{"id":"3d8a5b51-4f27-499c-9c23-88e3029c7111","arxiv_id":"2412.08482","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SAM-Mamba couples a Mamba-based prior with a frozen SAM encoder and adapter fine-tuning to achieve state-of-the-art polyp segmentation and cross-dataset zero-shot generalization.","lead":"This paper introduces SAM-Mamba, a deep learning model that combines the Segment Anything Model with Mamba state-space layers to segment polyps in colonoscopy images more accurately. It reports gains over prior models on seen and unseen datasets, which could help clinicians detect colorectal cancer earlier.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on borrowed baseline metrics that are never re-run under SAM-Mamba's protocol; without a like-for-like comparison, the reported margins are not yet established.","rationale":"I read the paper as claiming a new state of the art for generalized zero-shot polyp segmentation based on Table 1-3 numbers. The most load-bearing condition for that claim is that the comparison is apples-to-apples with prior methods. The paper does not re-run any baseline and even notes that one baseline's numbers were inherited from another paper, so protocol drift remains a live risk. This is exactly the concern the reader identified, and it is not resolved by the ablation study, which only compares internal configurations. I also note the paper's own table contradicts a specific textual claim: in CVC-ClinicDB, SAM-Mamba's mIoU is 88.7 while MEGANet reports 89.4 and Polyp-PVT 88.9, yet the text says SAM-Mamba outperforms peers in mIoU; this makes the comparative claims harder to trust. I am not arguing the method is invalid: the architecture is plausible, the code is promised, and the ablation shows the Mamba-Prior component adds value relative to an adapter-only variant. But until baselines are re-run under the same protocol, the reported margins should be treated as provisional, which matches the reader's CONDITIONAL verdict, so no change is needed.","tokens_in":13867,"tokens_out":6546,"duration_ms":71463,"concrete_test":"Using the released code, reproduce SAM-Mamba's predictions on Kvasir-SEG and CVC-ClinicDB for training and on CVC-300, CVC-ColonDB, and ETIS for testing; then re-run CTNet, MEGANet, Polyp-PVT, and a Medical SAM Adapter variant under that same 900+550 image, 352x352 protocol with the same evaluation script. If SAM-Mamba's mDice/mIoU advantage over these re-run baselines falls below about 1 point or reverses, the headline SOTA claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise of the abstract's SOTA claim is that all 14 baselines in Tables 1-3 were evaluated under exactly the same protocol as SAM-Mamba. The paper states in Section 4.3 that PEFNet's results are sourced from M2UNet and that all other baselines are taken directly from their original works, so no baseline was re-run with SAM-Mamba's 352x352 input, multi-scale augmentation, or 200-epoch training schedule. Differences in training split, preprocessing, and evaluation scripts are known to shift polyp-segmentation scores by several mIoU points, which is the same order as the reported gains (+4% and +3.8%). The abstract further claims superiority over Adapter-based models, but no adapter-based baseline from the literature is included; the only adapter comparison is the internal ablation. Independent support from the ablation study and qualitative figures does not close this gap because those comparisons are internal and do not calibrate the externally quoted baseline numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAM-Mamba, an adapter-based fine-tuning of the Segment Anything Model (SAM) for polyp segmentation. A Mamba-Prior module, composed of multi-scale spatial decomposition (MSD), channel saliency/context pooling, and Mamba channel interaction, injects domain priors into the frozen SAM encoder. The method is trained on Kvasir-SEG and CVC-ClinicDB and evaluated in a zero-shot generalization setting on CVC-300, CVC-ColonDB, and ETIS. The authors report that SAM-Mamba outperforms CNN-, ViT-, and adapter-based models, and they provide quantitative comparisons against 14 baselines plus ablation studies.","tokens_in":14086,"tokens_out":4272,"duration_ms":41630,"significance":"If the claimed results are protocol-equivalent, SAM-Mamba would represent a strong new state of the art for generalized zero-shot polyp segmentation, with notable margins on unseen datasets. The architectural idea of combining SSM-based Mamba layers with SAM adapters is motivated, and the ablation study provides some evidence that the MSD and Mamba components each help. The release of code and models via GitHub is a strength that supports reproducibility. However, the significance is currently conditional, because the headline comparison relies on borrowed baseline numbers and contains an internal contradiction.","major_comments":[{"comment":"The text states that on CVC-ClinicDB SAM-Mamba \"outperforms peers in key metrics such as mIoU, Fβ, and Sα,\" but Table 1 shows MEGANet at 89.4 mIoU versus 88.7 for SAM-Mamba. The mIoU claim is therefore contradicted by the paper's own table. Please correct either the table or the text, and ensure that all reported \"outperforms\" statements are consistent with the numbers shown.","section":"§4.3, Table 1"},{"comment":"The abstract and Section 4.3 claim superiority over \"Adapter-based models,\" but no adapter-based baseline appears in Tables 1-3; the only adapter comparison is the internal ablation in Table 4. Without at least one external adapter-based baseline (e.g., Medical SAM Adapter or a comparable SAM fine-tuning method) evaluated under the same protocol, the claim as stated is not supported by the presented evidence.","section":"Abstract, §4.3, Tables 1-3"},{"comment":"All baseline results are taken from the original papers, and PEFNet's numbers are taken from M2UNet, while SAM-Mamba is trained with 352x352 input, multi-scale augmentation, and 200 epochs. Since no baseline is re-run under SAM-Mamba's protocol, protocol differences in preprocessing, training split, and evaluation could account for part of the reported gains (e.g., +4.0 and +3.8 mIoU on CVC-ColonDB and ETIS). The zero-shot superiority claim needs either a like-for-like comparison for the most competitive baselines or an explicit justification that the borrowed numbers are protocol-equivalent.","section":"§4.2, §4.3"},{"comment":"The claim of \"outperforming all SOTA models\" on unseen datasets is stated as a margin over the next best method, but on several metrics the advantage is small or reversed (e.g., SAM-Mamba's Fβ on CVC-300 is 88.8 versus CTNet's 89.4, and its E^max on CVC-ClinicDB is below several baselines). Please report the comparison more precisely, and avoid selective metric reporting when summarizing results.","section":"§4.3, Tables 2-3"}],"minor_comments":[{"comment":"The notation f(θ) is used to describe the pseudo-mask extraction, but f(θ) is never formally defined in the text; please specify what this function and its parameter space θ represent.","section":"§3.2.3, Figure 1"},{"comment":"In Equation (3), the second concatenation operand is written as \"MC_o ⊙ M*\", which appears to be a typo for \"M^C_o ⊙ M*\"; please ensure superscript notation is consistent throughout.","section":"Equation (3)"},{"comment":"In the CVC-300 block, the CFA-Net row lists Fβ=93.8 and Sα=87.5; since Fβ is normally bounded by structural similarity, please verify these entries or explain why their relative ordering is plausible.","section":"Table 2"},{"comment":"The dagger symbol after \"SAM-Mamba\" in the abstract and the footnote marker do not appear to be resolved anywhere in the text; please either add the promised footnote or remove the symbol.","section":"Abstract, Title"}],"recommendation":"major_revision","confidential_remarks":"The core architecture and ablation are plausible, and the paper could become publishable after the comparison protocol is made rigorous and the internal contradictions are fixed. The central issue is that the SOTA claim currently rests on borrowed baseline numbers and is contradicted by one of the paper's own tables. This is within the manuscript's scope to repair, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, SAM-Mamba is a genuinely new combination: a frozen SAM encoder with a Mamba-Prior module (multi-scale spatial decomposition, channel saliency/context pooling, Mamba gating) plus cross-attention adapters and a two-stage pseudo-mask prompting scheme. That is not in the prior literature, and the ablation shows the MSD and Mamba components each add a couple of points on several datasets. Second, the headline SOTA claim is not established by the reported comparisons, because every baseline number is borrowed from other papers and none are re-run under the paper's protocol. That is a load-bearing weakness, not a cosmetic one.\n\nWhat the paper does well: the architecture is clearly motivated and the design choices are sensible. The two-stage training—train the encoder side-output first, then the decoder with the pseudo-mask as prompt—is an interesting way to remove SAM's prompt dependency. The authors release code and models, which is real evidence and will let others check the numbers. The experiments cover five standard benchmarks with the PraNet split, which is the field's default, and the reported gains on unseen datasets are large enough to be worth checking.\n\nThe soft spots. The CVC-ClinicDB mIoU contradiction is real: the text says SAM-Mamba outperforms peers on mIoU, but Table 1 has MEGANet at 89.4 and SAM-Mamba at 88.7. That is a factual error that needs correcting. The abstract's claim about beating Adapter-based models has no external adapter baseline in the tables; the only adapter comparison is the internal ablation, which does not support the claim as written. The borrowing of baselines is the deeper issue: PEFNet's numbers are taken from M2UNet's paper, and the rest come from their original works. With different input resolutions, training schedules, and augmentation, polyp segmentation scores can shift by several mIoU points—same order as the reported +4% gains. The paper provides no error bars or statistical testing, so we cannot tell whether the margins are real or protocol artifacts.\n\nWho this is for: researchers working on SAM-based medical segmentation and polyp segmentation specifically. They will find the architecture ideas useful. But the paper as a fresh submission needs revision before the SOTA claims are acceptable.\n\nRecommendation: send to peer review. The architecture is novel and the code release makes it verifiable. A referee should ask for re-run baselines under the same protocol, a corrected Table 1, and an actual adapter-based comparison. The method may well be as good as claimed, but the current evidence does not prove it.","headline":"A novel SAM+Mamba adapter architecture worth a serious look, but the SOTA claim is built on borrowed baseline numbers that were never re-run under the same protocol.","tokens_in":14616,"tokens_out":1837,"would_cite":false,"duration_ms":18287,"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":"A Mamba-Prior module injected into a frozen SAM encoder improves polyp segmentation on seen and unseen colonoscopy datasets, beating the CNN and ViT baselines it is compared with.","keywords":["polyp segmentation","zero-shot generalization","Segment Anything Model","Mamba","state space models","adapter fine-tuning","colonoscopy","multi-scale feature decomposition"],"falsifier":"Re-running CTNet, MEGANet, Polyp-PVT, and the other compared methods under SAM-Mamba's exact protocol (1,450 training images, 352x352 inputs, same augmentation, same evaluation protocol) and finding that any baseline reaches or exceeds 78.2% mIoU on ETIS or 77.1% mIoU on CVC-ColonDB would show the claimed zero-shot margins are not reproducible.","tokens_in":13670,"feed_emoji":"🩺","tokens_out":10762,"duration_ms":96644,"temperature":0.7,"pith_summary":"Polyp segmentation in colonoscopy has to cope with polyps that vary in size, color, and boundary clarity, and a clinically useful model should also transfer to images from devices it has never seen. This paper proposes SAM-Mamba, which keeps the Segment Anything Model's image encoder frozen and adds a Mamba-Prior module so the general-purpose SAM features are steered toward polyp-specific cues. On a five-dataset protocol in which 1,450 images from Kvasir-SEG and CVC-ClinicDB form the training set and three other datasets are held out, the paper reports that this design beats the CNN and ViT baselines it is compared with, with the largest gaps on the held-out datasets—margins the paper puts at about +4 points on CVC-ColonDB and +3.8 points on ETIS. If the results hold, the approach is a lightweight way to repurpose a large segmentation foundation model for a medical domain without full fine-tuning.","feed_headline":"Mamba-Prior steers SAM to better polyp segmentation on unseen data","feed_subtitle":"Freezing SAM's encoder and adding a Mamba prior lifts zero-shot polyp segmentation by about 4 points.","key_machinery":"The load-bearing component is the Mamba-Prior module, whose named pieces are Multi-scale Spatial Decomposition (MSD), Channel Saliency and Context Accumulation, and Mamba Channel Interaction. MSD builds a spatial feature pyramid with parallel 3x3, 5x5, and 7x7 convolutions; global max and average pooling then separate each scale into salient and contextual cues; and two Mamba layers—state-space models that process sequences with linear complexity—encode long-range dependencies over those channel distributions. The gated Mamba outputs are multiplied back into the multi-scale feature map and concatenated to form the prior, which enters SAM's frozen ViT blocks through cross-attention adapters. A pseudo-mask produced from the adapted encoder is later fed to SAM's decoder as the prompt, removing the need for user-supplied points or boxes, and the network is trained in two stages: encoder-side adapters first under deep supervision, then the decoder with the full mask loss.","core_discovery":"The central claim is that a Mamba-based prior, rather than unfreezing SAM or adding many parameters, is what unlocks strong polyp segmentation with zero-shot generalization. The Mamba-Prior module decomposes the input image at three spatial scales using 3x3, 5x5, and 7x7 convolutions, pools the scales into separate saliency and context channels, and runs both through Mamba layers that model long-range channel dependencies at linear cost; the gated outputs are then multiplied back into the original feature map to form a domain prior that is injected into SAM's frozen ViT blocks through cross-attention adapters. On the seen datasets the model reaches 92.4% mean Dice on Kvasir-SEG and 94.2% on CVC-ClinicDB, while on the unseen CVC-ColonDB and ETIS datasets it reports 77.1% and 78.2% mean intersection-over-union, ahead of every compared CNN and ViT baseline. Ablations attribute the gain to the combination of multi-scale decomposition and Mamba, with single-scale variants losing several points on unseen datasets.","pith_inferences":["A direct test the paper leaves open is whether the Mamba-Prior transfers to other SAM-based medical segmentation tasks, such as organ or lesion segmentation; the same style of gain would be expected if the prior is the active ingredient.","Because all baselines are quoted from prior publications rather than re-run, re-benchmarking them in one codebase with identical preprocessing and evaluation would either confirm or erode the reported margins.","The module's combination of multi-scale spatial decomposition and saliency/context separation is not tied to SAM in principle; applying it to other foundation-segmentation backbones could reveal whether the benefit comes from the prior or from SAM's frozen representation.","The abstract claims superiority over Adapter-based models, but no adapter-only baseline appears in the reported tables; a head-to-head comparison with existing SAM-adapter methods would make that claim testable."],"forward_implications":["If the reported numbers hold, a frozen SAM encoder plus a Mamba-Prior module beats specialized CNN and ViT polyp segmenters on both seen and unseen benchmark datasets.","The pseudo-mask prompting pipeline makes the model usable as a fully automatic colonoscopy screening tool, since no point, box, or text prompt is required at inference.","The module adds only about 9.5% more parameters (from 94M to 103M), so the accuracy gain does not depend on full fine-tuning of SAM.","According to the ablations, multi-scale decomposition is necessary for generalization: single-kernel variants lose roughly 3–4 points on the unseen datasets, while the full multi-scale-plus-Mamba configuration is best everywhere."],"supporting_citations":[{"why":"Supplies the Segment Anything Model backbone whose frozen image encoder and decoder SAM-Mamba adapts.","marker":"[16]"},{"why":"Supplies the cross-attention adapter design used to inject the Mamba-Prior into SAM's ViT blocks.","marker":"[5]"},{"why":"Supplies the Mamba/state-space-model layer used for long-range channel dependency modeling.","marker":"[38]"},{"why":"Defines the training/evaluation protocol (1,450 training images from Kvasir-SEG and CVC-ClinicDB) and the baseline results the paper compares against.","marker":"[10]"},{"why":"One of the two training datasets (Kvasir-SEG) that supplies 900 of the 1,450 training images.","marker":"[14]"},{"why":"One of the two training datasets (CVC-ClinicDB) that supplies 550 of the 1,450 training images.","marker":"[2]"},{"why":"The ETIS dataset used as an unseen zero-shot test set.","marker":"[25]"},{"why":"The CVC-ColonDB dataset used as an unseen zero-shot test set.","marker":"[26]"},{"why":"The EndoScene dataset whose CVC-300 split is used as an unseen test set.","marker":"[28]"},{"why":"Documents that SAM performs poorly on polyp segmentation without domain-specific adaptation, motivating the Mamba-Prior module.","marker":"[35]"}],"fun_headline_variants":["Mamba prior steers SAM to better zero-shot polyp segmentation","SAM-Mamba: Mamba-guided SAM for zero-shot polyp segmentation","Mamba-Prior lifts SAM's zero-shot polyp segmentation accuracy","Freeze SAM, add Mamba: zero-shot polyp segmentation improves"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed superiority rests on baseline numbers quoted from other papers, with the paper itself noting in Section 4.3 that PEFNet's numbers come from M2UNet, rather than on re-runs under SAM-Mamba's exact protocol; if those quoted numbers were obtained under different preprocessing, training, or evaluation details, the margins could shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Mamba prior steers SAM to better zero-shot polyp segmentation","SAM-Mamba: Mamba-guided SAM for zero-shot polyp segmentation","Mamba-Prior lifts SAM's zero-shot polyp segmentation accuracy","Freeze SAM, add Mamba: zero-shot polyp segmentation improves"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1676,"prompt_tokens":988,"completion_tokens":688,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":615}},"tokens_in":604,"tokens_out":688,"duration_ms":7191,"temperature":1.0,"reasoning_tokens":615,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:44:44.990794+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running CTNet, MEGANet, Polyp-PVT, and the other compared methods under SAM-Mamba's exact protocol (1,450 training images, 352x352 inputs, same augmentation, same evaluation protocol) and finding that any baseline reaches or exceeds 78.2% mIoU on ETIS or 77.1% mIoU on CVC-ColonDB would show the claimed zero-shot margins are not reproducible.","supporting_citations":[{"cited_title":"Pranet: Parallel reverse attention network for polyp segmentation","cited_arxiv_id":null,"evidence_quote":"Defines the training/evaluation protocol (1,450 training images from Kvasir-SEG and CVC-ClinicDB) and the baseline results the paper compares against."},{"cited_title":"Kvasir-seg: A segmented polyp dataset","cited_arxiv_id":null,"evidence_quote":"One of the two training datasets (Kvasir-SEG) that supplies 900 of the 1,450 training images."},{"cited_title":"Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs","cited_arxiv_id":null,"evidence_quote":"One of the two training datasets (CVC-ClinicDB) that supplies 550 of the 1,450 training images."},{"cited_title":"Toward embedded detection of polyps in wce images for early diagnosis of colorectal can- cer","cited_arxiv_id":null,"evidence_quote":"The ETIS dataset used as an unseen zero-shot test set."},{"cited_title":"Automated polyp detection in colonoscopy videos using shape and context information","cited_arxiv_id":null,"evidence_quote":"The CVC-ColonDB dataset used as an unseen zero-shot test set."},{"cited_title":"A benchmark for en- doluminal scene segmentation of colonoscopy images","cited_arxiv_id":null,"evidence_quote":"The EndoScene dataset whose CVC-300 split is used as an unseen test set."}],"review_version":1}