{"id":"393d59a7-410f-4d93-b9dd-59a58e8df3c3","arxiv_id":"2504.13754","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"CMSwinKAN, a lightweight Swin-KAN hybrid with contrastive multi-scale fusion, reportedly outperforms pathology foundation models on neuroblastoma classification and achieves 100% WSI accuracy on a private dataset.","lead":"A new model called CMSwinKAN combines Swin Transformers, KAN layers, and multi-scale contrastive learning to classify neuroblastoma tissue patches and whole-slide images. The authors report that their lightweight model, which needs no pretraining, beats large pathology foundation models and reaches 100% accuracy on their private WSI test set.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"WSI-level superiority rests on an unverifiable soft-vote protocol and an internally inconsistent Table IV; the 100% result is not independently established.","rationale":"The paper's most important claimed advantage is at WSI level: a 1.32M-parameter randomly initialized model with a clinical-prior soft vote reaching 100% on a private pediatric tumor WSI test set. That claim is what would justify clinical deployment and what distinguishes the method from ordinary patch classifiers. The mechanism producing it has three related red flags: the SVM is trained on CMSwinKAN's own features with no description of labels, counts, or validation; the text states alpha > beta but the experiments set alpha=1 and beta=8, so the soft-vote formula cannot be applied as described; and Table IV's identical metrics across 13 methods are numerically implausible and suggest the table was not derived from actual per-model predictions. The patch-level Table III and the ablation study are internally coherent, and the BreakHis results are plausible, so the architecture and patch-level experiments are not the main problem; the unsupported WSI-level comparison is the load-bearing failure. This supports the reader's REJECT verdict, though the emphasis here is shared between the missing SVM protocol and the implausible baseline table rather than on the SVM alone, hence partial agreement with the reader's stated weakest assumption.","tokens_in":14810,"tokens_out":7801,"duration_ms":74971,"concrete_test":"Reproduce Table IV using the released GitHub code: for each baseline and for CMSwinKAN-micro, extract patch probabilities on the 63 test WSIs, implement hard voting by majority label, and implement soft voting by training the tissue SVM only on training-WSI patches with pathologist-provided tissue labels, choosing alpha, beta, and gamma on a held-out validation set before seeing test labels. Record per-model ACC, BACC, KAPPA, and F1. If the 13 baseline rows are not all exactly 98.41/98.33/98.01/98.39, or if CMSwinKAN's soft-vote accuracy is not 100%, Table IV and the WSI-level conclusion are not reproducible as reported.","verdict_should_be":"REJECT","load_bearing_attack":"The central WSI-level claim depends on a soft-voting procedure that cannot be checked from the manuscript and on a comparison table that cannot be true as printed. Section III.E.2 states the SVM is trained on features extracted by CMSwinKAN and says alpha > beta, 'emphasizing the significance of neuropil', but Section IV.C sets alpha=1 and beta=8; the two statements contradict each other, so the clinical-prior weighting is not well-defined. The same section promises details of SVM training that never appear: no tissue-component labels, no annotation protocol, no SVM train/test split, and no SVM accuracy. Independent of that, Table IV reports exactly ACC=98.41, BACC=98.33, KAPPA=98.01, and F1=98.39 for 13 different methods under both hard and soft voting. With 63 test WSIs, identical four-metric tuples across heterogeneous models, including pathology foundation models and small randomly initialized CNNs, cannot arise from independent patch classifiers. The 100% soft-vote row for CMSwinKAN is therefore not supported by reproducible evidence: the SVM module is unspecified, the weights are contradictory, and the baseline table is implausible. Because the abstract's headline superiority at WSI level rests on that row, the central claim fails as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CMSwinKAN, a lightweight Swin-Transformer variant in which MLP layers are replaced by Kolmogorov-Arnold Network (KAN) layers, augmented with a contrastive multi-scale feature aggregation module (CMSA/CDFA), for patch-level classification of H&E histopathology images. The authors evaluate it on a private pediatric peripheral neuroblastic tumor dataset (PpNTs) and on the public BreakHis dataset, and they introduce a soft voting mechanism with an SVM tissue-component classifier to aggregate patch predictions into whole-slide image (WSI) labels. The central claims are that randomly initialized CMSwinKAN-micro outperforms pathology foundation models UNI, CONCH, and CTransPath at the patch level, and that with the proposed soft voting it reaches 100% accuracy, balanced accuracy, and KAPPA on the PpNTs WSI test set (Table IV).","tokens_in":15080,"tokens_out":3735,"duration_ms":35881,"significance":"If the claims were fully substantiated, the paper would be noteworthy: a 1.32M-parameter model trained from random initialization outperforming large pathology foundation models would matter for resource-constrained deployment, and the interpretability angle through KAN layers is timely. The authors also state that source code is available and they report an ethics-approved private dataset, which are positive features. However, the load-bearing WSI-level result is not currently verifiable: the SVM-based soft voting module is underspecified, the reported weighting coefficients contradict the method text, and Table IV reports identical four-metric tuples across many heterogeneous models, which is implausible for independent classifiers. These issues directly affect the headline superiority claim at the WSI level, so the paper as submitted does not meet the standard for acceptance.","major_comments":[{"comment":"The identical tuple (ACC=98.41, BACC=98.33, KAPPA=98.01, F1=98.39) is reported for numerous diverse method variants, including ResNet-50, DenseNet, ViT, Swin, UNI, CTransPath, CONCH, FMDNN, StoHisNet, HiFuse, and CMSwinKAN-micro under hard voting, with hard and soft voting giving the same numbers for most models. With only 63 test WSIs, identical four-metric tuples across such heterogeneous classifiers cannot plausibly arise from independent evaluations, so the table as printed cannot support the WSI-level comparison. The authors must report per-model confusion matrices or exact test-set predictions, and explain how these identical values were produced.","section":"§IV.D.1, Table IV"},{"comment":"The method text states that α > β, 'emphasizing the significance of neuropil,' but Section IV.C sets α = 1 for neuropil and β = 8 for Schwannian stroma, so the implemented weighting is β > α. This contradiction makes the clinical-prior weighting ill-defined, and Eq. (11) as configured would give Schwannian stroma eight times the weight of neuropil, opposite to the stated rationale. The authors must reconcile the text with the implementation and provide a sensitivity analysis over α and β, ideally with the weights chosen on a validation set rather than the test set.","section":"§III.E.2 vs. §IV.C"},{"comment":"The WSI soft voting depends on a multi-class SVM that recognizes neuropil, Schwannian stroma, and other tissue components, but the manuscript does not describe the SVM input features, the tissue-component labels, the annotation protocol, the number of annotated patches, the train/test split, or the SVM's own classification accuracy. Section III.D explicitly says 'Further details on the voting mechanism and SVM training will follow,' yet no such details appear anywhere in the paper. Without this information, the 100% soft-vote result for CMSwinKAN-micro in Table IV cannot be reproduced or independently checked, and the WSI-level claim is not established.","section":"§III.D and §III.E.2"},{"comment":"The patch-level improvements over UNI, CONCH, and CTransPath are reported only as point estimates, with no confidence intervals, significance tests, or per-patient statistics. Given the small test sets, differences of 0.2–0.3% in ACC and BACC are within plausible sampling noise, so the claim that a 1.32M-parameter randomly initialized model beats large pretrained pathology foundation models needs statistical substantiation.","section":"§IV.C, Table III"}],"minor_comments":[{"comment":"The sentence 'Further details on the voting mechanism and SVM training will follow' is inappropriate in a final manuscript; the details should either be fully integrated into Section III.E.2 or clearly stated as future work rather than deferred within the same paper.","section":"§III.D"},{"comment":"The term 'contrastive learning' appears in the model name and contribution list, but no contrastive loss is defined; the CDFA module compares attention maps across scales, which is not standard contrastive learning. Please clarify the terminology or specify the actual contrastive objective.","section":"§III.C"},{"comment":"The notation is not fully connected: w_{i,c} in Eq. (9) is not explicitly linked to w_patch in Eq. (11), and P1 and P2 are introduced without explaining how the SVM's multiclass probabilities are obtained or normalized. Please unify the notation and define all variables.","section":"Eqs. (8)–(11)"},{"comment":"The circle-size encoding of parameter count is hard to read because many points overlap; a table listing parameter counts and FLOPs for all compared models would be much clearer.","section":"Fig. 5"},{"comment":"The text says soft voting improves generalization on BreakHis, but Table V reports hard-voting results only; please either add soft-vote results for BreakHis or adjust the wording.","section":"Table V"}],"recommendation":"reject","confidential_remarks":"The paper appears to be an incomplete submission: Section III.D explicitly defers details of the SVM voting module, and Table IV contains an implausible pattern of identical metric tuples across heterogeneous models. These are not merely presentational gaps; they affect the central WSI-level claim of 100% accuracy. The authors would need to substantially revise the evaluation protocol and provide the missing SVM details for the result to be assessable. I would encourage a resubmission after these issues are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on arXiv:2504.13754. The paper is not ready as it stands, but it's not empty: the CMSwinKAN idea—replacing MLPs with KAN layers in a Swin backbone and adding contrastive multi-scale aggregation—is a plausible way to build a very small model for pediatric tumor patch classification. Patch-level numbers on PpNTs and BreakHis are not outlandish, and the ablation ties each module to a small gain. That part is worth a look.\n\nThe problems start at WSI-level. Table IV is not believable: thirteen different models, ranging from VGGNet to UNI/CONCH, all report exactly the same four metrics (98.41 ACC, 98.33 BACC, 98.01 KAPPA, 98.39 F1) under both hard and soft voting. On 63 test WSIs that cannot happen from independent classifiers. Either the table was copied in error or the comparison is degenerate. The text also says the soft-vote result for CMSwinKAN was 98.41, while Table IV shows 100.00; the two don't agree. And the clinical-prior weighting is self-contradictory: Section III.E.2 says alpha > beta to emphasize neuropil, but Section IV.C sets alpha=1, beta=8, which is the opposite. Meanwhile the SVM that drives the soft voting is never described: no tissue-label annotations, no train/test split, no SVM accuracy, only a promise that details 'will follow.' That is a load-bearing part of the method, not a footnote.\n\nSo the headline WSI superiority is not reproducible. The patch-level claim might be salvageable if the tables are corrected and the SVM protocol is fully specified, but as printed the evidence is too shaky to build on. I would flag this for verification before trusting any number in Table IV.\n\nAll that said, I'd still send it to peer review: the architecture is novel enough that a serious referee is needed to unpick whether the patch-level gains hold under proper evaluation, and the errors in Table IV need to be publicly corrected or explained. A desk rejection would bury the salvageable part. My inclination is a major-revision recommendation, pending a full explanation of Table IV and the SVM setup.","headline":"A plausible lightweight architecture with patch-level promise, but the WSI-level claims rest on an impossible comparison table and an under-specified, internally inconsistent voting protocol.","tokens_in":15636,"tokens_out":3507,"would_cite":false,"duration_ms":30252,"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":"A 1.32M-parameter, randomly initialized model is claimed to beat UNI, CONCH, and CTransPath on neuroblastoma patch classification, and to reach 100% WSI accuracy with soft voting.","keywords":["neuroblastoma","whole-slide image classification","Kolmogorov-Arnold Network","Swin Transformer","contrastive learning","multi-scale feature fusion","soft voting","histopathology"],"falsifier":"Re-run the WSI experiment with the tissue SVM trained on a held-out set of manually annotated neuropil and stroma patches, report the SVM's own accuracy, and freeze α, β, and γ before seeing the test WSIs; if the SVM is inaccurate or the frozen weights do not reproduce 100%, the central WSI claim fails.","tokens_in":14611,"feed_emoji":"🔬","tokens_out":4977,"duration_ms":45540,"temperature":0.7,"pith_summary":"This paper tries to establish that a lightweight model trained from scratch, without pretraining on massive pathology datasets, can outperform large pretrained pathology foundation models on neuroblastoma subtype classification. It argues that replacing the MLP layers in a Swin Transformer with Kolmogorov-Arnold Network (KAN) layers, combined with a contrastive multi-scale feature fusion module, improves both accuracy and interpretability. The paper also introduces a clinically guided soft voting mechanism intended to turn patch-level predictions into accurate whole-slide diagnoses. If these claims hold, small hospitals with limited data and computing resources could deploy accurate pediatric tumor classifiers without depending on large pretrained models.","feed_headline":"Random-start 1.3M-parameter model tops pathology giants","feed_subtitle":"CMSwinKAN adds KAN layers and multi-scale contrastive fusion to reach 93.58% patch accuracy and 100% WSI accuracy on neuroblastoma.","key_machinery":"The central mechanism is a Swin Transformer variant in which every MLP is replaced by a Kolmogorov-Arnold Network (KAN) layer, whose activations are learnable B-spline functions. A contrastive-driven multi-scale aggregation module (CMSA), with its CDFA submodule, aligns feature maps from the first three transformer stages and uses cross-scale attention to weight local windows, while the KAN classification head fuses stage-4 residuals with the aggregated multi-scale features. For whole-slide diagnosis, a multi-class SVM trained on the model's features recognizes neuropil and Schwannian stroma, and a formula converts those tissue probabilities into patch voting weights, with hand-set coefficients α=1, β=8, and γ=1.","core_discovery":"The paper claims that CMSwinKAN-micro, with only 1.32 million parameters and random initialization, reaches 93.58% accuracy on the five-class PpNTs patch classification task, beating UNI (93.35%), CONCH (92.17%), and CTransPath (92.22%), and reaches 87.79% accuracy on the public BreakHis dataset, again above the same baselines. With the proposed soft voting, CMSwinKAN-micro reaches 100% accuracy, balanced accuracy, and KAPPA on the private PpNTs whole-slide test set, while most baselines sit at 98.41% under hard or soft voting. The ablation studies attribute a 1.29 percentage point accuracy gain to the KAN-for-MLP replacement alone, and further gains to the contrastive multi-scale fusion components.","pith_inferences":["The perfect 100% WSI figure likely reflects a ceiling effect: nearly every baseline already reaches 98.41%, so the reported margin is concentrated in a few test slides, and a larger or harder test set would better separate methods.","Because the SVM that drives soft voting is trained on CMSwinKAN's own features, the WSI result is not an independent test of the patch classifier unless the SVM is trained on separate tissue annotations and its accuracy is reported.","The interpretability claim is architectural, resting on KAN splines, rather than demonstrated by a quantitative explanation metric; a follow-up examining which spline bases activate for neuropil versus stroma would test it.","The BreakHis WSI comparison uses CMSwinKAN-tiny while the patch experiments use CMSwinKAN-micro, so the cross-dataset gains mix model size with method; testing the same variant everywhere would isolate the contribution."],"forward_implications":["A randomly initialized 1.32M-parameter model can reach or exceed the patch-level accuracy of foundation models pretrained on over a million pathology images on a five-class pediatric tumor task.","KAN-based nonlinearity can serve as a drop-in replacement for MLPs in Swin-style blocks, with ablations attributing about 1.3 accuracy points and 1.2 KAPPA points to that replacement.","The clinical-prior soft voting scheme, if reproducible, lifts whole-slide accuracy to a perfect 100% on the private test set and also improves weaker models such as VGGNet.","The model's small size (0.22 GFLOPs) makes it feasible to run on edge devices, addressing a practical barrier to clinical deployment.","The ablation results indicate that each proposed component, KAN, CDFA, and CMSA, contributes positively, implying the full architecture is needed for the reported gains."],"supporting_citations":[{"why":"Supplies the Kolmogorov-Arnold Network layer that replaces MLPs in the Swin blocks and classification head.","marker":"[33]"},{"why":"Provides the UNI foundation model baseline that the paper claims to outperform.","marker":"[19]"},{"why":"Provides the CONCH vision-language foundation model baseline that the paper claims to outperform.","marker":"[20]"},{"why":"Provides the CTransPath baseline, a CNN-transformer hybrid pretrained on pathology, that the paper claims to outperform.","marker":"[18]"},{"why":"Supplies the Swin Transformer backbone that the paper modifies into the Swin KANsformer.","marker":"[39]"},{"why":"Provides the HiFuse multi-scale fusion baseline and motivates the CMSA module's global-local fusion design.","marker":"[30]"},{"why":"Provides the FMDNN pathology classification baseline that the paper compares against on both datasets.","marker":"[23]"},{"why":"Introduces the contrast-driven feature enhancement idea that the CDFA submodule is built upon.","marker":"[32]"}],"fun_headline_variants":["Tiny random-start model beats pathology giants with 100% WSI accuracy","1.3M-param model, random init, hits 100% on neuroblastoma WSI","Random-start micro model tops UNI, CONCH on five-class tumor typing","Small interpretable model beats pretrained giants on pathology"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole-slide 100% result rests on the soft-voting SVM that recognizes neuropil and Schwannian stroma, but the paper does not specify how that SVM is trained, on what annotations, or how its accuracy is validated; if the SVM is trained on the same model's features or the weights α=1, β=8, γ=1 are tuned on the test set, the perfect WSI accuracy would not be an independent result.","fun_headline_variants_meta":{"raw":{"variants":["Tiny random-start model beats pathology giants with 100% WSI accuracy","1.3M-param model, random init, hits 100% on neuroblastoma WSI","Random-start micro model tops UNI, CONCH on five-class tumor typing","Small interpretable model beats pretrained giants on pathology"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000782,"raw_usage":{"total_tokens":3475,"prompt_tokens":991,"completion_tokens":2484,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":2400}},"tokens_in":607,"tokens_out":2484,"duration_ms":17787,"temperature":1.0,"reasoning_tokens":2400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:00:30.471516+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the WSI experiment with the tissue SVM trained on a held-out set of manually annotated neuropil and stroma patches, report the SVM's own accuracy, and freeze α, β, and γ before seeing the test WSIs; if the SVM is inaccurate or the frozen weights do not reproduce 100%, the central WSI claim fails.","supporting_citations":[{"cited_title":"KAN: Kolmogorov–arnold net- works,","cited_arxiv_id":null,"evidence_quote":"Supplies the Kolmogorov-Arnold Network layer that replaces MLPs in the Swin blocks and classification head."},{"cited_title":"To- wards a general-purpose foundation model for computational pathology,","cited_arxiv_id":null,"evidence_quote":"Provides the UNI foundation model baseline that the paper claims to outperform."},{"cited_title":"A visual-language foundation model for computational pathology,","cited_arxiv_id":null,"evidence_quote":"Provides the CONCH vision-language foundation model baseline that the paper claims to outperform."},{"cited_title":"Transformer-based unsupervised con- trastive learning for histopathological image classification,","cited_arxiv_id":null,"evidence_quote":"Provides the CTransPath baseline, a CNN-transformer hybrid pretrained on pathology, that the paper claims to outperform."},{"cited_title":"Swin transformer: Hierarchical vision transformer using shifted windows,","cited_arxiv_id":null,"evidence_quote":"Supplies the Swin Transformer backbone that the paper modifies into the Swin KANsformer."},{"cited_title":"Fmdnn: A fuzzy-guided multigranular deep neural network for histopathological image classification,","cited_arxiv_id":null,"evidence_quote":"Provides the FMDNN pathology classification baseline that the paper compares against on both datasets."}],"review_version":1}