{"id":"423e34be-8983-4d05-a7f0-1a48fac77cce","arxiv_id":"2506.23086","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FMC-Net improves vertebrae segmentation in CT and MRI images by using wavelet transforms for blur-resistant downsampling and multi-scale state space modeling for context.","lead":"A new computer program that outlines individual bones of the spine in CT and MRI scans splits each scan into fine detail and overall shape, then processes the two parts separately to handle blurry images and similar-looking vertebrae. On two public medical datasets it reports better accuracy than previous programs, and the code is freely available.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No error bars, no significance tests, and a numerical mismatch between the results text and Table 1 leave the SOTA claim unverified; the true margins over the strongest baselines are about 1 DSC point.","rationale":"The reader's CONDITIONAL verdict focuses on missing error bars; I agree that is the central weakness, and the table-vs-text discrepancy strengthens it rather than replacing it. The paper has real strengths: public code, nnU-Net integration, ablations on LUMBAR, and evaluations on both CT and MRI. But the strongest claim is comparative, and the comparison is currently unverifiable: the reported improvements are around one DSC point over the correct best baselines, no uncertainty is provided, and the prose misidentifies the baseline. These issues are independently checkable by re-running with seeds and by recomputing Table 1's differences. If the error-bars test shows overlapping distributions, the SOTA claim fails; if it shows consistent gains, the claim may stand after correcting the text. Therefore the appropriate disposition remains CONDITIONAL rather than ACCEPT or REJECT: the paper is plausibly sound but needs statistical validation and a correction. My concern is the same in spirit as the reader's, with an added concrete inconsistency, hence 'partial' agreement.","tokens_in":7534,"tokens_out":7495,"duration_ms":80654,"concrete_test":"Using the released code (github.com/anaanaa/FMCNet), first recover the exact data split and evaluation protocol from the nnU-Net configuration files. Then retrain FMC-Net and the three Mamba baselines (UMamba, SegMamba, MambaClinix) on the same split with at least 5 random seeds each; compute mean and standard deviation of DSC and HD95, and run a paired Wilcoxon signed-rank test over test cases. Separately, recompute all pairwise differences from Table 1 to locate the '1.64%' margin; if it belongs to UMamba rather than SegMamba, the prose must be corrected and the SOTA comparison re-evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that FMC-Net achieves state-of-the-art on VERSE2019 and LUMBAR. Table 1 is the sole quantitative support, but it reports one run with no error bars, no repeated seeds, and no paired significance test. This is load-bearing because the margins over the actually strongest baselines are small: 1.64 DSC points over UMamba on VERSE2019 (78.28 vs 79.92) and 0.96 DSC points over SegMamba on LUMBAR (78.17 vs 79.13). In Section 3, the paper states 'Compared with the best-performing Mamba-based method, SegMamba, our approach improves the mean DSC per class by 1.64%', but Table 1 lists SegMamba at 72.22 on VERSE2019, making the gap 7.70, not 1.64. The 1.64 gap matches UMamba, not SegMamba, so either the text or the table is wrong. The train/test split for VERSE2019 and LUMBAR also is not stated, so the baseline numbers may not be protocol-comparable. A one-point advantage from a single seed is within seed noise for 3D medical segmentation, so as reported the SOTA claim is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FMC-Net, a 3D encoder-decoder for vertebrae segmentation that replaces standard downsampling with discrete wavelet transform, splits features into high- and low-frequency components, enhances them with a High-frequency Feature Refinement (HFR) module and a Multi-granularity State Space Model (MG-SSM), and reconstructs features via inverse wavelet transform. The method is evaluated on VERSE2019 (CT) and LUMBAR (MRI) datasets, reporting DSC and HD95 against CNN-, Transformer-, and Mamba-based baselines, together with an ablation study on LUMBAR. The central claim is that FMC-Net achieves state-of-the-art performance on both datasets.","tokens_in":7809,"tokens_out":2762,"duration_ms":27923,"significance":"If the reported results are reliable, the paper would make a modest but useful contribution to vertebrae segmentation: it combines two established ideas (wavelet-domain downsampling and state-space modeling) in a new way, addresses the clinically relevant problem of blurred and morphologically similar vertebrae, and provides a public codebase. The ablation study on LUMBAR shows large gains from the proposed components (DSC from 71.09 to 79.13), and the design is plausible and well-motivated. However, the significance as stated depends on the SOTA claim, which currently rests on a single training run and on a numerical mismatch between the text and Table 1. The gains over the strongest baselines are small (1.64 DSC over UMamba on VERSE2019 and 0.96 DSC over SegMamba on LUMBAR), and no evidence rules out seed or protocol variation. The manuscript is therefore of interest to the community, but the SOTA claim is not yet established at the level of certainty the paper asserts.","major_comments":[{"comment":"There is a direct numerical contradiction in the SOTA claim. The text states that compared with the best-performing Mamba-based method, SegMamba, the approach improves the mean DSC per class by 1.64% on VERSE2019. Table 1, however, lists SegMamba at 72.22 DSC and FMC-Net at 79.92 DSC on VERSE2019, a difference of 7.70 percentage points, not 1.64. The 1.64-point gap corresponds to UMamba (78.28 vs. 79.92), not SegMamba. Either the text identifies the wrong baseline or the table reports different numbers. This is load-bearing because the paper's central claim is state-of-the-art performance, and the reader cannot tell which baseline is actually the strongest. The authors must correct the text/table inconsistency and clearly state which baseline is the best prior method on each dataset.","section":"Section 3, Table 1 and text"},{"comment":"No error bars, confidence intervals, repeated runs, or significance tests are reported for any of the quantitative results. The margins over the strongest baselines are approximately 1.64 DSC points on VERSE2019 and 0.96 DSC points on LUMBAR; in 3D medical segmentation, single-seed differences of this size are typically within random seed variation and hyperparameter sensitivity. Without at least three independent runs with mean and standard deviation, or a paired statistical test, the claim that FMC-Net outperforms state-of-the-art methods is not established. The same issue affects the ablation study in Table 2: for example, adding HFR improves DSC by 3.55 points (71.75 to 75.30) but worsens HD95 from 22.12 to 24.52, and adding MG-SSM improves DSC by 1.79 points (76.56 to 78.35?) while worsening HD95 from 16.72 to 17.13 in the final row. Without variance information, the claim that each component 'significantly enhances' performance is unsupported.","section":"Section 3, Tables 1 and 2"},{"comment":"The experimental protocol is incompletely specified. The paper does not state how the VERSE2019 (160 scans) and LUMBAR (156 scans) datasets were split into training, validation, and test sets, nor whether the baseline numbers in Table 1 were obtained under the same split and preprocessing as the proposed method. If baseline results are quoted from their original papers and those papers use different splits or evaluation protocols, the cross-method comparison in Table 1 is not valid. The authors should state the exact split, the number of folds if cross-validation is used, and either re-evaluate all baselines under the identical protocol or justify that the cited numbers are protocol-comparable.","section":"Section 3, Datasets and Implementation Details"}],"minor_comments":[{"comment":"Equation (8) appears to have a missing closing parenthesis: X2 = phi(Conv([Fd1, Fd2, Fd3])) is presumably intended. Please fix the typo.","section":"Section 2.3, Eq. (8)"},{"comment":"The checkmark pattern in Table 2 is ambiguous. The rows with the same number of checkmarks ('✓ ✓') cannot be reliably mapped to components without carefully cross-referencing the text; for example, the row at 76.56 could be Baseline+DWT+MG-SSM or Baseline+HFR+MG-SSM. Please label each row explicitly or use separate columns with Y/N entries.","section":"Section 3, Table 2"},{"comment":"The title claims 'Efficient' segmentation, but no efficiency metrics (parameters, FLOPs, inference time, GPU memory) are reported, and the only efficiency evidence is the linear-complexity statement about SSMs. Please add quantitative efficiency comparisons with the baselines, or soften the efficiency claim.","section":"Title and Section 2"},{"comment":"The dilation rates d1, d2, d3 in MG-SSM and the group counts 2^i in HFR are fixed hyperparameters, but no sensitivity analysis or ablation is provided for them. Since the multi-granularity argument depends on these choices, a short ablation or at least a statement of how they were chosen would strengthen the paper.","section":"Section 2.2 and Table 1"},{"comment":"There are several typographical errors: 'he evaluation results' should be 'The evaluation results' in Section 3; 'are in illustrated Fig. 1 (a)' should be 'are illustrated in Fig. 1 (a)' in Section 1; and Eq. (1) lists eight sub-bands but uses a notation where F_i_{llh,...,hhh} is written with only one representative subscript. These should be cleaned up.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competent empirical architecture study with a clear motivation and a plausible design, but the central SOTA claim is currently not verifiable because of the text/table mismatch, missing error bars, and an underspecified evaluation protocol. The numerical mismatch in Section 3 is especially concerning because it suggests that the authors may have compared against different baselines in the text and the table. I would like to see a careful revision that resolves the inconsistency, reports multi-seed statistics, and specifies the data split, before the manuscript is reconsidered. The scope of the work is suitable for the venue, and the availability of code is a positive signal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a decent incremental architecture paper for vertebrae segmentation, but the SOTA claim is not yet established because the reported margins over the strongest baselines are about one DSC point and come from a single run. There is also a numerical mismatch in the text that needs fixing.\n\nWhat's new: the paper adapts wavelet transform downsampling from semantic segmentation (Xu et al.) to 3D vertebrae segmentation and combines it with a high-frequency refinement module and a multi-granularity state space model. The specific combination is new, and the ablation on LUMBAR shows each component adds something: baseline 71.09 DSC, adding DWT gives 71.75, HFR 75.30, MG-SSM 76.56, full 79.13. That internal evidence is the strongest part of the paper. The code is public, baselines include nnU-Net, Swin-UNetR, CoTR, and three Mamba variants, and the evaluation covers both CT and MRI.\n\nSoft spots: the biggest is a concrete error in Section 3. The text claims improvement of 1.64% over 'the best-performing Mamba-based method, SegMamba' on VERSE2019, but Table 1 has SegMamba at 72.22 mean DSC and FMC-Net at 79.92; the 1.64 gap actually matches UMamba (78.28). Either the table is wrong or the text names the wrong baseline. Second, no error bars, repeated seeds, or significance tests anywhere. Given that the true margins over UMamba (CT) and SegMamba (MRI) are 1.64 and 0.96 DSC points, the SOTA claim is within typical seed noise for 3D medical segmentation. Third, the data split for both datasets is not stated, so protocol comparability with published baselines is unclear. Fourth, MRI evaluation is limited to T1-weighted LUMBAR, with no T2 or contrast-enhanced T1, so the generalization to MRI is narrow.\n\nThe paper is coherent and the thinking is clear; the flaws are missing statistics rather than bad logic. A serious referee should see this, but the authors need to correct the SegMamba/UMamba discrepancy, report variance over multiple runs, and specify the split before the SOTA claim can be taken at face value.","headline":"Solid incremental architecture paper, but the SOTA claim is undercut by a numerical slip and single-run margins of about 1 DSC point.","tokens_in":8352,"tokens_out":2942,"would_cite":false,"duration_ms":28129,"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 wavelet-based segmentation network achieves state-of-the-art vertebrae segmentation on CT and MRI by processing high- and low-frequency signal components separately, restoring blurred detail and distinguishing look-alike vertebrae.","keywords":["vertebrae segmentation","wavelet transform","state space model","multi-granularity context","high-frequency feature refinement","computed tomography","magnetic resonance imaging","medical image segmentation"],"falsifier":"Run the released code on both benchmarks under the stated settings with at least five random seeds; the central claim would be undermined if the seed-to-seed variation in FMC-Net's mean DSC overlaps the reported scores of the best prior method in Table 1 (a 1.64-point margin on VERSE2019 and a 0.96-point margin on LUMBAR), or if the DWT-Sample component changes the DSC by less than the 0.66 points reported in the ablation.","tokens_in":7323,"feed_emoji":"🦴","tokens_out":12509,"duration_ms":112965,"temperature":0.7,"pith_summary":"FMC-Net is designed to solve two specific failures in vertebrae segmentation: blur and noise that smudge fine detail in CT and MRI, and the near-identical appearance of adjacent vertebrae that makes them hard to tell apart. The paper's central claim is that replacing ordinary downsampling with a wavelet transform, then treating the high- and low-frequency components differently, fixes both problems. For high frequencies it restores edges and texture while filtering noise; for low frequencies it builds multi-scale context models that capture long-range spatial dependencies cheaply. On the VERSE2019 CT benchmark the method reports a mean Dice score of 79.92% (HD95 of 3.04 mm), and on the LUMBAR MRI benchmark 79.13% DSC (HD95 17.13 mm), both above the prior methods compared. If the reported gains hold up, the practical payoff is more reliable automated labeling of individual vertebrae for surgical planning and diagnosis.","feed_headline":"79.9% Dice: wavelet network tops vertebrae segmentation","feed_subtitle":"Frequency-split networks restore blurry detail and separate look-alike vertebrae, beating prior methods on CT and MRI.","key_machinery":"The engine of the argument is the discrete wavelet transform used as a lossless downsampling step (WTD): it splits each 3D feature map into one low-frequency band and seven high-frequency sub-bands instead of pooling away information. The high-frequency bands go through a High-frequency Feature Refinement (HFR) module with two attention paths, one amplifying salient structures (max-pooling plus softmax weighting) and one smoothing noise (average pooling); the low-frequency band goes through a Multi-granularity State Space Model (MG-SSM) that runs three parallel dilated depthwise convolutions into three state-space modules and merges their outputs, capturing long-range dependencies at linear complexity. A Wavelet Transform Upsampling (WTU) step reuses encoder wavelet coefficients during decoding. Each named component does specific work: WTD avoids information loss, HFR restores fine detail, and MG-SSM supplies the spatially-varying multi-scale context the paper argues is needed to tell similar vertebrae apart.","core_discovery":"The discovery this paper puts forward is a segmentation architecture whose performance gains come from where information is lost, not from a bigger model. FMC-Net replaces max-pooling and interpolation with a discrete wavelet transform, which splits each feature map into one low-frequency band (the bulk of the image content) and seven high-frequency bands (edges, texture, noise) without discarding information. High-frequency bands pass through a High-frequency Feature Refinement (HFR) module with two spatial-attention paths — one that amplifies salient structure via max-pooling, one that smooths noise via average pooling. Low-frequency bands pass through a Multi-granularity State Space Model (MG-SSM): three dilated depthwise convolutions feed three state-space modules whose outputs are merged, capturing contexts at multiple receptive fields with linear complexity in sequence length. The wavelet-based sampling is also used in the decoder (Wavelet Transform Upsampling), so encoder detail is reused during upsampling. The author's claim is that this combination specifically targets the two failure modes — blur and vertebra similarity — and that it outperforms previous convolutional, transformer, and state-space-model methods on both CT and MRI vertebrae benchmarks.","pith_inferences":["My inference: if the failure-mode analysis is right, the accuracy gains should concentrate on blurry slices and on the boundaries between adjacent vertebrae; a stratified error analysis by sharpness and inter-vertebral border distance would confirm where the method actually helps.","My inference: the component-wise ablation numbers come from single runs, so the exact size of each module's contribution is not yet established; multi-seed ablations would turn the observed ordering of gains into a quantitative claim.","My inference: the frequency-split recipe (wavelet downsampling plus separate high/low treatment) is a general remedy for repetitive-anatomy segmentation with blur, and could be tested on ribs, teeth, or finger bones where the same two failure modes dominate.","My inference: the method's reported improvements over the closest prior model are in the 1–2 DSC point range; whether that margin is clinically meaningful for vertebrae labeling is a separate question the paper does not address, since DSC gains of that size may or may not change surgical planning decisions."],"forward_implications":["Because the ablation isolates each piece, the wavelet sampling, the high-frequency refinement, and the multi-granularity state-space module each can be added to an existing U-shaped baseline independently and raise the Dice score.","The same network, with only patch size and epoch count changed, produces the reported results on both CT and MRI, so the design is not tied to one imaging modality.","The underlying state-space formulation keeps long-range context capture at linear complexity, avoiding the quadratic cost of self-attention that limits transformer-based segmentation of large 3D volumes.","Public release of the code means the reported numbers on both benchmarks can be checked directly against the paper's tables."],"supporting_citations":[{"why":"Supplies the Haar wavelet downsampling idea that FMC-Net adapts to 3D encoder features for lossless downsampling.","marker":"[12]"},{"why":"Is the VERSE2019 CT benchmark whose mean DSC and HD95 are the paper's primary comparison table.","marker":"[15]"},{"why":"Is the LUMBAR MRI dataset providing the second benchmark and the lumbar vertebra labels used in the evaluation.","marker":"[16]"},{"why":"Is the CNN baseline and the self-configuring framework whose preprocessing and training pipeline the method is integrated with.","marker":"[17]"},{"why":"Is the strongest state-space-model prior method; the paper's headline margin of 1.64 DSC points is measured against it on VERSE2019.","marker":"[21]"},{"why":"Provides a further state-space-model baseline compared on both datasets.","marker":"[20]"},{"why":"Is the selective state-space model that the VSSM modules inside MG-SSM are built on.","marker":"[11]"},{"why":"Supplies the earlier simplified state-space layer design motivating the linear-complexity sequence modeling component.","marker":"[10]"}],"fun_headline_variants":["Wavelet split sharpens blurry vertebrae in CT and MRI","Frequency-aware network tops vertebrae Dice with wavelet downsampling","Multi-granularity state space boosts vertebrae segmentation accuracy","FMC-Net: wavelet detail and context beat prior vertebrae methods"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported performance gaps over the prior methods are genuine effects of the architecture and not artifacts of a single training run: the paper reports no error bars, repeated trials, or significance tests, so the observed 1–2 point DSC advantages are assumed to be stable.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet split sharpens blurry vertebrae in CT and MRI","Frequency-aware network tops vertebrae Dice with wavelet downsampling","Multi-granularity state space boosts vertebrae segmentation accuracy","FMC-Net: wavelet detail and context beat prior vertebrae methods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000261,"raw_usage":{"total_tokens":1628,"prompt_tokens":1016,"completion_tokens":612,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":551}},"tokens_in":632,"tokens_out":612,"duration_ms":6983,"temperature":1.0,"reasoning_tokens":551,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:49:58.506633+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code on both benchmarks under the stated settings with at least five random seeds; the central claim would be undermined if the seed-to-seed variation in FMC-Net's mean DSC overlaps the reported scores of the best prior method in Table 1 (a 1.64-point margin on VERSE2019 and a 0.96-point margin on LUMBAR), or if the DWT-Sample component changes the DSC by less than the 0.66 points reported in the ablation.","supporting_citations":[{"cited_title":"Pattern Recognition, 143, 109819 (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the Haar wavelet downsampling idea that FMC-Net adapts to 3D encoder features for lossless downsampling."},{"cited_title":"E., Bayat, A., Löffler, M., Liebl, H., Li, H., Kirschke, J","cited_arxiv_id":null,"evidence_quote":"Is the VERSE2019 CT benchmark whose mean DSC and HD95 are the paper's primary comparison table."},{"cited_title":"A., Becherucci, E","cited_arxiv_id":null,"evidence_quote":"Is the LUMBAR MRI dataset providing the second benchmark and the lumbar vertebra labels used in the evaluation."},{"cited_title":"F., Kohl, S","cited_arxiv_id":null,"evidence_quote":"Is the CNN baseline and the self-configuring framework whose preprocessing and training pipeline the method is integrated with."},{"cited_title":"In: Medical Image Computing and Computer-Assisted Intervention, pp","cited_arxiv_id":null,"evidence_quote":"Is the strongest state-space-model prior method; the paper's headline margin of 1.64 DSC points is measured against it on VERSE2019."},{"cited_title":"In: First Conference on Language Modeling","cited_arxiv_id":null,"evidence_quote":"Is the selective state-space model that the VSSM modules inside MG-SSM are built on."},{"cited_title":"T., Warrington, A., Linderman, S","cited_arxiv_id":null,"evidence_quote":"Supplies the earlier simplified state-space layer design motivating the linear-complexity sequence modeling component."}],"review_version":1}