{"id":"ed251b67-d788-49e6-bf8c-68c9e8546a9b","arxiv_id":"2504.17515","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Mamba-Sea applies global and local sequence-level augmentation to a Mamba U-Net and reports state-of-the-art domain-generalized segmentation on fundus, prostate, and skin lesion benchmarks.","lead":"A new image segmentation framework, Mamba-Sea, combines Mamba state-space layers with global appearance changes and local style perturbations to handle hospital-to-hospital image differences. It reports top average Dice scores on three medical segmentation benchmarks, including the first average above 90% on the Prostate dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 'first/SOTA' claim needs a comparison to the cited same-group START [36], which is never evaluated in Tables II, III, or IX; without it the headline claim is not established.","rationale":"The reader's verdict is CONDITIONAL, with the stated weakest assumption being the Gaussian model in LSA (Eqs. 8-12). I agree that the Gaussian assumption is unvalidated, but it is not the most load-bearing threat to the central claim: the empirical SOTA results are supported by ablations, parameter sensitivity checks, and paired t-tests against the compared baselines. Even if the Gaussian assumption fails, LSA could still be a useful generic regularizer, so the paper's practical contribution might survive. The more direct threat is the omission of START [36], which the paper cites as the basis for its LSA motivation and domain-gap evaluation. START is a same-group SSM-DG method whose title and cited context indicate token-aware transformation to suppress domain-specific information in state-space models. If START already produced DG results on medical segmentation benchmarks, the paper's 'first Mamba DG medical segmentation' and 'SOTA' statements are not just unproven but potentially false. This is a concrete, checkable concern, not a stylistic one. The reader's rationale did mention the START conflict, so there is partial agreement, but the reader's formal weakest-assumption field focused on the Gaussian assumption instead. I would keep the verdict CONDITIONAL rather than moving it, because the paper's internal evidence (Tables II-IV, VI-IX) supports substantial gains over the listed baselines; the missing comparison should be a required revision rather than a ground for rejection. The proposed concrete test would settle whether the headline novelty claim survives.","tokens_in":23738,"tokens_out":9360,"duration_ms":91619,"concrete_test":"Obtain START [36] (code or published results) and run it under exactly the protocol of Tables II, III, and IX: leave-one-domain-out, same VM-UNet/VMamba-S initialization, same preprocessing, and same Dice/ASD metrics on Prostate, Fundus, and the ESP-MedSAM skin-lesion split. If START's average Dice on those benchmarks is at least 90.02%, 89.68%, and 93.11%, or if START already reports these benchmarks, the 'first' and 'SOTA' claims in the abstract and Section I are invalidated. A minimal version of this check is to add START as a row to Tables II, III, and IX and to the paired t-tests in Table VIII.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central assertion is that Mamba-Sea is the first Mamba-based framework for domain-generalized medical image segmentation and that it surpasses prior SOTA (90.02% average Dice on Prostate, 89.68% on Fundus, 93.11% on skin lesion). This claim is only secure if all relevant prior work is included in the comparison. The paper itself cites START [36], a NeurIPS 2024 state-space-model DG method from the same group, and explicitly uses it to motivate the LSA design (Sec. III.C.3: input-dependent matrices accumulate domain-specific information) and to justify the domain-gap evaluation protocol (Sec. IV.C.3: 'we conduct additional quantitative experiments following the previous work [36]'). START is never compared against in Tables II, III, or IX. If START already applies Mamba/SSM to medical DG and reports Dice on Prostate/Fundus/skin lesion, then Mamba-Sea is not the first and its SOTA numbers are not established against the most relevant prior work. This is not a generic novelty quibble: the paper's own text indicates that the design and evaluation protocol are direct extensions of START, so the omission could conceal that the reported gains are incremental. The LSA Gaussian assumption flagged by the reader is a weaker concern because even if the Gaussian model is incorrect, LSA could still act as an effective regularizer; the missing START comparison directly threatens the headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Mamba-Sea, a domain-generalization framework for medical image segmentation built on a Mamba/VM-UNet backbone. It combines a global appearance variation augmentation (GVA) with a local sequence-wise style transformation (LSA) that resamples channel-wise feature statistics under a Gaussian assumption over a random continuous token mask, plus a semantic consistency loss between original and globally augmented views. The method is evaluated on Fundus, Prostate, and skin-lesion benchmarks under leave-one-domain-out protocols, reporting average Dice of 89.68%, 90.02%, and 93.11%, respectively. The paper claims to be the first Mamba-based DG framework for medical image segmentation and the first to exceed 90% average Dice on the Prostate dataset, with code released.","tokens_in":24048,"tokens_out":7578,"duration_ms":73898,"significance":"If the empirical results hold, Mamba-Sea would be a useful extension of Mamba to domain generalization, with a small computational overhead (GVA adds 2.58e-4 M parameters and training time increases by about 0.004 s per iteration), public code, and reported paired t-tests on all three datasets. The ablations and comparisons against CNN, ViT, and SAM-based baselines are extensive. However, the headline novelty and SOTA claims are not yet established because the cited prior state-space-model DG method START [36] is not included in any comparison table, and the hyperparameter selection protocol is unspecified. The Gaussian resampling assumption is a modeling choice; the paper's empirical value would remain if LSA were positioned as a stochastic style perturbation regularizer rather than a faithful model of unseen-domain statistics.","major_comments":[{"comment":"The manuscript claims to be the first Mamba-based DG framework for medical image segmentation and to achieve SOTA, yet it never compares against START [36], which is cited as the source of the LSA motivation (Sec. III.C.3) and of the domain-gap evaluation protocol (Sec. IV.C.3). Because [36] is titled 'A generalized state space model with saliency-driven token-aware transformation' and is prior work from the same group, it appears to be a directly relevant SSM-based DG method. Without a quantitative comparison on Fundus, Prostate, or skin lesion, the 'first' and SOTA claims are not established. Please add START to all comparison tables or explain why it is not a comparable prior method, and revise the novelty statements accordingly.","section":"Sec. IV.C; Tables II, III, IX"},{"comment":"Hyperparameters are reported as fixed after sensitivity analysis: tau differs between Fundus (0.4) and Prostate (0.03), p is set to 0.75, and lambda to 0.1. The paper does not state whether these values were chosen by validation on held-out source domains or by observing target-domain test performance. In a DG evaluation, tuning on the target test set can inflate results and violates the held-out protocol. Furthermore, tau is not included in the sensitivity analysis of Table VII, so its effect is not quantified. Please specify the exact validation procedure used to select tau, p, and lambda, and if target-domain information was used, rerun the comparisons with source-only hyperparameter selection.","section":"Sec. IV.B and IV.C.4"},{"comment":"All Fundus and Prostate results are reported as means over three runs with no standard deviations, so the 1.41% Prostate margin over DCAC cannot be assessed against run-to-run variability. The paired t-tests in Table VIII are computed at the sample level; for Prostate, 2D slices from the same 3D volume are not independent, which can make sample-level p-values artificially small. Please report mean plus/minus standard deviation over runs, provide run-level comparisons, and use volume-level or clustered significance tests for Prostate.","section":"Sec. IV.C, Tables II-III, VIII"},{"comment":"The claim 'first to surpass a Dice coefficient of 90% on the Prostate dataset' is ambiguous: the 90.02% is the average over six leave-one-out tasks, while the baseline VM-Unet already exceeds 90% on Domain 1 (91.02) and Domain 6 (90.91), and DCAC reaches 91.76 on Domain 1. Please state explicitly that the record is the average across domains, and compare against the prior best average of 88.61% as given in the text.","section":"Abstract and Table III"}],"minor_comments":[{"comment":"The notation f_g(x) is used both for the input features to GVA and for the transformed output in Eq. (7), which is confusing; please use separate symbols for the input and the augmented output.","section":"Sec. III.C.2, Eq. (5)-(7)"},{"comment":"Several reference numbers in Table IX are incorrect: WT-PSE is [62] not [37], SAM-Med2D is [63] not [62], and Med-SA is [7] not [63]; please correct the citations.","section":"Table IX"},{"comment":"The t-SNE and MMD analyses measure feature alignment, which may partly reflect the effect of the consistency loss rather than genuine domain-invariant semantic learning; consider an additional analysis that separates alignment from semantic content preservation.","section":"Sec. IV.C.3"},{"comment":"The architecture of the GVA lightweight network is not described in enough detail (number of enhancement blocks, channel widths, activation functions), and the skin lesion experiment does not state whether the same training configuration and hyperparameters as Fundus/Prostate were used; please provide these reproducibility details.","section":"Sec. III.C.4 and IV.B"},{"comment":"The choice of continuous sub-sequence perturbation over randomly scattered token perturbation is asserted to preserve token dependencies, but no ablation compares the continuous mask in Eq. (12) with an i.i.d. random mask of the same proportion; please add such an ablation or soften the claim.","section":"Sec. III.C.3"},{"comment":"There are several typos and reference inconsistencies, including 'pervious' for 'previous', 'matrics' for 'matrices', and the duplicated LKM-UNet description in Sec. II.B; a careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the omission of START [36], which is from the same group and appears to be a direct predecessor of this work. Please ask the authors for a direct quantitative comparison with START on all three benchmarks, and for a clear statement of the hyperparameter selection protocol. The 'first' and 'SOTA' claims should be softened unless the comparison supports them."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: careful, incremental engineering paper that reports real gains on three medical DG benchmarks, but the headline \"first/SOTA\" claim is not established because the authors cite a same-group SSM-based DG method (START, NeurIPS 2024) that they never compare against. The rest is mostly solid.\n\nWhat's new: a Mamba-based DG segmentation pipeline with two augmentation modules (GVA at image level, LSA at sequence level) plus a consistency loss. LSA—resampling channel-wise feature statistics on random continuous token sub-sequences—is a sensible adaptation of MixStyle/ALOFT-style augmentation to Mamba's sequential scanning. Ablations show each component matters, and gains over the VM-UNet baseline are consistent across Fundus, Prostate, and skin lesion. t-SNE and domain-gap analyses are suggestive, though partly self-confirmatory. Code is released.\n\nSoft spots. The biggest one is the missing START comparison. The paper cites START [36] for the claim that Mamba accumulates domain-specific info, and Section IV.C.3 says the domain-gap protocol follows [36]. START is a state-space-model DG method from the same group, likely on medical segmentation. If START already reports Dice on these benchmarks, then \"first Mamba DG in medical segmentation\" is false and the SOTA numbers are not directly established against the most relevant prior work. The authors need to either compare against START on all three datasets or explain why it is excluded. That is not a minor novelty quibble; it bears on the central claim.\n\nSecondary issues: tables report means over three runs without std devs; hyperparameters tau, p, lambda are tuned per dataset with no statement that selection used only source-domain validation. The Prostate \"first above 90%\" is ambiguous because baselines already exceed 90% on several individual domains; the average is what crosses 90. The Gaussian resampling assumption in LSA is a modeling choice—it could still act as a regularizer even if the Gaussian is wrong, so I don't treat that as fatal.\n\nOverall: the empirical work is extensive and the incremental gains look real. The paper deserves a serious referee, but the referee should insist on the START comparison and a clearer hyperparameter selection protocol before acceptance.","headline":"Solid incremental engineering with real benchmark gains, but the missing comparison to the same-group START paper undermines the first/SOTA claim.","tokens_in":24571,"tokens_out":1841,"would_cite":false,"duration_ms":16781,"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":"Mamba-Sea, a Mamba-based framework with global-to-local sequence augmentation, achieves state-of-the-art domain-generalized medical image segmentation, including the first Prostate Dice score above 90%.","keywords":["domain generalization","medical image segmentation","state space models","Mamba","sequence augmentation","style transformation","semantic consistency training","cross-domain robustness"],"falsifier":"Run the released implementation on the Prostate leave-one-domain-out protocol with the Gaussian resampling in Eq. (10) replaced by uniform or bootstrap resampling of the same per-batch statistics; if average Dice does not drop below 90%, the Gaussian modeling is not what carries the improvement.","tokens_in":23545,"feed_emoji":"🩺","tokens_out":6436,"duration_ms":56666,"temperature":0.7,"pith_summary":"The paper is trying to establish that Mamba, a selective state space model, can be turned into a strong backbone for domain-generalized medical image segmentation rather than remaining limited to supervised same-distribution tasks. It proposes Mamba-Sea, which augments images globally and token sequences locally during training, and enforces semantic consistency between original and augmented views. On leave-one-domain-out benchmarks the framework is reported to reach 90.02% average Dice on the Prostate dataset, 89.68% on the Fundus dataset, and 93.11% on a large skin-lesion dataset, each above the previous state of the art. If the claim holds, clinically deployed segmentation models could be made more robust to new scanners and protocols without quadratic attention cost.","feed_headline":"First Mamba model cracks 90% Dice on prostate MRI segmentation","feed_subtitle":"Global-to-local sequence augmentation lifts unseen-domain segmentation on prostate, fundus, and skin-lesion benchmarks.","key_machinery":"The load-bearing machinery is the local sequence-wise style transformation (LSA) placed inside the Mamba scanning blocks, coordinated with the global appearance variation augmentation (GVA) and a semantic consistency loss. LSA computes per-instance channel-wise means and standard deviations, estimates their uncertainty across the batch, resamples them from a Gaussian model, and mixes the transformed features with the original features using a random continuous mask over tokens (Eqs. 8-13). This is designed to perturb exactly the regions of the input-dependent state matrices where, according to the paper's hypothesis, domain-specific information accumulates, while the continuous-mask choice preserves local semantic structure. GVA supplies whole-image appearance diversity through a lightweight gating network, and the consistency loss makes the two augmentation levels work together.","core_discovery":"The paper's central claim is that the input-dependent matrices of Mamba's selective scanning mechanism accumulate domain-specific information during training, and that this accumulated bias is the main obstacle to cross-domain generalization. Mamba-Sea therefore inserts a local sequence-wise style transformation (LSA) into the scanning blocks, which models per-batch channel statistics as Gaussian, resamples them to synthesize new style shifts, and applies them only to random continuous subsequences of tokens via a mask. A global appearance variation augmentation (GVA) uses a learnable gating network to simulate site-to-site illumination differences over entire images, and a semantic consistency loss forces predictions on original and augmented inputs to match. The paper reports that this combined recipe yields state-of-the-art averages of 90.02% Dice on Prostate, 89.68% on Fundus, and 93.11% on skin lesion segmentation, with GVA and LSA discarded at inference time so that test-time cost is unchanged.","pith_inferences":["Editorial inference: the Gaussian assumption on feature statistics is likely not the only source of the gains; comparing it against bootstrap or uniform resampling of the same statistics would isolate how much the distributional model matters.","Editorial inference: the GVA brightness threshold is hand-set per dataset (0.4 for Fundus, 0.03 for Prostate), so a learned or data-dependent threshold is a natural extension the paper leaves open.","Editorial inference: the reported 32% reduction in average feature-space domain gap in the decoder is a quantitative handle that could be applied to other SSM backbones to test whether the augmentation genuinely removes domain-specific accumulation.","Editorial inference: if the input-dependent matrices are indeed the carrier of domain bias, architectural changes to the state update itself, rather than augmentation only, should further improve generalization."],"forward_implications":["Mamba-based models can serve as competitive backbones for domain-generalized medical segmentation, matching or surpassing CNN- and ViT-based DG methods without quadratic self-attention cost.","The per-batch Gaussian resampling of feature statistics and continuous token masking give a concrete, reusable recipe for style augmentation inside state space models.","Because GVA and LSA are training-only and discarded at inference, the framework adds negligible test-time latency and parameter overhead, easing clinical deployment.","Removing either augmentation module or the consistency loss measurably lowers average Dice on both benchmarks, so the gains depend on the combination rather than on the Mamba backbone alone."],"supporting_citations":[{"why":"The Mamba-based U-Net that serves as the baseline backbone and the competitive performance that Mamba-Sea improves upon.","marker":"[8]"},{"why":"Supplies the VSS block and SS2D bidirectional scanning mechanism whose input-dependent matrices are the target of the local augmentation.","marker":"[34]"},{"why":"Introduces the selective state space model that motivates the architectural choice and the linear complexity assumption.","marker":"[18]"},{"why":"Defines the selective SSM module used inside the LSA augmentation pipeline.","marker":"[51]"},{"why":"Provides the Fundus benchmark protocol and the DoFE baseline, one of the main prior DG methods compared.","marker":"[32]"},{"why":"Provides the style-augmentation and dual-normalization baseline (RAM-DSIR) that LSA is compared against on the same backbone.","marker":"[12]"},{"why":"The previous state-of-the-art on Prostate (DCAC) whose 88.61% Dice is the threshold Mamba-Sea claims to exceed.","marker":"[37]"},{"why":"The MixStyle baseline that represents channel-level style augmentation, contrasted with LSA's sequence-wise continuous-mask approach.","marker":"[52]"},{"why":"The ALOFT baseline representing ViT-based DG augmentation that LSA is benchmarked against on the VM-UNet backbone.","marker":"[53]"},{"why":"Supplies the saliency-driven token-aware transformation and domain-gap measurement used to argue that input-dependent matrices accumulate domain-specific information.","marker":"[36]"}],"fun_headline_variants":["Mamba-Sea: first Mamba to crack 90% Dice on prostate","Global-to-local augmentation lifts Mamba across medical domains","Mamba model hits 90% Dice on unseen prostate MRI","Sequence-wise style augmentation boosts Mamba generalization","First Mamba framework for domain-general medical segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that per-batch feature statistics (channel-wise means and standard deviations) are distributed close enough to Gaussian that resampling them, masked over random continuous subsequences, produces realistic unseen-domain style shifts without damaging semantic content.","fun_headline_variants_meta":{"raw":{"variants":["Mamba-Sea: first Mamba to crack 90% Dice on prostate","Global-to-local augmentation lifts Mamba across medical domains","Mamba model hits 90% Dice on unseen prostate MRI","Sequence-wise style augmentation boosts Mamba generalization","First Mamba framework for domain-general medical segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000284,"raw_usage":{"total_tokens":1729,"prompt_tokens":1054,"completion_tokens":675,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":594}},"tokens_in":670,"tokens_out":675,"duration_ms":6648,"temperature":1.0,"reasoning_tokens":594,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:37:56.953056+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released implementation on the Prostate leave-one-domain-out protocol with the Gaussian resampling in Eq. (10) replaced by uniform or bootstrap resampling of the same per-batch statistics; if average Dice does not drop below 90%, the Gaussian modeling is not what carries the improvement.","supporting_citations":[{"cited_title":"VMamba: Visual state space model,","cited_arxiv_id":null,"evidence_quote":"Supplies the VSS block and SS2D bidirectional scanning mechanism whose input-dependent matrices are the target of the local augmentation."},{"cited_title":"Dofe: Domain-oriented feature embedding for generalizable fundus image segmentation on unseen datasets,","cited_arxiv_id":null,"evidence_quote":"Provides the Fundus benchmark protocol and the DoFE baseline, one of the main prior DG methods compared."},{"cited_title":"Generalizable cross- modality medical image segmentation via style augmentation and dual normalization,","cited_arxiv_id":null,"evidence_quote":"Provides the style-augmentation and dual-normalization baseline (RAM-DSIR) that LSA is compared against on the same backbone."},{"cited_title":"Domain and content adaptive convolution based multi-source domain generalization for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"The previous state-of-the-art on Prostate (DCAC) whose 88.61% Dice is the threshold Mamba-Sea claims to exceed."},{"cited_title":"Aloft: A lightweight mlp- like architecture with dynamic low-frequency transform for domain generalization,","cited_arxiv_id":null,"evidence_quote":"The ALOFT baseline representing ViT-based DG augmentation that LSA is benchmarked against on the VM-UNet backbone."},{"cited_title":"Start: A generalized state space model with saliency-driven token-aware transformation,","cited_arxiv_id":null,"evidence_quote":"Supplies the saliency-driven token-aware transformation and domain-gap measurement used to argue that input-dependent matrices accumulate domain-specific information."}],"review_version":1}