{"id":"679d711e-9d8d-419d-bfde-051995760847","arxiv_id":"2507.20056","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FaRMamba adds a multi-scale frequency transform module and a reconstruction auxiliary encoder to a Vision Mamba segmentation encoder, reporting improved Dice and mIoU on CAMUS, Mouse-cochlea, and Kvasir-Seg datasets.","lead":"The paper presents FaRMamba, a Mamba-based medical image segmentation network that adds frequency-domain processing and a reconstruction branch to recover detail and spatial structure. If the reported gains hold, it offers a modest accuracy boost on three medical imaging benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SSRAE's attention masks are built from ground-truth segmentation labels during training; the paper never specifies how masks are obtained at inference, so the reported gains may stem from label leakage rather than the proposed mechanism.","rationale":"I read the paper as claiming that MSFM and SSRAE improve segmentation on three benchmarks beyond CNN, Transformer, and Mamba baselines. For that claim to hold, the evaluation must reflect genuine test-time behavior. The paper's own Section 3.4 says SSRAE derives binary region masks from segmentation labels and adds them to attention logits during training. Since the Mamba encoder is shared across the segmentation and reconstruction branches, this label-derived mask can leak ground-truth region information into encoder features during training, inflating the reported Dice. No test-time protocol is described: the paper does not say whether SSRAE is disabled at inference, whether masks come from predictions, or whether ground-truth masks are used at test. Without that information, the improvements in Tables 1 and 2 cannot be attributed to the proposed mechanisms. The missing comparisons against GlobalMamba, EM-Net, and P-Mamba, and the absence of computational-overhead numbers, are secondary and fixable. The label-leak ambiguity is primary because it threatens the internal validity of all empirical claims. I therefore recommend REJECT as written; if a controlled experiment shows that masks use only predicted labels at both train and test, the concern would be resolved and the earlier CONDITIONAL verdict could stand. The reader's weakest assumption about added parameters or transform choice is related but not identical; the deeper issue is the train/test asymmetry in the label-guided attention.","tokens_in":10711,"tokens_out":6878,"duration_ms":71265,"concrete_test":"Inspect the implementation (if released, or ask the authors) to determine whether SSRAE's region-attention masks are computed from ground-truth labels at train time and from predicted masks or nothing at test time. Independently, train two FaRMamba variants on CAMUS 2CH: one exactly as described, and one where the region mask is either omitted or computed from the segmentation branch's own soft prediction at both train and test. If the original variant beats the label-free variant by more than roughly 1 DSC, or if the label-free variant no longer outperforms U-Mamba, the reported gain is explained by label leakage rather than the proposed mechanism.","verdict_should_be":"REJECT","load_bearing_attack":"Section 3.4 states: 'Binary region masks are first derived from segmentation labels and then resized to match the feature map resolution... this mask is added directly to the attention logits, thereby constraining attention to remain within each labeled region.' This label-derived mask is used during training, while at inference no ground-truth masks exist. The paper does not state whether SSRAE is dropped at inference, whether masks are replaced by the segmentation branch's own predictions, or how the train/test mismatch is handled. Because the Mamba encoder is shared between the segmentation branch and SSRAE, label-guided attention can inject region information into the encoder's learned representation during training, directly inflating validation Dice relative to label-free baselines. The ablation in Table 2 attributes roughly 2.5 DSC (89.81 vs 87.25) to SSRAE plus MSFM; if this gain is due to label-derived masks, the central claim of consistent superiority is not supported. This concern is internal to the paper's own description, making it more load-bearing than missing baselines or absent error bars. The module is also called 'self-supervised' while using segmentation labels, which is at minimum a terminological inconsistency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FaRMamba, a Vision Mamba variant for medical image segmentation. It argues that Mamba's patch tokenization and 1D serialization cause local high-frequency information deficiency (LHICD) and 2D spatial structure degradation (2D-SSD), and addresses these with two modules: a Multi-Scale Frequency Transform Module (MSFM) that processes features through DWT, FFT, or DCT in separate bands, and a Self-Supervised Reconstruction Auxiliary Encoder (SSRAE) that reconstructs degraded encoder outputs with label-derived region attention. Experiments on Mouse-cochlea (private MRI), CAMUS (echocardiography), and Kvasir-Seg (endoscopy) report Dice and MIoU improvements over CNN, Transformer, and Mamba baselines, with an ablation on CAMUS 2CH attributing gains to both modules. The manuscript also claims modality-dependent superiority of the three frequency transforms.","tokens_in":10937,"tokens_out":3985,"duration_ms":42082,"significance":"If validated, FaRMamba would be a useful and practical frequency-aware augmentation for Mamba-based medical segmentation backbones, with the interesting observation that different frequency transforms suit different imaging modalities. The paper contributes detailed architectural descriptions, ablations of both modules, and evaluations on three datasets, including a public benchmark. The main empirical claim is plausible but not yet established: the SSRAE module is trained using ground-truth segmentation labels with no specified inference-time treatment, which raises a label-leakage concern; the quantitative comparison lacks variance information, statistical tests, and a fixed selection protocol across variants; and the closest frequency-based Mamba competitors cited in Section 2.1 are not included in Table 1. The central derivation is internally consistent, but these load-bearing issues require resolution before the reported gains can be attributed to the proposed mechanism.","major_comments":[{"comment":"The paper states that 'Binary region masks are first derived from segmentation labels and then resized to match the feature map resolution' and that this mask 'is added directly to the attention logits, thereby constraining attention to remain within each labeled region.' It does not specify how these masks are obtained at inference time, where ground-truth labels are unavailable. Because the Mamba encoder is shared between the segmentation branch and SSRAE, label-derived attention masks can inject region/class information into the encoder's learned representation during training, making the roughly 2.5 DSC gain attributed to 'SSRAE + MSFM' in Table 2 (89.81 vs 87.25) ambiguous evidence for the proposed mechanism. The same section also calls this process 'self-supervised' and says it trains 'without extra labels,' which is contradicted by the use of segmentation labels. Please state the exact inference-time behavior (drop SSRAE, use predicted masks, or another mechanism) and provide an ablation with a label-free version of the region attention.","section":"Section 3.4, SSRAE"},{"comment":"No standard deviations, confidence intervals, or repeated-run information are reported, and many reported gaps are small (e.g., 95.98 vs 95.75 on CAMUS 4CH, 60.89 vs 59.82 on Mouse-cochlea, 88.97 vs 87.39 on Kvasir-Seg). Moreover, the 'Ours' entry is selected per dataset as the best of FFT/DCT/DWT, which inflates the apparent advantage over fixed baselines. The claim that FaRMamba 'consistently outperforms' competitive methods needs means and standard deviations over at least three seeds, a fixed variant or an appropriate multi-comparison correction, and ideally significance tests.","section":"Tables 1 and 2, Section 4.2-4.3"},{"comment":"The related work discusses GlobalMamba, EM-Net, and P-Mamba as frequency-based Mamba methods, but Table 1 contains no comparison against any of them. Since these are the closest competitors on the same design axis, their omission prevents the reader from assessing whether MSFM provides gains beyond existing frequency-aware Mamba approaches. Please add these baselines under the same training protocol, or explicitly explain why a direct comparison is not possible.","section":"Section 2.1 and Section 4.2"},{"comment":"The abstract and conclusion claim 'superior boundary accuracy, detail preservation, and global coherence,' but experiments report only DSC and MIoU. No boundary-specific metric (e.g., HD95, boundary F1, or boundary IoU) and no reconstruction-quality metric are provided, so the specific qualitative claims are not directly supported by the reported measurements. Please add such metrics or soften the claims.","section":"Abstract and Section 4.2"}],"minor_comments":[{"comment":"The DCT formula is missing the v variable in the second cosine factor: it should be cos(pi(2y+1)v/2N), not cos(pi(2y+1)/2N).","section":"Equation (4)"},{"comment":"There are several typos and inconsistencies: 'aafrequency' in Section 3.4, 'UMamba' vs 'Umamba' in Table 1 and elsewhere, and 'First,' at the start of the Introduction. A careful proofread is needed.","section":"Throughout"},{"comment":"The Mouse-cochlea dataset is described as private with the placeholder 'xxUniversity'; the institution name and ethical approval details should be provided, and any public release plan stated, to support reproducibility.","section":"Section 4.1"},{"comment":"The formatting of Table 2 is difficult to parse: the meaning of the checkmark columns and the reported DSC/MIoU triples should be clarified in the caption or table structure.","section":"Table 2"},{"comment":"The SSRAE description mentions 'multi-head self-attention' even though the architecture is Mamba-based and the main encoder uses SS2D; please clarify whether SSRAE contains attention layers or whether this wording is a typo.","section":"Section 3.4"}],"recommendation":"major_revision","confidential_remarks":"The label-leakage concern in Section 3.4 is the decisive issue for this paper. I would not recommend acceptance until the authors specify the inference-time treatment of SSRAE and provide a label-free ablation that isolates the proposed mechanism. The missing error bars and the per-dataset selection of the best transform also need to be addressed, since they directly affect the 'consistently outperforms' claim. The private Mouse-cochlea dataset further limits reproducibility, so a public-data demonstration or a release plan would strengthen the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know about FaRMamba is that the architecture is a sensible, incremental combination of known ideas, and the experiments are directionally encouraging—but there is one ambiguity in the method section that you need to resolve before trusting the numbers. Section 3.4 describes SSRAE as 'self-supervised' yet builds its attention masks from ground-truth segmentation labels: 'Binary region masks are first derived from segmentation labels and then resized... this mask is added directly to the attention logits, thereby constraining attention to remain within each labeled region.' The paper never states whether these masks are dropped at inference, replaced by the segmentation branch's own predictions, or handled some other way. Because the Mamba encoder is shared between the reconstruction branch and the segmentation branch, label-guided attention during training can inject region information into the features used at test time. That makes the ablation gains—about 2.5 Dice for SSRAE+MSFM in Table 2—hard to attribute to the proposed mechanism. This is not a missing-baseline issue; it is internal to the paper's own description.\n\nWhat is genuinely new: the dual-module integration—MSFM for frequency-band restoration and SSRAE for reconstruction-based spatial coherence—on a Mamba backbone is a combination I have not seen exactly this way. The paper also does a fair job of motivating why different transforms might suit different modalities, and the per-dataset transform ranking (DWT on ultrasound, FFT on MRI, DCT on endoscopy) is a nice empirical observation even if it was selected post hoc.\n\nNow the soft spots, in proportion. The label-leakage ambiguity is the biggest. Second, there are no error bars or significance tests anywhere; given the reported gains are often around one Dice point, that matters. Third, the paper cites GlobalMamba, EM-Net, and P-Mamba as related frequency-Mamba work but does not compare against them in Table 1. Fourth, reporting Ours-FFT, Ours-DCT, Ours-DWT and then cherry-picking the best per dataset inflates the apparent consistency of the gains; the abstract's 'consistently outperforms' claim is stronger than the table supports.\n\nWho is this for? Readers working on Mamba-based medical segmentation who want a concrete recipe for adding frequency and reconstruction branches. The architecture idea is worth taking seriously, but the current write-up needs one clarif\n","headline":"A clean, incremental Mamba+frequency architecture with a potentially load-bearing ambiguity: the auxiliary 'self-supervised' branch is trained with ground-truth masks and the paper never says what happens at inference.","tokens_in":11421,"tokens_out":1644,"would_cite":false,"duration_ms":18715,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U10","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"FaRMamba claims to fix Vision Mamba's loss of high-frequency detail and 2D spatial structure by adding a multi-scale frequency transform module and a self-supervised reconstruction auxiliary encoder, outperforming CNN–Transformer hybrids…","keywords":["Medical image segmentation","Vision Mamba","Frequency domain analysis","Self-supervised reconstruction","State-space models","Multi-scale frequency transform","Discrete wavelet transform","Fast Fourier transform"],"falsifier":"Compare FaRMamba against a Mamba baseline augmented with a same-parameter-count convolutional high-frequency enhancement branch: if the Dice gain matches or exceeds FaRMamba's, the low-pass/1D-serialization mechanism is not needed to explain the results; alternatively, compute the high-frequency spectral energy of feature maps before and after the first SS2D block to see whether patch embedding and recurrence actually suppress it.","tokens_in":10523,"feed_emoji":"🩻","tokens_out":6152,"duration_ms":53015,"temperature":0.7,"pith_summary":"The paper argues that Vision Mamba, although efficient at modeling long-range dependencies, degrades medical image segmentation because its patch tokenization and one-dimensional scanning act like a low-pass filter, suppressing boundary and texture detail, and because it linearizes 2D structure. To counter this, FaRMamba adds two modules: MSFM, which isolates and reconstructs multi-band spectra via wavelet, cosine, or Fourier transforms, and SSRAE, a parameter-sharing Mamba encoder that reconstructs high-frequency features from a deliberately degraded input. The paper reports that on cardiac ultrasound, MRI cochlea, and endoscopic polyp datasets, FaRMamba variants beat CNN–Transformer hybrids and existing Mamba models on Dice and mIoU, with the best frequency transform depending on the imaging modality. A sympathetic reader would take this as evidence that Mamba's efficiency can be combined with explicit frequency restoration to get both global coherence and fine detail.","feed_headline":"Mamba segmentation gets sharper edges with frequency-boosted model","feed_subtitle":"On cardiac ultrasound, MRI cochlea, and endoscopy, the model beats CNN–Transformer hybrids and Mamba variants.","key_machinery":"The central objects are the two modules. MSFM projects feature maps into a frequency domain (DWT, FFT, or DCT), applies masks to isolate spectral bands, processes them with scale-adapted convolutions, inverse-transforms, and fuses them back into the spatial feature stream, restoring high-frequency cues that the patch embedding and 1D recurrence suppressed. SSRAE is a Mamba-based auxiliary encoder that shares architecture and parameters with the main segmentation encoder, takes a degraded copy of the input, and is trained to reconstruct the main encoder's high-frequency feature maps, with label-derived binary region masks constraining attention inside each annotated region; a progressive composite loss of L1, cosine similarity, and gradient terms supervises it. The load-bearing physical claim underneath is that the state-space recurrence implements a stable linear dynamical system whose impulse response decays exponentially, so it behaves as a low-pass filter that attenuates high-frequency components during scanning.","core_discovery":"On its own terms, the paper's central discovery is a two-part failure analysis and remedy. First, it identifies two concrete deficiencies in Vision Mamba for medical images: local high-frequency capture deficiency (LHICD), caused by patch embedding averaging and the exponential-decay low-pass behavior of the state-space recurrence, and two-dimensional spatial structure degradation (2D-SSD), caused by serializing the image into a 1D causal sequence. Second, it claims both are substantially repaired by MSFM and SSRAE, and that the combination outperforms the compared CNN, transformer, and Mamba baselines on CAMUS, Mouse-cochlea, and Kvasir-Seg, delivering better boundary accuracy, detail preservation, and global coherence without prohibitive computational overhead. The paper also claims a modality-dependent ranking of the three transforms: DWT best on noisy ultrasound, FFT best on MRI because of native k-space sampling, DCT best on textured endoscopy.","pith_inferences":["The paper's low-pass story is plausible but not directly measured; one could test it by comparing the high-frequency energy of feature maps before and after Mamba blocks, and by replacing MSFM with an equally sized convolutional high-pass block.","Because no single transform always wins, a learned soft combination or input-adaptive selection of DWT, FFT, and DCT might beat the fixed per-dataset choice.","SSRAE uses ground-truth region labels to build attention masks; in low-label regimes, pseudo-labels or unsupervised reconstruction losses could let the same idea work with fewer annotations.","If the 1D-serialization loss of 2D adjacency is the mechanism, 3D or video Mamba segmentation should suffer similar or larger 2D structure degradation, making the reconstruction aid more valuable there."],"forward_implications":["Vision Mamba encoders can recover boundary detail by injecting frequency-band-restored features, so Mamba-based segmentation need not trade global modeling for fine edges.","The transform choice matters by modality: use DWT on speckle-heavy ultrasound, FFT on MRI, DCT on textured endoscopy, suggesting modality-aware frequency design.","A shared-encoder reconstruction auxiliary task improves segmentation without a separate large network or extra labels beyond the segmentation mask.","MSFM and SSRAE are complementary; ablations show each contributes Dice gains and the full configuration is best.","The approach extends the same architecture to multiple modalities without redesign, since only the transform changes."],"supporting_citations":[{"why":"Supplies the selective state-space model that the backbone extends.","marker":"[10]"},{"why":"Supplies the VSS block and SS2D bidirectional scanning used as the main encoder.","marker":"[41]"},{"why":"Supplies the U-Mamba decoder and a primary Mamba baseline for medical segmentation.","marker":"[25]"},{"why":"Provides the exponential-decay impulse response argument for low-pass behavior of linear dynamical systems.","marker":"[29]"},{"why":"Provides evidence that state-space models have a frequency bias toward low frequencies.","marker":"[38]"},{"why":"Provides the CAMUS echocardiography dataset used for evaluation.","marker":"[19]"},{"why":"Provides the Kvasir-SEG polyp dataset used for evaluation.","marker":"[16]"},{"why":"Provides the reconstruction-aided segmentation evidence motivating SSRAE.","marker":"[26]"}],"fun_headline_variants":["FaRMamba adds frequency boosts to Mamba for sharper medical segmentation","Frequency-aware Mamba variant restores high-frequency details in medical images","FaRMamba fixes Mamba's high-frequency loss for better segmentation boundaries","Dual-module Mamba enhances medical segmentation with frequency and reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the observed losses of high-frequency detail and spatial structure in Vision Mamba come specifically from patch tokenization and 1D serialization, so that fixing those two mechanisms is what produces the accuracy gains rather than the added parameters or the choice of frequency transform.","fun_headline_variants_meta":{"raw":{"variants":["FaRMamba adds frequency boosts to Mamba for sharper medical segmentation","Frequency-aware Mamba variant restores high-frequency details in medical images","FaRMamba fixes Mamba's high-frequency loss for better segmentation boundaries","Dual-module Mamba enhances medical segmentation with frequency and reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000596,"raw_usage":{"total_tokens":2826,"prompt_tokens":1020,"completion_tokens":1806,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":1730}},"tokens_in":636,"tokens_out":1806,"duration_ms":13296,"temperature":1.0,"reasoning_tokens":1730,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:50:00.348689+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare FaRMamba against a Mamba baseline augmented with a same-parameter-count convolutional high-frequency enhancement branch: if the Dice gain matches or exceeds FaRMamba's, the low-pass/1D-serialization mechanism is not needed to explain the results; alternatively, compute the high-frequency spectral energy of feature maps before and after the first SS2D block to see whether patch embedding and recurrence actually suppress it.","supporting_citations":[{"cited_title":"In: Proceed- ings of the 41st International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Supplies the VSS block and SS2D bidirectional scanning used as the main encoder."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the exponential-decay impulse response argument for low-pass behavior of linear dynamical systems."},{"cited_title":"In: Frangi, A.F., Schnabel, J.A., Davatzikos, C., Alberola-López, C., Fichtinger, G","cited_arxiv_id":null,"evidence_quote":"Provides the reconstruction-aided segmentation evidence motivating SSRAE."}],"review_version":2}