{"id":"53d9a059-0eed-43d7-b030-35e6b5107776","arxiv_id":"2505.16304","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A combined SAM2, Mamba, and UNet architecture reports state-of-the-art Dice of 0.9103 on the ACDC cardiac MRI benchmark, with no code or error bars yet released.","lead":"This paper combines the SAM2 vision model, the Mamba state-space model, and a UNet decoder into one cardiac MRI segmentation network, reporting slightly better scores on a public benchmark. A generalist should care because automatic heart segmentation could reduce manual labeling time and support surgical robotics, though the claimed robotic benefits are not tested here.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 0.71pp mDice advantage may be an artifact of uncontrolled baseline comparison; Table I mixes retrained and literature-reported results without specifying splits or preprocessing.","rationale":"The reader's weakest_assumption identifies the uncontrolled Table I comparison as the key threat, and I agree this is the single most load-bearing concern. The paper's central quantitative claim is the 0.9103 mDice and the 0.71pp margin over LeViT-UNet-384. That claim stands or falls on whether the baselines were evaluated under identical conditions. The paper's own description admits a mixture of retrained results and literature-reported numbers, but it does not say which methods are which or specify the preprocessing, split, and metric implementation. Because ACDC has no canonical 2D split and reported scores vary by several points with preprocessing decisions, the margin could easily be an artifact. Without error bars or multiple seeds, one cannot tell whether the advantage is real. The impossible software stack (PyTorch 1.10 with Python 3.12 and CUDA 12.1) and the undefined 'GCAA' ablation entry further undermine confidence in the experimental reporting, but they are secondary to the comparison-protocol issue. If a controlled re-evaluation shows the margin persists, the conditional acceptance could be upgraded; if it disappears, the headline claim collapses. Thus the reader's CONDITIONAL verdict is appropriate, and my stress-test does not change it; it sharpens the specific test that would resolve the concern.","tokens_in":10870,"tokens_out":6367,"duration_ms":52821,"concrete_test":"Run LeViT-UNet-384, Mamba-UNet, and SAMba-UNet on the same ACDC preprocessed slices, same patient split, same augmentation, and the same evaluation script, with at least 3 random seeds. If the mean SAMba-UNet mDice minus the best baseline mean is below about 0.3pp or within 1 standard deviation, the claimed 0.71pp improvement is not significant; report per-seed numbers and the exact split and preprocessing used.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-D states that models with official implementations were 'retraining and testing on an A800 server (or using official weights directly)', while others use 'the officially reported results from their respective papers'. The paper never lists which baselines fall in each category, nor does it state the ACDC split, 2D preprocessing (cropping, resampling, slice selection), or the exact metric code used for the retrained models. ACDC has no standardized 2D train/val split, and published Dice on it varies by several points with preprocessing choices. With no error bars, confidence intervals, or multiple-seed results, the reported 0.71pp mDice advantage over LeViT-UNet-384 is within the range of protocol and seed variation. The claim that SAMba-UNet 'surpasses all comparative models' is therefore not established as an architectural gain; it could be an artifact of an uncontrolled comparison. This is the load-bearing assumption because the paper's headline result is the 0.9103 mDice against these baselines. A further reproducibility red flag: the stated stack (PyTorch 1.10, Python 3.12, CUDA 12.1) is not a valid released combination, and Table II ablates 'GCAA' although no such module is defined in Section III, weakening confidence that the experimental description maps to the architecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAMba-UNet, a U-shaped dual-encoder architecture that combines a frozen SAM2 Hiera encoder, a VMamba encoder, and a VMamba decoder, with two new modules: a Dynamic Feature Fusion Refiner (DFFR) and a Heterogeneous Omni-Attention Convergence Module (HOACM). The method is evaluated on the ACDC cardiac MRI segmentation benchmark, reporting an mDice of 0.9103 and HD95 of 1.0859 mm, and the authors claim it surpasses all compared baselines. The paper also states that the model is directly applicable as a perception module for intelligent medical and surgical robotic systems. The experiments include a comparison table, two ablation tables, and qualitative visualizations.","tokens_in":11094,"tokens_out":3710,"duration_ms":29311,"significance":"If the performance claim were established under a controlled protocol, the architecture would be a useful contribution: integrating a frozen SAM2 encoder with a linear-complexity Mamba branch for medical segmentation is a plausible and timely direction, and the use of a public benchmark (ACDC) is a strength. The paper also promises open-sourced code, which would support reproducibility. However, the central empirical claim is currently not established because the baseline comparison in Table I is not controlled, the boundary metrics that are highlighted in the abstract are not compared against any baseline, and several modules referenced in the ablations are undefined. The claimed robotic-perception applicability is not tested by the experiments. The contribution is therefore more of a proposal with promising but unverified results than a validated method at this stage.","major_comments":[{"comment":"The main performance claim rests on an uncontrolled baseline comparison. The text says that models with official implementations were 'retraining and testing on an A800 server (or using official weights directly)', while others use 'the officially reported results from their respective papers', but it never lists which baseline falls into which category, what ACDC train/validation split was used, what 2D preprocessing was applied (cropping, resampling, slice selection), or what metric implementation was used. ACDC has no standardized 2D split, and published Dice values on it vary by several points with preprocessing choices. No standard deviations, confidence intervals, or multiple-seed results are reported for any method. Consequently, the 0.71 percentage-point mDice advantage over LeViT-UNet-384 (Section IV-D) is within the range of protocol and seed variation and cannot be interpreted as an architectural gain. The authors need to retrain all baselines under the same split, preprocessing, and metric code, report per-class Dice for all, and provide statistical significance or at least multiple-seed variance.","section":"IV-D, Table I"},{"comment":"The abstract and conclusion emphasize boundary localization and report HD95 as a headline result, but Table I lists only mDice, RV, MYO, and LV for the baselines. HD95 and ASD, which are defined in Section IV-C, are reported for the proposed method only through the ablation tables. This makes it impossible to support the claim of 'notably improving boundary localization' over prior work. Baseline HD95/ASD values must be included for a fair boundary-performance comparison.","section":"Table I and Abstract"},{"comment":"The ablations contain an undefined module: Table II includes a row 'w/o GCAA', but no module named GCAA is defined in Section III; the defined modules are OCA and BSEA, and a 'gated spatial attention (GSA)' appears inside OCA. As written, the reader cannot determine what is being ablated. In addition, the ablation differences are small (e.g., ALL 0.9103 vs. w/o BSEA 0.9064) and no variance or significance measures are reported, so it is unclear whether these differences are meaningful. The authors should define every ablated component, align the naming with the method section, and report multiple-seed results.","section":"Table II, Table III, Section III-C"},{"comment":"The implementation details are not reproducible as stated. The paper reports Python 3.12.0 with PyTorch 1.10 and CUDA 12.1, but PyTorch 1.10 does not support Python 3.12, and the combination as written is not a valid released configuration. Additionally, the paper says 'the preprocessed ACDC dataset' was used but does not specify the preprocessing steps, image resolution, slice selection, or data split. This blocks reproduction. The authors must provide the exact software versions, preprocessing pipeline, split definitions, and either the released code or a precise configuration file.","section":"IV-B"},{"comment":"The paper repeatedly claims that SAMba-UNet is 'directly applicable as a perception module within intelligent medical and surgical robotic systems' and includes a robotic workflow figure, but no experiment evaluates robotic perception, inference latency, real-time constraints, or integration with a robot controller. The ACDC segmentation benchmark alone does not test this claim. Either the robotic claim should be removed from the title, abstract, and conclusion, or supporting experiments (e.g., latency, deployment on a robotic platform) should be added.","section":"Abstract, Fig. 1, Conclusion"}],"minor_comments":[{"comment":"The symbol M is reused for both the attention weight matrix in Eq. (10) and the spatial attention output in Eq. (12), which is confusing; please use distinct notation for these quantities.","section":"Eq. (12)"},{"comment":"In Eq. (4), X_cat has shape B×2C while X_sp is not given an explicit channel dimension in Eq. (3); the addition X_cat + X_sp appears dimensionally ambiguous and should be clarified.","section":"Eq. (4)"},{"comment":"The text 'we integrate design principles from FE-UNet and Medical SAM Adapter' references FE-UNet without a citation; a reference or a brief description is needed.","section":"III-A"},{"comment":"Reference [37] is cited to support the claim about SAM2's window-based absolute position encoding, but the cited paper is about mitigating position bias in large language models and does not appear relevant; please cite the appropriate source.","section":"References"},{"comment":"The caption contains a typo: 'eterogeneous' should be 'Heterogeneous'.","section":"Figure 4 caption"},{"comment":"The claim of being 'the first pioneering framework that successfully synergizes SAM2, Mamba, and UNet' should be tempered or supported by a more thorough recent-literature comparison, since several SAM+Mamba works already exist (e.g., SAM-Mamba [33]) and the novelty claim as stated is difficult to verify.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the uncontrolled baseline comparison in Table I; this is not a matter of style but of the central claim's validity. The robotic-perception framing may also be a scope concern for the journal, since the experiments do not address it. If the authors can provide a fully matched comparison with variance estimates and fix the reproducibility issues, the paper could become a solid empirical contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the architecture is new and clearly described, but the empirical claim that it beats LeViT-UNet-384 by 0.71 pp on ACDC is not established as a real gain, because the baseline comparison is uncontrolled and the reported setup doesn't add up.\n\nWhat's actually new: pairing a frozen SAM2 Hiera encoder with a VMamba encoder and a single VMamba decoder, with the DFFR and HOACM fusion blocks, is a genuine combination not found in the works cited (SAM-Mamba uses the original SAM, Mamba-UNet has no SAM2). The module descriptions in Section III are explicit—equations for the pooling, channel gating, attention, etc.—so an experienced reader can reconstruct the architecture. The ablations in Tables II and III are the right shape: every component is removed and mDice, HD95, etc. are reported. That's more than many architecture papers do.\n\nThe soft spots are in the evaluation, and they're load-bearing. Table I mixes models that were retrained on an A800 with numbers taken from other papers, but the paper never says which baselines fall in which category, nor what ACDC split/preprocessing was used. ACDC has no standard 2D split, and published Dice can swing by a few points with preprocessing choices. With no error bars, no multiple seeds, and no significance tests, the reported 0.71 pp gap is within the noise of protocol differences. So 'surpasses all comparative models' is not demonstrated as an architectural gain. HD95 is a headline metric but is not in Table I at all. The robotic perception claim in the title and abstract is not tested—no robot, no latency, no integration experiment. There are also two concrete red flags in the text: the stated stack (PyTorch 1.10 + Python 3.12 + CUDA 12.1) is not a valid released combination, and Table II ablates 'GCAA' although no such module is defined anywhere. These don't necessarily kill the architecture, but they make the experimental description hard to trust as written.\n\nWho this is for: someone working on SAM/Mamba hybrids for medical segmentation, who wants to see the design space explored. It's a plausible engineering contribution that needs a proper head-to-head comparison—same split, same preprocessing, same metric code, multiple runs—before the performance claim can be believed.\n\nRecommendation: send it to peer review, but the referee should demand the split, per-baseline protocol, error bars, and a fix for the software-stack and GCAA inconsistencies. If those are resolved, the paper could be a solid niche contribution; as it stands the headline number is unsupported.","headline":"New architecture, uncontrolled comparison: the 0.71pp ACDC gap over LeViT-UNet-384 is not established as a real gain.","tokens_in":11699,"tokens_out":2514,"would_cite":false,"duration_ms":18360,"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":"Fusing SAM2 with Mamba in a UNet skeleton yields the best reported ACDC cardiac MRI segmentation, with a 0.9103 mDice.","keywords":["SAM2","Mamba","UNet","cardiac MRI segmentation","ACDC","Dynamic Feature Fusion Refiner","HOACM","medical robotic perception"],"falsifier":"Re-run all ten baselines plus SAMba-UNet on a single, pre-registered ACDC protocol with identical preprocessing and seeds, and check whether the 0.71-point mDice lead over LeViT-UNet-384 survives; if it does not, the headline gain is protocol artifact.","tokens_in":10609,"feed_emoji":"🫀","tokens_out":5141,"duration_ms":32924,"temperature":0.7,"pith_summary":"This paper tries to establish that a single segmentation architecture can profitably combine three apparently different ideas: the frozen vision foundation model SAM2, the linear-complexity state-space model Mamba, and the classic UNet encoder–decoder skeleton. On the ACDC cardiac MRI benchmark, the resulting SAMba-UNet reports a mean Dice of 0.9103 and a 95th-percentile Hausdorff distance of 1.0859 mm, beating ten published competitors and placing its biggest gains on the right ventricle and myocardium. If the architecture really delivers that margin, it would give medical robotic perception a ready-made high-fidelity segmentation module for pre-operative planning, intraoperative navigation, and post-operative screening.","feed_headline":"SAMba-UNet tops cardiac MRI rivals by 0.71 points","feed_subtitle":"It fuses SAM2's boundary detail with Mamba's long-range context to sharpen right-ventricle and myocardium segmentation.","key_machinery":"The load-bearing mechanism is the Heterogeneous Omni-Attention Convergence Module (HOACM), a fusion block that combines Omniscient Contextual Attention (OCA), which re-establishes global pixel-level positional semantics for SAM2's hierarchical features, with Bifurcated Selective Emphasis Attention (BSEA), which dynamically weights Mamba's global and locally salient features before cross-attention merges the two streams. The Dynamic Feature Fusion Refiner, a multi-scale pooling and channel–spatial calibration adapter, carries the domain-transfer work of pulling SAM2's natural-image representations toward MRI.","core_discovery":"The paper's central claim is that SAM2's window-based positional detail and Mamba's long-range global context are complementary, and that a dedicated fusion module—not just a concatenation—lets a UNet exploit both at once. To that end it introduces a Dynamic Feature Fusion Refiner that adapts SAM2's natural-image features to MRI via multi-scale pooling and channel–spatial calibration, and a Heterogeneous Omni-Attention Convergence Module (HOACM) whose Omniscient Contextual Attention and Bifurcated Selective Emphasis Attention jointly reweight the two encoder streams. The reported results on ACDC—mDice 0.9103, mHD95 1.0859 mm—and the ablation tables are offered as evidence that every component contributes, with the right ventricle and myocardium showing the clearest gains.","pith_inferences":["The paper argues the segmentation maps are 'directly applicable' to robotic perception, but it reports no robot-in-the-loop or latency experiment; treating that as a transfer claim rather than a demonstrated result would be my reading.","The headline comparison mixes re-trained baselines with numbers copied from original papers; if protocols differ, the 0.71-point gap could shrink or vanish, so an independent re-benchmark under one protocol would be the natural stress test.","A testable next step is to freeze the trained encoder and probe whether the fusion module transfers to other cardiac MRI views or other modalities (e.g., CT), since the domain-shift argument is generic.","The 'first to synergize SAM2, Mamba, and UNet' claim is about architectural novelty; whether that combination is necessary—versus a well-tuned single encoder—is only answered by the ablations, which show each module adds a small increment."],"forward_implications":["If the reported margin is real, SAMba-UNet is the strongest known single model on the ACDC benchmark, with a 0.71-percentage-point mDice lead over the next-best entry.","The gains concentrate on the right ventricle and myocardium—the structures with the most complex boundaries—so the approach is aimed at exactly the cases where manual annotation is hardest.","Because the Mamba branch processes sequences in linear time, the added global-context capacity does not force quadratic attention cost, keeping the fused model practical for clinical image volumes.","The frozen SAM2 encoder with a lightweight adapter means the method inherits foundation-model priors without full fine-tuning, a recipe that could transfer to other MRI targets with modest data.","The authors position the model as a perception module for surgical robotic systems, where its boundary precision would support navigation and complication screening."],"supporting_citations":[{"why":"SAM2, the frozen vision foundation model whose Hiera encoder provides the position-sensitive branch.","marker":"[19]"},{"why":"VMamba, the visual state-space model that supplies the long-range global context branch.","marker":"[24]"},{"why":"Mamba, the selective state-space model whose linear-complexity sequence modeling motivates the Mamba encoder.","marker":"[30]"},{"why":"Mamba-UNet, the prior UNet-like Mamba segmentation model whose VSS block and decoder design SAMba-UNet reuses.","marker":"[31]"},{"why":"UNet, the canonical encoder-decoder skeleton and the segmentation architecture being extended.","marker":"[2]"},{"why":"LeViT-UNet-384, the strongest baseline whose 0.9032 mDice is the margin SAMba-UNet claims to beat by 0.71 points.","marker":"[41]"},{"why":"Segment Anything Model, the earlier foundation model whose adapter-based medical fine-tuning motivates the refiner.","marker":"[15]"}],"fun_headline_variants":["SAM2+Mamba+UNet: cardiac MRI segmentation sharpens RV and myocardium","SAMba-UNet hits 0.91 Dice in cardiac MRI with hybrid fusion","Dual-encoder fusion lifts cardiac MRI boundary precision","SAM2 and Mamba combine for cardiac MRI robot perception"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported 0.71-point margin assumes every baseline result in Table I comes from a comparable protocol—same ACDC split, preprocessing, and metric implementation—so the gap reflects architecture, not measurement differences.","fun_headline_variants_meta":{"raw":{"variants":["SAM2+Mamba+UNet: cardiac MRI segmentation sharpens RV and myocardium","SAMba-UNet hits 0.91 Dice in cardiac MRI with hybrid fusion","Dual-encoder fusion lifts cardiac MRI boundary precision","SAM2 and Mamba combine for cardiac MRI robot perception"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000855,"raw_usage":{"total_tokens":3718,"prompt_tokens":952,"completion_tokens":2766,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":2700}},"tokens_in":568,"tokens_out":2766,"duration_ms":14893,"temperature":1.0,"reasoning_tokens":2700,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:03:11.061961+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run all ten baselines plus SAMba-UNet on a single, pre-registered ACDC protocol with identical preprocessing and seeds, and check whether the 0.71-point mDice lead over LeViT-UNet-384 survives; if it does not, the headline gain is protocol artifact.","supporting_citations":[{"cited_title":"Levit-unet: Make faster encoders with transformer for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"LeViT-UNet-384, the strongest baseline whose 0.9032 mDice is the margin SAMba-UNet claims to beat by 0.71 points."}],"review_version":1}