{"id":"5bf4dcae-fab5-4518-a3a0-48d45fa6f23b","arxiv_id":"2505.06210","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Topo-VM-UNetV2 encodes persistence-based topology attention maps into VM-UNetV2's SDI module and improves polyp segmentation Dice by 1.2 to 3.3 points on five public datasets.","lead":"This paper adds a topology-aware attention module to a Mamba-based polyp segmentation network and reports moderate accuracy gains on five public datasets. It tests whether persistence-based attention maps, computed from a network's own probability outputs, can improve boundary and structure quality in medical image segmentation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sublevel-filtration birth mapping may highlight background minima rather than polyps, so the reported gains may not come from topology; a superlevel or exact-birth re-run would settle it.","rationale":"The paper's central empirical claim is that persistence-derived topology attention maps improve a Mamba U-Net segmentation model. For that claim to be true, the attention maps must actually encode the births of relevant topological features of the probability maps. That condition is least secure at the birth-location mapping: the sublevel filtration direction means significant 0D classes are background basins, and the 'very close' tolerance is undefined. This is not a stylistic complaint; it determines whether the method's mechanism can operate as stated. I agree with the reader's conditional verdict: code, error bars, and the sensitivity checks above are needed. I would not reject the paper: the ablation versus PM-SDI is a useful control, the Gudhi-based runtime is reported, and five-dataset consistency is some evidence. But the empirical gain is not yet attributable to topology.","tokens_in":9661,"tokens_out":7093,"duration_ms":75552,"concrete_test":"Re-run the reported Stage 2 protocol on ETIS and ColonDB with three attention variants: (1) exact birth-cell assignment (no 'very close' tolerance) for the current sublevel filtration; (2) the same pipeline on a superlevel filtration of the probability maps, so 0D births occur at high-probability polyp maxima; (3) a non-topological control such as a Canny edge map with the same spatial support. If variants 1-2 do not match or beat the paper's tolerance-based Topo-SDI, or if Canny matches it, the claimed topological mechanism is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption is correct, and it can be sharpened. Section II-A2 Step 1 explicitly builds a sublevel filtration: pixels are activated when p_ij <= t_n. Under a sublevel filtration of a grayscale image, 0-dimensional homology classes are born at local minima of the probability map, i.e., within low-probability background regions, not at the high-probability polyp regions that the segmentation task cares about. The birth-location mapping in Step 3 assigns persistence to 'cells (pixels) whose filtration values closely match the feature's birth value,' but the tolerance is never defined, and for a 0D class the actual birth cell is a single local-minimum pixel (or a plateau with a tie-breaking rule), not a band of close-valued pixels. The 50th-percentile significance filter is also parameter-free and untested. If the filtration direction or the birth-location heuristic is wrong, the topology attention maps may simply be a boundary/background signal, and the consistent 1.0-3.3 DSC gains over VM-UNetV2 could come from the extra test-time probability-map conditioning in the SDI module rather than from genuine topological encoding. The ablation against PM-SDI partially controls for the extra PM input, but PM-SDI is a weak baseline (sigmoid of raw PM) and does not establish that persistence-derived attention is what drives the gains.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Topo-VM-UNetV2, a two-stage method for polyp segmentation that injects topology-aware attention into VM-UNetV2, a Mamba-based U-Net. In Stage 1, a trained VM-UNetV2 produces probability maps (PMs) for training and test images; these PMs are then processed via persistent homology to generate topology attention maps. The pipeline constructs a sublevel filtration of each PM, computes persistence diagrams for 0- and 1-dimensional features, filters features by persistence at the 50th percentile, assigns each surviving feature's persistence to its birth-location pixels, and applies a sigmoid to obtain attention weights in [0,1]. In Stage 2, these maps are multiplied with the multi-scale features inside the SDI module of VM-UNetV2, forming a Topo-SDI module. Experiments on Kvasir-SEG, ClinicDB, ColonDB, ETIS, and CVC-300 report consistent DSC, IoU, and MAE improvements over VM-UNetV2, with the largest gain on ETIS (3.29% DSC). An ablation compares Topo-SDI against a PM-SDI variant that uses sigmoid-transformed raw PMs as attention maps, showing larger gains for Topo-SDI. The authors state that each experiment is run five times with different seeds, but only averaged results are reported.","tokens_in":9999,"tokens_out":3311,"duration_ms":34427,"significance":"If the mechanism is correct, the work is a useful and inexpensive way to incorporate global topological structure into Mamba-based segmentation networks, and it addresses a recognized limitation of SSM architectures in capturing connected components and loops. The main strengths are the clear two-stage design, the evaluation across five public datasets with consistent gains, the inclusion of a PM-SDI ablation as a first-step control for the extra probability-map input, and the reporting of topology-attention computation cost. These elements make the empirical claim concrete and reproducible if code is released. The principal risk is that the topology attention maps may not actually encode the polyp's topology, because the birth-location heuristic in the sublevel filtration is conceptually questionable and empirically unvalidated; this concern must be resolved before the contribution can be fully credited.","major_comments":[{"comment":"The sublevel filtration defined in Step 1 activates pixels when p_ij <= t_n, so 0-dimensional homology classes are born at local minima of the probability map. In polyp segmentation, these minima are located in the background, not on the polyp. The birth-location mapping in Step 3 then assigns persistence values to 'cells whose filtration values closely match the feature's birth value,' but for a 0D class the birth cell is a single local-minimum pixel or a tied plateau, not a band of close-valued pixels, and the matching tolerance is never defined. Consequently, the topology attention maps may highlight background minima rather than polyp structures, which would mean the reported gains do not come from encoding polyp topology. Please validate this step by (i) overlaying birth locations on the PMs, (ii) re-running the pipeline with a superlevel filtration (activating p_ij >= t_n, where 0D classes are born at local maxima), or (iii) otherwise demonstrating that the birth-localization heuristic is correct for the claimed purpose.","section":"II-A2, Steps 1-3"},{"comment":"The persistence significance filter is fixed at the 50th percentile with no sensitivity analysis, and the PM-SDI ablation does not isolate the persistence geometry. PM-SDI uses a smooth sigmoid of the raw PM, whereas the topology maps, as the paper's own Fig. 5 states, emphasize boundaries and transitions; the observed Topo-SDI gains could therefore arise from a boundary-emphasis or high-frequency signal rather than from persistent homology. Please add at least one additional control whose spatial distribution matches the topology maps but whose values are not persistence-derived (e.g., an edge-filtered or randomly shuffled version of the same attention maps), and report results across a small grid of persistence thresholds (e.g., 40th, 50th, 60th percentiles) to show the 50th percentile choice is not load-bearing.","section":"II-A2, Step 3 and Table II"},{"comment":"The manuscript reports that each experiment is run five times with different seeds, but Table I and Table II list only point estimates with no standard deviations, confidence intervals, or significance tests. Given that the gains on Kvasir-SEG (1.20% DSC) and ClinicDB (1.76% DSC) are small relative to typical seed-to-seed variability in medical segmentation models, the claim of consistent improvement is not statistically supported as presented. Please report per-seed results or standard deviations, and, if possible, paired statistical tests between Topo-SDI and VM-UNetV2 on the same data splits.","section":"III-B and Table I"}],"minor_comments":[{"comment":"The comparison table omits PraNet, SANet, and Polyp-PVT with only a sentence explaining that U-Net v2 already outperforms them; please state this explicitly in the table caption or in the main text so readers do not interpret the table as a complete SOTA comparison.","section":"III-C / Table I"},{"comment":"The caption says 'As the threshold value increases or decreases,' but the sublevel filtration only uses an increasing threshold; please adjust the wording to avoid ambiguity.","section":"Fig. 3 caption"},{"comment":"The computational complexity section reports wall-clock time for 1450 images but not the per-image breakdown or the memory footprint; adding those details would make the cost claim easier to assess.","section":"III-E"},{"comment":"The terms 'sub-level' and 'sublevel' are used inconsistently; please use one spelling throughout.","section":"II-A2"},{"comment":"Reference [22] is formatted inconsistently with the other entries (missing conference/journal and year formatting); please check the bibliographic style.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The central empirical result (consistent gains over VM-UNetV2) is plausible, but the paper's claimed mechanism—that the gains come from encoding topological features—rests on the sublevel-filtration birth-location heuristic, which is conceptually suspect for polyp segmentation. If the superlevel or exact-birth re-run shows the attention maps are essentially background/boundary signals, the contribution reduces to an extra input channel with a hand-designed filter, and the title claim would need revision. I would like the editor to ensure the revision includes the proposed controls and preferably per-seed statistics before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent engineering paper with consistent DSC gains, but the topology story is probably wrong. The sublevel filtration maps birth locations to low-probability background minima, not to the polyp, so the attention maps may be a boundary/background signal. The gains over VM-UNetV2 are still real, but they may come from the extra probability-map conditioning rather than from topology. The paper deserves peer review because the empirical result is worth checking, but it needs major revision.\n\nWhat's new: the specific design of computing persistence diagrams from a trained model's probability maps, thresholding persistence at the 50th percentile, mapping persistence to birth pixels, and injecting the result into the SDI module of VM-UNetV2. The ablation against PM-SDI is a reasonable control, and the gains are consistent across five datasets. The computational cost is reported (187 seconds for 1450 images), which is useful.\n\nThe soft spots: first, the filtration direction. The paper explicitly constructs a sublevel filtration: pixels activate when p_ij <= t_n. Under that filtration, 0D homology classes are born at local minima of the probability map, which are typically background regions. The birth-location mapping then assigns high persistence to those background minima, not to the polyp. The paper's Figure 5 shows attention at boundaries, which is consistent with a boundary artifact rather than a genuine topological signal. This is not a minor concern; it undercuts the paper's central claim that topology is being encoded. The authors should re-run with a superlevel filtration (or at least justify why sublevel minima are the right birth locations), and they should define the 'closely match' tolerance.\n\nSecond, the evaluation is missing error bars. They ran five seeds but only report averages, with no standard deviations or significance tests. Some of the reported differences are small (e.g., PM-SDI vs Topo-SDI on ClinicDB is 0.79 DSC points), and without variance we cannot tell whether the topology part contributes meaningfully.\n\nThird, no code is released, and the persistence threshold (50th percentile) and sigmoid scaling are not sensitivity-tested. The claim of SOTA is also overstated given that they compare against only four baselines.\n\nWho this is for: anyone working on topology-aware segmentation will be interested, but should read the ablation table with care. It is a useful empirical data point, not a settled method.\n\nRecommendation: send to peer review, but require the authors to fix or reframe the filtration issue, add error bars/significance tests, release code, and run a sensitivity analysis. If the gains survive those changes, the paper becomes a solid application; if not, it is a cautionary tale about conflating extra inputs with topological understanding.","headline":"Consistent empirical gains, but the topology story is likely wrong; needs revision before it should be trusted.","tokens_in":10485,"tokens_out":3475,"would_cite":false,"duration_ms":34159,"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":"Persistence-based attention maps improve Mamba U-Net polyp segmentation on five public datasets.","keywords":["polyp segmentation","persistent homology","topology attention","Mamba","state space models","VM-UNetV2","medical image segmentation"],"falsifier":"Recompute the topology attention maps while sweeping the 'very close' tolerance and the 50th-percentile significance filter across a coarse grid; if the reported Dice gains vanish or the maps reduce to generic edge detectors on synthetic images with known component counts, the improvement does not stem from persistent homology.","tokens_in":9509,"feed_emoji":"🩺","tokens_out":2564,"duration_ms":27065,"temperature":0.7,"pith_summary":"The paper tries to show that encoding topological features into a Mamba U-Net improves polyp segmentation beyond what the base architecture achieves. It proposes a two-stage method: first train VM-UNetV2 to produce probability maps, convert those maps into topology attention maps via persistent homology, then re-train the network with those attention maps integrated into its feature-fusion module. The claim is that this yields state-of-the-art Dice, IoU, and MAE scores on all five datasets, with up to a 3.29% DSC improvement over VM-UNetV2 on ETIS. If correct, it shows that persistence-based topology is a useful, cheaply computable prior for medical image segmentation.","feed_headline":"Topology maps lift polyp segmentation accuracy","feed_subtitle":"Persistence-based attention in a Mamba U-Net sharpens boundaries and cuts false positives, gaining up to 3.29% Dice.","key_machinery":"The central mechanism is the Topo-SDI module: a per-scale Hadamard product between the encoder's features (after CBAM attention and a 1x1 convolution) and a resized topology attention map, followed by the same resolution-aligned multi-scale fusion as the original SDI. The attention map itself is built from a persistence diagram: significant 0- and 1-dimensional features of a probability map are selected by a 50th-percentile filter, each feature's persistence value is placed at the pixel(s) whose filtration value is 'very close' to the feature's birth value, and the resulting score map is sigmoid-activated into [0,1].","core_discovery":"Topo-VM-UNetV2 claims that replacing the plain SDI module in VM-UNetV2 with a topology-guided version that multiplies each multi-scale feature map by a persistence-derived attention map consistently improves polyp segmentation. The attention maps are generated by computing persistence diagrams of the model's probability maps under a sublevel filtration, filtering to the top 50% by persistence, assigning each surviving feature's persistence value to its birth location, and passing those scores through a sigmoid. The paper reports that this addition outperforms VM-UNetV2 on all five datasets and also beats a control in which raw probability maps, rather than topology-derived maps, are used for the same attention mechanism.","pith_inferences":["The same persistence-to-attention recipe might transfer to other medical imaging tasks like tumor or organ segmentation, where boundary accuracy and the number of connected components are clinically meaningful.","A natural testable extension is to sweep the 'very close' tolerance and the 50th-percentile filter per dataset; if the improvements are stable across a wide range, the method is robust, and if they collapse, the reported gains may be implementation-specific.","The attention mechanism effectively acts as a boundary-and-structure prior, so it could be combined with other priors such as shape constraints or anatomical atlases to push further.","One implicit consequence is that the quality of the Stage-1 probability maps bounds the quality of the topology; retraining Stage 1 with better segmenters might yield even larger gains."],"forward_implications":["The reported gains suggest that a lightweight topological prior computed once per image can sharpen boundaries and reduce false positives without changing the backbone or the inference-time architecture.","The same Topo-SDI idea could be grafted onto other U-Net-style segmenters, since the attention maps are model-agnostic once probability maps exist.","Because the topology maps are computed offline in Stage 1, adding them does not increase the model's parameter count or inference cost during deployment.","The method's behavior on hard datasets such as ETIS and ColonDB indicates the largest payoff appears exactly where the base model struggles most with under- and over-segmentation.","The ablation against PM-SDI, where raw probability maps replace the topology maps, indicates that the improvement comes from persistence information rather than from simply adding any auxiliary spatial signal."],"supporting_citations":[{"why":"Provides VM-UNetV2, the base architecture whose probability maps and SDI module are modified; all comparisons are against this model.","marker":"[10]"},{"why":"Supplies the SDI fusion module that Topo-SDI augments, and the U-Net v2 baseline that established the multi-scale fusion design.","marker":"[17]"},{"why":"Defines Mamba, the state space model underlying the backbone's linear-complexity sequence modeling.","marker":"[18]"},{"why":"Defines VMamba and the SS2D cross-scan mechanism that makes Mamba suitable for 2D image features.","marker":"[19]"},{"why":"Provides the computational-topology foundation for persistence diagrams and filtrations used to build the attention maps.","marker":"[21]"},{"why":"Supplies the modern computational-topology treatment of persistence, referenced for the theoretical details of the filtration and diagram construction.","marker":"[22]"},{"why":"Supplies CBAM, the channel-spatial attention applied before the topology maps are multiplied into the features.","marker":"[23]"}],"fun_headline_variants":["Topology boosts polyp segmentation in Mamba U-Net","Persistence-derived attention sharpens polyp boundaries","Topo-VM-UNetV2: topology-aware Mamba for polyps","Mamba U-Net gains 3.29% Dice via topology attention","Topology attention fixes polyp boundaries in Mamba U-Net"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the pixels whose probability values are very close to a topological feature's birth value are exactly where that feature is born, yet that tolerance is never defined and a sublevel filtration actually births a 0-dimensional component at a single local minimum pixel, not across a plateau.","fun_headline_variants_meta":{"raw":{"variants":["Topology boosts polyp segmentation in Mamba U-Net","Persistence-derived attention sharpens polyp boundaries","Topo-VM-UNetV2: topology-aware Mamba for polyps","Mamba U-Net gains 3.29% Dice via topology attention","Topology attention fixes polyp boundaries in Mamba U-Net"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1298,"prompt_tokens":997,"completion_tokens":301,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":212}},"tokens_in":613,"tokens_out":301,"duration_ms":3145,"temperature":1.0,"reasoning_tokens":212,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:46:17.844449+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the topology attention maps while sweeping the 'very close' tolerance and the 50th-percentile significance filter across a coarse grid; if the reported Dice gains vanish or the maps reduce to generic edge detectors on synthetic images with known component counts, the improvement does not stem from persistent homology.","supporting_citations":[{"cited_title":"VM-UNetV2: Rethinking vision Mamba U-Net for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides VM-UNetV2, the base architecture whose probability maps and SDI module are modified; all comparisons are against this model."},{"cited_title":"VMamba: Visual state space model,","cited_arxiv_id":null,"evidence_quote":"Defines VMamba and the SS2D cross-scan mechanism that makes Mamba suitable for 2D image features."},{"cited_title":"Edelsbrunner and J","cited_arxiv_id":null,"evidence_quote":"Provides the computational-topology foundation for persistence diagrams and filtrations used to build the attention maps."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the modern computational-topology treatment of persistence, referenced for the theoretical details of the filtration and diagram construction."}],"review_version":1}