outperforms,\" and the protocol in Section 4.1 can't support that. One 4:1 random split per dataset, no fixed seed, no held-out test set, no significance test. The three \"independent runs\" are presumably on the same split, so the plus/minus reflects only initialization noise. On Kvasir-Seg the IoU gap to Seg. U-KAN is 1.1 points with overlapping stds; on GlaS it's 0.28, deep inside one std. Only BUSI (2.2 points) looks clearly better. Also, we are not told how baselines were tuned on these splits, so part of the gap could be baseline under-tuning. The title and abstract overstate what the table shows.\n\nMinor soft spots: no code, no seed listing, and the BoA ablation (Table 4) shows 'None' vs 'Single (ReLU)' vs 'BoA' but does not isolate spatial attention or the skip connection separately. Still, the ablation that exists is coherent and supports the general direction.\n\nBottom line: this is a paper for readers interested in U-KAN-style backbones for medical segmentation. It is not a breakthrough, but it is a plausible incremental step. The empirical claim needs stronger evaluation before it can be trusted. I'd send it to a serious referee — the architecture is concrete and testable — but I'd expect a major revision or rejection if the evaluation isn't fixed.\n\nWho gets value: people building on U-KAN or Mamba-based segmentation backbones; also a good negative example for a methods class on why one split is not enough.","headline":"KAMBA block is a real but modest new assembly of known parts; the central performance claim rests on a single uncontrolled validation split.","tokens_in":9270,"tokens_out":2553,"would_cite":false,"duration_ms":23699,"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":"KAN-Mamba FusionNet claims top IoU and F1 across BUSI, Kvasir-Seg, and GlaS by combining learnable nonlinearities with state-space long-range modeling.","keywords":["medical image segmentation","Kolmogorov-Arnold networks","state space models","Mamba","U-KAN","bag of activation functions","spatial attention","KAMBA block"],"falsifier":"Train the five compared methods and KAN-Mamba FusionNet on identical multi-seed splits of BUSI, Kvasir-Seg, and GlaS and compute IoU and F1 on held-out test folds; if Seg. U-KAN or U-Mamba reaches or exceeds the reported numbers on those folds, the central claim fails.","tokens_in":8340,"feed_emoji":"🩺","tokens_out":6634,"duration_ms":54855,"temperature":0.7,"pith_summary":"The paper argues that medical image segmentation can be improved by joining two ideas that have complementary weaknesses: Kolmogorov-Arnold networks capture nonlinear structure through learnable activations but miss long-range context, while state-space models capture long-range dependencies but rely on fixed activations. It assembles both, together with spatial attention and a learned mixture of activation functions, into a KAMBA block placed inside a U-KAN-style encoder-decoder, yielding KAN-Mamba FusionNet. On breast ultrasound (BUSI), polyp (Kvasir-Seg), and gland (GlaS) datasets, the model reports the highest IoU and F1 among U-Net, U-NeXt, Rolling-UNet, U-Mamba, and Seg. U-KAN. The practical interest is that the gains come at a modest computational cost compared with transformer- or Mamba-heavy alternatives.","feed_headline":"KAN-Mamba hybrid tops U-KAN and Mamba on three medical sets","feed_subtitle":"A KAMBA block adds learnable activations and state-space context to U-KAN, lifting IoU and F1 on three datasets.","key_machinery":"The load-bearing object is the KAMBA block, whose output is $Z''_{\\mathrm{out}} \\oplus \\Psi(Z_{\\mathrm{in}}) \\oplus Z_{\\mathrm{in}}$, where $Z''_{\\mathrm{out}}$ is the result of KAN, activation mixture, state-space model, and spatial attention in sequence. The activation mixture, called bag of activation functions (BoA), is $\\Psi(Z) = \\sum_{p=1}^{P} \\alpha_p \\psi_p(Z)$ with $\\sum_p \\alpha_p = 1$ and learnable weights, replacing a single fixed activation. The state-space model supplies linear-complexity long-range dependency modeling, and the spatial attention module uses a $7\\times7$ convolution over pooled features to emphasize boundaries of irregular structures. Together these components are what the paper credits for the reported accuracy gains.","core_discovery":"The central claim is that a single block can jointly handle the nonlinear appearance of anatomical structures and the long-range spatial context needed to segment them, and that doing so improves measured segmentation quality. The KAMBA block processes features through a KAN layer, a learned weighted sum of five activation functions (ReLU, tanh, Softplus, GELU, SiLU), a state-space model, and a spatial attention module, then fuses the result with a parallel activation mixture and the original input through a skip connection. On the three datasets the paper reports IoU/F1 of 65.84/79.17, 76.95/86.66, and 85.73/92.29, each above the corresponding values for all five compared methods. Ablation results show that adding KAN, BoA, and spatial attention to a classical Mamba backbone raises IoU from 60.87 to 65.84 on BUSI.","pith_inferences":["Inference: the paper's own comparison is on validation splits from a single random 4:1 split, so the most direct extension is to measure test-set performance over multiple fixed seeds to see whether the reported margins persist out-of-sample.","Inference: the BoA weights could be inspected after training; if one activation dominates on a given dataset, the bag may be acting mainly as a regularizer rather than as a genuinely diversified nonlinearity.","Inference: the same block structure should transfer to 3D segmentation or video frames, where long-range temporal or volumetric context is even more important and the state-space component has a larger relative role.","Inference: replacing the fixed sigmoid in the spatial attention module with a KAN-style learnable activation would remove the last static nonlinearity, a direct test of whether the nonlinearity argument extends to attention itself."],"forward_implications":["The KAMBA block can be inserted into a U-KAN pipeline and improve validation IoU and F1 on ultrasound, endoscopic, and histology images without adding much computational cost (13.96 GFLOPs and 6.37 million parameters).","The reported margins, from 0.3 to 2.2 IoU points over Seg. U-KAN, suggest that the main benefit is a consistent small gain rather than a large architectural reset.","The ablation evidence implies that each added component, KAN, BoA, and spatial attention, contributes positively, with the full KAMBA block outperforming every intermediate configuration.","Because the model uses a linear-complexity state-space model, it is a candidate for segmenting large images or 3D volumes where transformer attention would be too expensive."],"supporting_citations":[{"why":"Provides the U-KAN backbone that KAN-Mamba FusionNet extends and is the strongest baseline it must beat.","marker":"[11]"},{"why":"Supplies the Kolmogorov-Arnold network construction used in the KAN block.","marker":"[14]"},{"why":"Supplies the selective state-space model that gives the block long-range dependency modeling.","marker":"[6]"},{"why":"Earlier Mamba-based segmentation baseline and source of the state-space component for biomedical images.","marker":"[15]"},{"why":"Defines the spatial attention module used to extract fine-grained boundary features.","marker":"[20]"},{"why":"The breast ultrasound dataset (BUSI) on which the model and ablations are evaluated.","marker":"[1]"},{"why":"The polyp segmentation dataset (Kvasir-Seg) used as a second evaluation benchmark.","marker":"[9]"},{"why":"The gland segmentation dataset (GlaS) used as a third evaluation benchmark.","marker":"[18]"},{"why":"The U-Net baseline that anchors the comparison family.","marker":"[16]"}],"fun_headline_variants":["KAN-Mamba FusionNet beats five methods on three medical sets","New KAMBA block boosts medical segmentation IoU and F1","Fusing KAN and Mamba improves medical image segmentation","One block to rule them: KAN-Mamba tops medical segmentation","KAN-Mamba FusionNet: Better non-linearity and long-range context"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed superiority rests on validation scores from a single unseeded 4:1 random split of each dataset, so if that split is unrepresentative, the reported advantages may not reproduce.","fun_headline_variants_meta":{"raw":{"variants":["KAN-Mamba FusionNet beats five methods on three medical sets","New KAMBA block boosts medical segmentation IoU and F1","Fusing KAN and Mamba improves medical image segmentation","One block to rule them: KAN-Mamba tops medical segmentation","KAN-Mamba FusionNet: Better non-linearity and long-range context"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000955,"raw_usage":{"total_tokens":4094,"prompt_tokens":990,"completion_tokens":3104,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":3013}},"tokens_in":606,"tokens_out":3104,"duration_ms":21764,"temperature":1.0,"reasoning_tokens":3013,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:32:55.512815+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the five compared methods and KAN-Mamba FusionNet on identical multi-seed splits of BUSI, Kvasir-Seg, and GlaS and compute IoU and F1 on held-out test folds; if Seg. U-KAN or U-Mamba reaches or exceeds the reported numbers on those folds, the central claim fails.","supporting_citations":[{"cited_title":"Gu and T","cited_arxiv_id":null,"evidence_quote":"Supplies the selective state-space model that gives the block long-range dependency modeling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the spatial attention module used to extract fine-grained boundary features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The polyp segmentation dataset (Kvasir-Seg) used as a second evaluation benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The gland segmentation dataset (GlaS) used as a third evaluation benchmark."},{"cited_title":"Ronneberger, P","cited_arxiv_id":null,"evidence_quote":"The U-Net baseline that anchors the comparison family."}],"review_version":1}