{"id":"afe5ef78-0bf1-4cee-b304-e489bd842b9a","arxiv_id":"2507.01323","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A new Mamba-based vascular segmentation network with serpentine window tokenization reports the best average scores on three public datasets, though without error bars or code.","lead":"SWinMamba segments blood vessels by scanning overlapping serpentine windows with a bidirectional Mamba state space model and fusing spatial and frequency features. If the reported gains hold, it offers a low-parameter option for continuity-aware vascular segmentation in medical images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported superiority claim rests on four-fold single-run point estimates; the 0.14% mean Dice gain is within expected fold noise, and Table 1 itself shows SWinMamba is not best on CHASE β0 (1.546 vs DSCNet's 1.434) or DCA1 Dice (0.7659 vs DE-DCGCN-EE's 0.7660).","rationale":"In good faith, SWinMamba is a plausible incremental architecture: SWToken adaptively samples overlapping windows, BAM imposes bidirectional sequential scanning, and SFFU adds a frequency-domain branch. The ablations in Table 2 are internally consistent, with the full model generally best. The central claim, however, is empirical: SWinMamba 'outperforms SOTA models on most metrics with low computational cost' (Section 3.5). For that claim to hold, the reported margins must survive variance in training and evaluation. The paper provides no variance information: the protocol is a single four-fold cross-validation, Table 1 contains only point estimates, and the training loss is never specified. The reported margins are very small, especially 0.39% clDice and 0.14% Dice, which are far below the fold-to-fold noise typical on datasets with 28, 300, and 134 samples. The reader's weakest_assumption identifies exactly this issue, and my reading agrees. I add a concrete internal check: Table 1 has at least two cells where SWinMamba is not actually best—CHASE β0 (DSCNet 1.434 vs SWinMamba 1.546, lower is better) and DCA1 Dice (DE-DCGCN-EE 0.7660 vs SWinMamba 0.7659). Thus the paper's own numbers already undermine the universal 'best results' caption. This is not an accusation of misconduct; it is a verification failure. A rerun with seeds, confidence intervals, and a stated loss could confirm the claim, which is why the appropriate disposition remains CONDITIONAL rather than REJECT. Since the reader already reached CONDITIONAL, my verdict is UNCHANGED, but I would emphasize the Table 1 bookkeeping error and the loss-function omission as concrete items to fix.","tokens_in":8315,"tokens_out":8659,"duration_ms":92199,"concrete_test":"Run every method in Table 1 with at least 5 independent seeds on the same four-fold splits, state the loss function in Section 3.2, and report mean ± std for clDice, β0, and Dice per dataset. Then compute paired bootstrap 95% confidence intervals (or Wilcoxon signed-rank tests over folds×seeds) between SWinMamba and each baseline. As a separate bookkeeping check, re-verify Table 1's bolded cells: DSCNet's CHASE β0 (1.434) and DE-DCGCN-EE's DCA1 Dice (0.7660) are better than SWinMamba's values (1.546 and 0.7659). If any per-dataset confidence interval overlaps the best baseline, or the bolding is corrected to remove those wins, the 3.15%/0.39%/0.14% average-improvement claim should be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that SWinMamba's Table 1 margins are real and repeatable. That premise is not supported by the evidence as reported. Section 3.2 specifies augmentation, optimizer, epochs and batch size but never states the training loss; Section 3.5 reports only point estimates from a single four-fold split, with no seed variance, confidence intervals, or significance tests. The differences in question are small: average gains are 0.39% in clDice and 0.14% in Dice. On CHASE-DB1, with 28 images split into four folds (roughly 7 test images per fold), a per-image Dice standard deviation of even 3 points gives a per-fold standard error above 1 point, an order of magnitude larger than the headline Dice gain. The internal table also contradicts the caption's 'best results': CHASE β0 is 1.546 for SWinMamba versus 1.434 for DSCNet (lower is better), and DCA1 Dice is 0.7659 versus 0.7660 for DE-DCGCN-EE. The aggregate 3.15% β0 improvement therefore hides a per-dataset regression on the key connectivity metric. Since clDice is commonly used as a loss function and Section 3.2 omits the loss, it is impossible to tell whether reported gains reflect the architecture or a different training objective. No code is supplied, so the point estimates cannot be independently checked. Without repeated-seed statistics and a stated loss function, the central superiority claim is not established; a rerun could easily place SWinMamba within noise of several baselines.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SWinMamba, a U-shaped encoder-decoder for vascular segmentation with three main components: a Serpentine Window Tokenizer (SWToken) that constructs learnable overlapping window strings along vascular structures, a Bidirectional Aggregation Module (BAM) that feeds these token sequences into a bidirectional state space model, and a Spatial-Frequency Fusion Unit (SFFU) that combines spatial- and frequency-domain features. The method is evaluated on CHASE-DB1, OCTA-500, and DCA1 against ten published baselines using clDice, Betti number error β0, Dice, FLOPs, and parameter counts. The central claim is that SWinMamba outperforms state-of-the-art methods on most metrics with low computational cost, achieving average improvements of 3.15% in β0, 0.39% in clDice, and 0.14% in Dice across the three datasets (Section 3.5, Table 1).","tokens_in":8715,"tokens_out":11257,"duration_ms":119484,"significance":"The architectural idea is timely and the evaluation scope is a strength: three datasets of different modalities, ten baselines spanning convolutional, hybrid, and Mamba-based families, and the stated use of official codes. If the reported superiority is robust, SWinMamba would be a useful low-cost SSM-based alternative for tubular structure segmentation. However, the significance is currently conditional because the main evidence consists of single-run point estimates: the reported Dice margins are far smaller than the expected fold-to-fold variability on a dataset as small as CHASE-DB1, and Table 1 contains cells in which SWinMamba is not the best despite the caption claiming the best results. The submission does not provide code, trained models, or per-fold results, so the empirical claim is not independently verifiable as presented.","major_comments":[{"comment":"The central superiority claim is not supported by the reported statistics. Four-fold cross-validation on CHASE-DB1 (28 images) leaves roughly 7 test images per fold; with a per-image Dice standard deviation of even 3 points, the per-fold standard error is above 1 point, which is roughly an order of magnitude larger than the headline 0.14% average Dice gain. Report repeated-seed results as mean±std, confidence intervals, and paired significance tests across folds, and provide per-fold breakdowns for all methods. The same limitation applies to the ablation study in Table 2, where component contributions are single-point estimates and the word 'significant' is used informally.","section":"Section 3.5, Table 1"},{"comment":"Table 1 is internally inconsistent with its caption stating that SWinMamba achieves the best results. On CHASE-DB1, the β0 of SWinMamba (1.546, lower is better) is worse than DSCNet's 1.434, and on DCA1 the Dice of SWinMamba (0.7659) is worse than DE-DCGCN-EE's 0.7660. Since β0 is the connectivity metric that the paper's premise targets, the aggregate '3.15% improvement in β0' conceals a per-dataset regression on CHASE-DB1; the average is largely carried by OCTA-500, where SWinMamba's 2.636 is well ahead of the next best 3.042. The bolding and the claim of consistent connected-vessel improvement should be corrected.","section":"Section 3.3, Section 3.5, Table 1"},{"comment":"The training loss is never stated. Because clDice is both an evaluation metric and a commonly used training objective for tubular-structure segmentation, omitting the loss makes it impossible to determine whether the reported gains come from the architecture or from a different training objective. State the loss function, any weighting or auxiliary losses, and, if possible, retrain all baselines with the same objective for the main comparison.","section":"Section 3.2"},{"comment":"The comparison protocol is underspecified. The text says all baselines are implemented using official codes, but it does not state whether they were retrained under the same four-fold split, augmentation, random-crop size, optimizer schedule, and epoch count, or whether Table 1 reuses published numbers. It also does not describe the test-time inference procedure, which is especially important because training uses 256×256 random crops while input images are as large as 999×960. These choices can easily account for differences of the magnitude reported here, so the full protocol must be documented and per-fold results or code should be supplied.","section":"Section 3.2, Section 3.5"},{"comment":"The core tokenizer depends on three hand-set parameters (L=9, s=8, α=2), described only as 'set empirically.' No sensitivity analysis is provided, so it is unclear whether SWinMamba's performance is stable in the parameter region or tuned specifically to these datasets. Add an ablation over L, s, and α, or provide a robustness argument.","section":"Section 2.1"}],"minor_comments":[{"comment":"Define precisely how the Betti number error β0 is computed, including whether it is a per-image mean, whether it is normalized, and how connected components are counted. This metric is central to the connectivity claim.","section":"Section 3.2"},{"comment":"Clarify the notation in Eq. (2): specify the range and construction of the predicted offset field Δy_k, and explain how coordinates that fall outside the feature map are handled during bilinear sampling.","section":"Section 2.1, Eq. (2)"},{"comment":"Explain how the aggregate 'average improvements' are computed across datasets and metrics, especially for β0 where lower is better and the numerical scale differs between datasets.","section":"Section 3.5"},{"comment":"The submission does not provide code, trained models, or an explicit reproducibility statement. For an empirical point-estimate comparison, releasing these resources would substantially strengthen the paper.","section":"General"},{"comment":"Figure 1 contains several short labels and fragments (for example, the text boxes under the challenge illustrations) that are difficult to read in the submitted version; the figure should be redrawn for legibility.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is an empirical benchmark result, and the requested revision is not cosmetic: it requires repeated-seed experiments, a full statement of the training and evaluation protocol, and correction of the internal inconsistencies in Table 1. The architectural idea is plausible and the OCTA-500 β0 gap is large enough to be credible, so I would not reject on the current evidence. I would require the missing statistics before publication, and I suggest asking the authors to provide per-fold and per-seed results as part of the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: SWinMamba is a genuine architectural contribution, but its headline numerical claim is not supported by the evidence as reported. The serpentine window tokenizer with learned offsets plus bidirectional Mamba over window sequences is new, and comparing on three public benchmarks is the right kind of experiment. But the paper undercuts itself by reporting single point estimates with no variance, omitting the training loss, and bolding numbers that are not actually best.\n\nWhat is good: the idea is coherent. SWToken adaptively samples overlapping windows along predicted vessel-aligned paths, BAM scans them bidirectionally, and SFFU mixes spatial and frequency features. The ablation is structured and shows each component helps, with SFFU recovering the decline from adding FFT alone. The parameter/FLOP count is competitive: 2.78M parameters and 8.20 GFLOPs, lower than most baselines. If the connectivity gains are real, this is a useful option for vascular segmentation.\n\nWhere it is soft: the central claim is the average improvement of 3.15% in beta0, 0.39% in clDice, and 0.14% in Dice. On CHASE-DB1 with 28 images and four-fold CV, a per-image Dice standard deviation of 3 points gives a per-fold standard error above 1 point, an order of magnitude larger than the headline Dice gain. That makes 0.14% meaningless without repeated-seed statistics. Table 1 itself shows SWinMamba is not best on CHASE beta0 (1.546 vs DSCNet's 1.434) or on DCA1 Dice (0.7659 vs DE-DCGCN-EE's 0.7660). The aggregate beta0 improvement hides a regression on the key connectivity metric on one dataset. More importantly, the loss function is never stated; clDice is commonly used as a loss, so without knowing it you cannot attribute gains to the architecture. No code is provided, so the point estimates cannot be independently checked. These are fixable: report error bars over seeds, state the loss, correct the bolding, release code.\n\nWho this is for: researchers working on Mamba-based or tubular-structure segmentation. The architecture is worth knowing about, but the empirical superiority should not be taken at face value. A serious referee could push for the missing statistics and code, and with those the paper could be solid.\n\nRecommendation: send it to peer review rather than desk reject, but expect heavy revision on experimental reporting. If the authors cannot show variance or state the loss, the superiority claim stays unproven.","headline":"Real architecture, unproven numeric claim: the serpentine-window Mamba idea is worth a look, but the reported gains are point estimates without variance or a stated loss on tiny datasets.","tokens_in":9238,"tokens_out":2290,"would_cite":false,"duration_ms":25601,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A serpentine window ordering through a bidirectional state space model keeps slender vessels connected, beating ten baselines on three medical datasets while using only 2.78M parameters and 8.20 GFLOPs.","keywords":["vascular segmentation","state space model","Mamba","serpentine window tokenizer","bidirectional aggregation","spatial-frequency fusion","medical image segmentation"],"falsifier":"Retrain SWinMamba and the strongest baseline from each family with at least ten random seeds under the identical augmentation, optimizer, and schedule, and compare the distributions of $\\beta_0$, clDice, and Dice; the claimed $3.15\\%$, $0.39\\%$, and $0.14\\%$ improvements would be falsified if the gaps fall within one standard deviation of the paired differences.","tokens_in":8096,"feed_emoji":"🩺","tokens_out":8006,"duration_ms":68881,"temperature":0.7,"pith_summary":"Vascular segmentation in medical images fails when slender vessels are split into disconnected pieces, and this paper claims that the cause is how the image is turned into tokens. The proposed SWinMamba network orders overlapping local windows into serpentine sequences that follow the vessel direction, then feeds them through a bidirectional state space model so the segmenter sees a vessel as one continuous tubular object. On three datasets --- retinal fundus, OCT angiography, and X-ray coronary angiography --- the authors report that SWinMamba beats ten published methods on most metrics, including the largest average gain of $3.15\\%$ on $\\beta_0$, the connectivity error, while using only $2.78$M parameters and $8.20$ GFLOPs. The paper's contribution, read sympathetically, is a tokenization-and-scanning recipe that replaces rigid raster scans with anatomy-aligned sequences, making state space models usable for continuity-critical segmentation.","feed_headline":"Vessel breakage drops 3.15% with serpentine Mamba windows","feed_subtitle":"A 2.78M-parameter model beats ten baselines on retinal, OCT, and X-ray datasets using roughly 8 GFLOPs.","key_machinery":"The load-bearing machinery is the Serpentine Window Tokenizer (SWToken) combined with the Bidirectional Aggregation Module (BAM). SWToken computes strings of anchor points whose lateral offsets are predicted by a convolution, so the $s\\times s$ windows overlap and bend along the vessel direction instead of sitting on a fixed grid; fractional sampling is carried out by bilinear interpolation. BAM arranges the tokens of each window into a sequence, runs a forward and a backward state space scan across windows, and reassembles the processed patches into a feature map. A third component, the Spatial-Frequency Fusion Unit (SFFU), runs the same windowed processing in the Fourier domain and fuses it with the spatial branch under attention weights. Together, these convert slender, meandering vessels into ordered local contexts that a state space model can reason about sequentially.","core_discovery":"The paper's central claim is that continuity, not just pixel accuracy, is what a vessel segmenter should be structured around, and that a state space model can deliver it if the scanning path follows the vessel. SWinMamba is said to produce complete and connected vascular segmentation by replacing the rigid 1D flattening of Vision Mamba with overlapping serpentine window sequences: anchor points are shifted by learnable offsets so local windows snake along vessel contours, and a bidirectional aggregation module scans those windows forward and backward to integrate coherent local features. A Spatial-Frequency Fusion Unit then fuses spatial-domain and frequency-domain features under attention weights. With $2.78$M parameters and $8.20$ GFLOPs, the reported results on CHASE-DB1, OCTA-500, and DCA1 beat all ten compared methods on most metrics, with average improvements of $3.15\\%$ in $\\beta_0$, $0.39\\%$ in clDice, and $0.14\\%$ in Dice.","pith_inferences":["A natural extension is 3D: the same anchor-string construction could snake through CT or MR angiography volumes, where vessels are curves in three dimensions and connectivity errors are even harder to measure.","The frequency-domain branch suggests that phase information carries continuity cues; one testable ablation would remove only the phase of the Fourier features and watch what happens to the $\\beta_0$ gain.","Because the training loss and seed variance are not reported, the most direct test of the contribution is to retrain the best baseline and SWinMamba under several seeds with the identical loss and schedule, turning the claimed averages into distributions."],"forward_implications":["If SWinMamba is right, connectivity errors in vessel segmentation can be reduced by aligning sequence order with anatomy rather than by increasing model size.","The network outputs connected vessels directly, so downstream analysis of diameter, curvature, and bifurcation angles would not need a separate repair step.","At $2.78$M parameters and $8.20$ GFLOPs, the approach is light enough for clinical workstations, unlike transformer baselines that require hundreds of GFLOPs.","Since the largest reported gain is on $\\beta_0$, the design targets the topological error that matters most for quantifying vascular geometry."],"supporting_citations":[{"why":"Supplies the selective state space model that BAM builds on for bidirectional scanning.","marker":"[8]"},{"why":"Provides the Vision Mamba block that serves as the ablation baseline and comparison method.","marker":"[24]"},{"why":"Defines clDice, the topology-aware metric used to measure vessel completeness.","marker":"[20]"},{"why":"Supplies the CHASE-DB1 retinal fundus dataset used in evaluation.","marker":"[6]"},{"why":"Supplies the OCTA-500 OCT angiography dataset used in evaluation.","marker":"[11]"},{"why":"Supplies the DCA1 X-ray coronary angiography dataset used in evaluation.","marker":"[1]"},{"why":"Acts as the UNet baseline that the comparison must beat.","marker":"[17]"},{"why":"Acts as the DSCNet baseline, a snake-convolution method for tubular structures.","marker":"[16]"},{"why":"Acts as the VM-UNet baseline, a Mamba-based segmentation network.","marker":"[18]"}],"fun_headline_variants":["Serpentine Mamba scan keeps vessel segments connected","Vessel breakage down 3.15% via adaptive serpentine windows","SWinMamba weaves continuous vessels with state space model","Bidirectional serpentine windows stitch together vessel maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported gains over ten baselines are larger than run-to-run noise, since the evidence is four-fold cross-validation on small datasets with single point estimates, no variance or significance tests, and no stated training loss.","fun_headline_variants_meta":{"raw":{"variants":["Serpentine Mamba scan keeps vessel segments connected","Vessel breakage down 3.15% via adaptive serpentine windows","SWinMamba weaves continuous vessels with state space model","Bidirectional serpentine windows stitch together vessel maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000566,"raw_usage":{"total_tokens":2690,"prompt_tokens":959,"completion_tokens":1731,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":1662}},"tokens_in":575,"tokens_out":1731,"duration_ms":89490,"temperature":1.0,"reasoning_tokens":1662,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:55:39.190722+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain SWinMamba and the strongest baseline from each family with at least ten random seeds under the identical augmentation, optimizer, and schedule, and compare the distributions of $\\beta_0$, clDice, and Dice; the claimed $3.15\\%$, $0.39\\%$, and $0.14\\%$ improvements would be falsified if the gaps fall within one standard deviation of the paired differences.","supporting_citations":[{"cited_title":"In: Proceedings of the International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Provides the Vision Mamba block that serves as the ablation baseline and comparison method."},{"cited_title":"IEEE Transactions on Biomedical Engineering59(9), 2538–2548 (2012)","cited_arxiv_id":null,"evidence_quote":"Supplies the CHASE-DB1 retinal fundus dataset used in evaluation."},{"cited_title":"Applied Sciences 9(24), 5507 (2019)","cited_arxiv_id":null,"evidence_quote":"Supplies the DCA1 X-ray coronary angiography dataset used in evaluation."}],"review_version":1}