{"id":"76370763-90e0-40dd-b606-9b46930062ef","arxiv_id":"2508.00442","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A two-stage test-time adaptation framework, TopoTTA, uses topology-aware difference convolutions and pseudo-break consistency training to improve tubular structure segmentation under domain shift.","lead":"This paper introduces TopoTTA, a test-time adaptation method that helps tubular structure segmentation models (blood vessels, roads, neurons) perform better when applied to new datasets with different image properties. It combines topology-aware convolutions with consistency training on generated pseudo-break images and reports large average gains in topological continuity metrics across ten datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage 2's topological gains rest on unvalidated pseudo-label accuracy: Eq. 9-10 reinforce the teacher's confident foreground pixels, and no experiment measures whether those pseudo-labels are correct in the selected pseudo-break regions.","rationale":"Reader's conditional verdict is appropriate. I agree with the reader that pseudo-label reliability in Stage 2 is the weakest load-bearing assumption. The paper's own baseline (CoTTA) also uses a teacher-student scheme, so the comparison does not isolate whether TopoHG's gain is due to better topology handling or to a more aggressive self-training schedule. The proposed concrete test directly measures the accuracy of the teacher in the exact regions that receive high-weight supervision. If pseudo-labels are accurate in those regions, the mechanism is sound; if not, the claimed topological continuity improvement may be an artifact of reinforcing confident errors. I do not see an internally inconsistent derivation or a clear violation of the TTA protocol; the remaining issues (table formatting, baseline tuning disclosure, variance) are secondary and already covered by the conditional recommendation. Therefore the reader's verdict should remain unchanged.","tokens_in":24444,"tokens_out":13822,"duration_ms":138657,"concrete_test":"On a held-out target subset with ground truth (e.g., CHASE in DRIVE to CHASE, Neub2 in Neub1 to Neub2), instrument Stage 2 to log every selected keypoint and its foreground window P_i^fg. Compute precision and F1 of the teacher pseudo-label Y_i^t against ground truth inside the weighted region P_i^* and Y_i^t. Stratify the per-image Stage-2-minus-Stage-1 clDice and Betti change by this precision. If gains appear only where precision is high, the concern lands; if gains persist when precision is low, the method does not depend on pseudo-label correctness. As a stronger diagnostic, rerun Eq. 9 with oracle ground-truth labels in the weighted region only; if gains improve materially, the reported results are bounded by pseudo-label quality, and the paper should report this bound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Stage 2's entire contribution is supervised by teacher pseudo-labels: Eq. 9 applies cross-entropy between the student's hard-sample prediction and Y_i^t, with Eq. 10 putting weight gamma exactly on pseudo-label-foreground pixels inside the pseudo-break patches. The only selection safeguard is the confidence threshold tau=0.5 (Sec. 3.4, Step 1). Under domain shift, confidence and accuracy decouple; overconfident false positives are precisely the pixels that satisfy tau and will be reinforced. The paper reports no measurement of pseudo-label precision on the selected keypoints or windows, so the clDice/Betti gains attributed to TopoHG (Table 3: clDice 62.20 to 66.61) could reflect fitting the teacher's, possibly wrong, topology rather than correcting real breaks. This is load-bearing because the topological continuity refinement claim depends on the correctness of these self-generated targets. A secondary issue is that the headline 31.81% clDice improvement is an absolute gain over Source Only (Table 2: 42.19 to 74.00), not over the best prior TTA (CoTTA: 69.34), which overstates the SOTA margin, though the average comparison to CoTTA remains positive.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TopoTTA, a two-stage test-time adaptation method for tubular structure segmentation. Stage 1 defines eight Topological Meta Difference Convolutions built from the pre-trained vanilla kernels and learns per-patch router weights (about 1,280 new parameters) via entropy minimization, leaving the pre-trained weights untouched. Stage 2 generates hard samples with local pseudo-breaks by swapping low-frequency components between foreground and nearby background windows selected from high-confidence teacher predictions, and applies teacher-student consistency (cross-entropy with a spatial weight map) to refine topological continuity. The method is evaluated on ten cross-domain source-target pairs spanning retinal vessel, road, neuronal, and OCT-A vessel segmentation, using UNet, CS2Net, and a DSCNet variant, against seven prior TTA methods. The authors report consistent gains in Dice, clDice, and Betti errors, and an average clDice improvement of 31.81%.","tokens_in":24709,"tokens_out":10164,"duration_ms":99861,"significance":"Test-time adaptation specifically for tubular structures is an underexplored and practically relevant problem, and the paper has several concrete strengths: the TopoMDC design is parameter-efficient, the TopoHG hard-sample generation is original, the evaluation covers more datasets and backbones than is typical in the TTA segmentation literature, the ablations isolate the two stages, and the appendix includes paired t-tests and synthesis-quality analysis. I find no circularity: the target labels come from external benchmarks, no constants are fitted to those labels, and the pseudo-label supervision is a standard TTA mechanism. If the claimed results hold after the revisions below, the method would be a useful plug-and-play contribution to tubular structure segmentation under domain shift.","major_comments":[{"comment":"Stage 2 is supervised entirely by teacher pseudo-labels, yet the paper provides no measurement of pseudo-label correctness in the selected keypoints or pseudo-break regions. The confidence threshold tau = 0.5 does not guarantee accuracy under domain shift, and overconfident false-positive pixels can be selected and then reinforced by the cross-entropy term weighted by gamma in Eq. (10). This is load-bearing for the TopoHG contribution (Table 3: clDice increases from 62.20 to 66.61). I request a direct analysis: report pseudo-label precision/recall or agreement with ground truth at the selected keypoints and pseudo-break patches on the target-domain test sets, or provide an oracle-quality ablation (for example, comparing against a variant that uses ground-truth masks as pseudo-labels) to separate the effect of reliable supervision from the effect of the TopoHG operation. Without this, the claim that TopoHG improves topological continuity by correcting real breaks remains a plausible interpretation rather than an established one.","section":"Sec. 3.4, Eq. (9)-(10)"},{"comment":"The advertised 'average improvement of 31.81% in clDice' is the absolute difference between Source Only and TopoTTA for the UNet baseline (Table 2: 42.19 to 74.00), not an average over prior TTA methods and not an average over both backbones. Against the strongest prior method (CoTTA), the average clDice gains are 4.66 points for UNet and 2.86 points for CS2Net; the average absolute gain over Source Only across both backbones is approximately 26.7 points. The claim as stated overstates the state-of-the-art margin and should be rewritten in percentage points with the reference baseline clearly identified.","section":"Abstract; Table 2"},{"comment":"The main results table is not numerically legible in the version under review. In the DRIVE to STARE block, the Source Only row reads '48.05 / 106.30', and several cells (105.20, 106.00, 298.74) are hard to assign to Dice, clDice, or Betti-error columns; the caption explains the '/' marker, but the column alignment makes it impossible to verify which values were averaged. This matters because Table 2 reports averages over ten datasets while at least three Source Only clDice entries are marked '/'; the paper never states how missing clDice values are handled in the average (available-case mean, imputation, or exclusion). Please regenerate Table 1 with unambiguous column separators, state the missing-value policy for computing averages, and recompute any headline numbers if the averaging rule changes the results.","section":"Table 1; Table 2"}],"minor_comments":[{"comment":"The sentence preceding Eq. (8) reads 'we update only update pixels'; it should be 'we update only pixels'.","section":"Eq. (8)"},{"comment":"Eq. (10) introduces a weight gamma and Step 1 introduces a keypoint-selection coefficient, but the hyperparameter list in Sec. 4.1 only reports s, theta_bg, and the mask threshold; please give the values used for gamma and the keypoint coefficient in the main text.","section":"Sec. 4.1; Eq. (10); Sec. 3.4 Step 1"},{"comment":"Many p-values in Table C.3 are rendered as unreadable placeholders ('������'); please replace them with actual numeric values so the statistical significance claim can be checked.","section":"Table C.3"},{"comment":"TopoHG is introduced as a three-step procedure, but Sec. 4.3 later refers to 'the above four steps'; please renumber consistently.","section":"Sec. 4.3"},{"comment":"The code availability link is an anonymous placeholder; a stable release link should be provided if the paper is accepted.","section":"Supplementary Material, Sec. B.1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely problem and the experimental scope is a genuine strength. The main risk is that Stage 2's pseudo-label supervision is not validated directly, and the headline improvement is framed against Source Only rather than prior TTA methods. Both are fixable with additional analysis and revised presentation. I see no evidence of circular evaluation or data contamination, but the main table must be regenerated before the numeric claims can be fully verified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"TopoTTA is the first TTA method aimed specifically at tubular structure segmentation, and the paper gives you a wide, consistent set of experiments for that claim: ten datasets across vessels, roads, neurons, and OCT-A, two baselines plus a DSCNet variant, and ablations that isolate each stage. I think the core contribution is real. TopoMDCs are a sensible extension of central difference convolution to directional dual-pixel differences, with only 1,280 router parameters per sample, and the router-reset-per-sample design is clean. TopoHG's low-frequency swap to create pseudo-breaks is also a nice idea, and the ablation against blur/noise/spatial swap shows it is doing something specific. The evaluation uses external labeled benchmarks, so circularity is not a concern. The paper deserves a serious referee.\n\nThe real soft spot is Stage 2's reliance on teacher pseudo-labels. Eq. 9-10 put consistency weight exactly on pseudo-label foreground pixels inside pseudo-break regions, selected by confidence tau=0.5. Under shift, confidence and accuracy decouple, and overconfident false positives are exactly the pixels that pass tau. The paper never measures pseudo-label precision in the selected keypoints/windows. That is a fair request, and an easy one: since target test labels exist in these benchmarks, they can report how often selected keypoints fall on real foreground, or vary tau and show the gain is robust. I don't think this sinks the paper, because the consistency mechanism helps even with imperfect targets, and the gains are consistent, but the mechanism as stated would be strengthened by that check.\n\nThe other issues are presentation and framing. The abstract's 31.81% clDice improvement is absolute over Source Only (42.19 to 74.00); against the best prior CoTTA the margin is 4.66 clDice on UNet. Still positive, but the headline overstates. Main tables have rendering artifacts (values like 106.30, stray slashes) and no error bars; the appendix has paired t-tests, which helps. Baseline tuning is described as \"assigning optimal ones\" with no per-method detail, which is standard but deserves disclosure. The DSCNet variant only runs Stage 2, which is fine but should be stated in the main text.\n\nWho is this for: anyone working on TTA for segmentation or topology-aware losses. It is a subfield contribution, not a field-shifter, but it is solid and the first in its niche. I would accept it for peer review and ask for the pseudo-label analysis, clean tables, and an explicit comparison against the strongest baseline rather than Source Only. My own verdict would be conditional accept, with code release as part of the condition.","headline":"First TTA method for tubular structures with broad, consistent experiments; the main open question is Stage 2 pseudo-label reliability, and the headline metric overstates the margin over the strongest baseline.","tokens_in":25274,"tokens_out":2896,"would_cite":true,"duration_ms":30346,"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":"TopoTTA proposes the first test-time adaptation framework tailored to tubular structure segmentation, reporting an average 31.81% clDice improvement across ten cross-domain datasets.","keywords":["test-time adaptation","tubular structure segmentation","topological continuity","clDice","central difference convolution","pseudo-break generation","domain shift","consistency regularization"],"falsifier":"Take a target domain whose tubular appearance is systematically different from anything in the source training data, so the teacher produces confident but wrong pseudo-labels in whole regions; run Stage 2 alone and compare Betti errors against Stage 1 alone. If Betti errors increase or the clDice gain disappears, the pseudo-label reliability premise is what failed.","tokens_in":24213,"feed_emoji":"🩺","tokens_out":8751,"duration_ms":81871,"temperature":0.7,"pith_summary":"TopoTTA argues that test-time adaptation for tubular structure segmentation should optimize topology itself, not just pixel accuracy. The paper proposes a two-stage pipeline in which Stage 1 rewires the encoder's convolutions into eight directional difference operators whose combination is reweighted per test sample, and Stage 2 manufactures local pseudo-breaks in confidently segmented regions and trains the model to restore continuity against teacher pseudo-labels. Across retinal, road, microscopic-neuron, and OCT-angiography datasets, TopoTTA reports an average clDice gain of 31.81% over seven prior test-time adaptation methods, with higher Dice in most settings. If the claim holds, topology-aware self-supervision becomes a practical recipe for online adaptation of thin-structure segmenters without source data or label access.","feed_headline":"TopoTTA lifts tubular segmentation by 31.81% clDice","feed_subtitle":"Two-stage adaptation repairs broken vessels, roads, and neurons with directional routing and pseudo-break consistency.","key_machinery":"The load-bearing machinery is the pair of topology-specific modules. TopoMDC is a family of eight directional difference convolutions built by extending central difference convolution to a two-pixel difference along each of eight directions; because local tube segments are directional and elongated, the router can select the combination of directions that matches the test patch's trajectory, and resetting the router per sample lets each test image receive its own topological prior. TopoHG is the counterpart that creates targeted training signal: it turns a confident foreground patch into a pseudo-broken hard sample by low-frequency Fourier swapping with a nearby low-confidence background patch, then applies consistency regularization with a position-dependent weight map so the model learns the local foreground/background cues needed to reconnect broken predictions.","core_discovery":"The central claim is that a source-trained CNN can be adapted to an unseen tubular-structure image at test time by using topological structure as the adaptation signal. Stage 1 replaces each vanilla $3\\times3$ convolution in the encoder with eight Topological Meta Difference Convolutions (TopoMDCs), each comparing the central pixel with one of eight neighboring directions; a learnable router reweights these operators per image patch, and only the 1,280 router parameters are updated, leaving pretrained weights untouched. Stage 2, called Topology Hard sample Generation (TopoHG), selects high-confidence foreground key points from teacher pseudo-labels, finds the neighboring background window with lowest pseudo-label confidence, swaps low-frequency Fourier components between the foreground and background patches, and overlays the edit only on pseudo-foreground pixels; the student is then aligned to the teacher's pseudo-label with a consistency loss that is up-weighted inside the pseudo-break regions, the teacher being updated by exponential moving average. The paper reports that this two-stage procedure beats seven prior test-time adaptation methods on clDice across all ten cross-domain pairings and on Dice in most settings, including cases where every comparison method degrades.","pith_inferences":["The paper's experiments are all in 2D; a testable extension is to lift the eight-direction TopoMDC family and the sliding-window pseudo-break search to 26-neighborhood 3D convolutions for airways or volumetric vessels, where continuity failures are arguably more damaging.","The contribution of TopoHG's low-frequency swap suggests that high-frequency texture carries cross-domain foreground identity; one could isolate this by ablating only the high-frequency preservation while keeping the same pseudo-break locations, rather than swapping whole augmentation schemes.","The teacher pseudo-labels are used only as supervision targets; the same per-patch confidence statistics that TopoHG computes (least-confident background windows near confident foreground) could be repurposed as an uncertainty or pseudo-label-quality signal for filtering, not just for hard-sample generation.","The router-reset design makes each test sample an independent adaptation episode, so streaming or batch settings with correlated samples may behave differently from the per-image iterations evaluated here; measuring TopoTTA on video or volume sequences would test that boundary."],"forward_implications":["Across ten cross-domain dataset pairs and two standard backbone families (plus a DSCNet variant), TopoTTA reports the highest clDice among all compared methods, with an average 31.81% improvement and gains over the second-best method of 4.66 clDice and 3.95 Dice on UNet.","The improvement is not bought by sacrificing segmentation accuracy: Dice also rises on most pairings, and in the DeepGlobe-to-CNDS case the method improves where every comparison method declines.","Because only router parameters are updated in Stage 1 and pseudo-break consistency needs no source labels in Stage 2, the pipeline fits the plug-and-play, per-sample online setting where target data arrive one image at a time.","The design applies to CNN-based tubular segmenters generally rather than a single architecture, which the paper demonstrates by reporting results with UNet, CS2Net, and a Stage-2-only DSCNet variant."],"supporting_citations":[{"why":"Supplies the central-difference convolution primitive that TopoMDCs generalize into eight directional operators.","marker":"[70]"},{"why":"Provides the entropy-minimization objective used to update router parameters in Stage 1 and is one of the seven baselines.","marker":"[58]"},{"why":"Provides the teacher-student consistency and exponential-moving-average update used in Stage 2 and is the strongest baseline beaten.","marker":"[60]"},{"why":"Defines clDice, the topological-continuity metric used to report the headline 31.81% improvement.","marker":"[52]"},{"why":"Provides the low-frequency Fourier swap used by TopoHG to create pseudo-breaks while preserving high-frequency foreground details.","marker":"[34]"},{"why":"Is one of the seven test-time adaptation baselines, representing medical-image segmentation methods.","marker":"[3]"},{"why":"Is a semantic-segmentation test-time adaptation baseline compared across all ten datasets.","marker":"[61]"},{"why":"Is a robustness-oriented test-time adaptation baseline compared across all ten datasets.","marker":"[43]"},{"why":"Is a batch-normalization-based test-time adaptation baseline compared across all ten datasets.","marker":"[72]"},{"why":"Is a robust batch-normalization baseline compared across all ten datasets.","marker":"[44]"}],"fun_headline_variants":["TopoTTA: topology-driven test-time adaptation for tubes","TopoTTA boosts clDice by 31.81% in unseen domains","Plug-and-play TTA for tubular structure segmentation","Two-stage TopoTTA adapts CNNs to new image domains","TopoTTA: first TTA framework for tubular structures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Stage 2's continuity gain assumes the teacher's pseudo-labels are trustworthy in the high-confidence regions where pseudo-breaks are inserted; if those labels are wrong, the consistency update teaches the model to reinforce the mistake and can worsen topological continuity.","fun_headline_variants_meta":{"raw":{"variants":["TopoTTA: topology-driven test-time adaptation for tubes","TopoTTA boosts clDice by 31.81% in unseen domains","Plug-and-play TTA for tubular structure segmentation","Two-stage TopoTTA adapts CNNs to new image domains","TopoTTA: first TTA framework for tubular structures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1688,"prompt_tokens":1020,"completion_tokens":668,"prompt_tokens_details":{"cached_tokens":896},"prompt_cache_hit_tokens":896,"prompt_cache_miss_tokens":124,"completion_tokens_details":{"reasoning_tokens":582}},"tokens_in":124,"tokens_out":668,"duration_ms":25376,"temperature":1.0,"reasoning_tokens":582,"cache_read_input_tokens":896,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:09:01.408900+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a target domain whose tubular appearance is systematically different from anything in the source training data, so the teacher produces confident but wrong pseudo-labels in whole regions; run Stage 2 alone and compare Betti errors against Stage 1 alone. If Betti errors increase or the clDice gain disappears, the pseudo-label reliability premise is what failed.","supporting_citations":[{"cited_title":"Searching Central Difference Convolutional Networks for Face Anti-Spoofing","cited_arxiv_id":null,"evidence_quote":"Supplies the central-difference convolution primitive that TopoMDCs generalize into eight directional operators."},{"cited_title":"TENT: Fully Test-Time Adap- tation by Entropy Minimization","cited_arxiv_id":null,"evidence_quote":"Provides the entropy-minimization objective used to update router parameters in Stage 1 and is one of the seven baselines."},{"cited_title":"Con- tinual Test-Time Domain Adaptation","cited_arxiv_id":null,"evidence_quote":"Provides the teacher-student consistency and exponential-moving-average update used in Stage 2 and is the strongest baseline beaten."},{"cited_title":"Pluim, Ulrich Bauer, and Bjoern H Menze","cited_arxiv_id":null,"evidence_quote":"Defines clDice, the topological-continuity metric used to report the headline 31.81% improvement."},{"cited_title":"FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Con- tinuous Frequency Space","cited_arxiv_id":null,"evidence_quote":"Provides the low-frequency Fourier swap used by TopoHG to create pseudo-breaks while preserving high-frequency foreground details."},{"cited_title":"Each Test Image Deserves a Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Seg- mentation","cited_arxiv_id":null,"evidence_quote":"Is one of the seven test-time adaptation baselines, representing medical-image segmentation methods."},{"cited_title":"Dynamically Instance- Guided Adaptation: A Backward-Free Approach for Test- Time Domain Adaptive Semantic Segmentation","cited_arxiv_id":null,"evidence_quote":"Is a semantic-segmentation test-time adaptation baseline compared across all ten datasets."},{"cited_title":"Towards Stable Test-Time Adaptation in Dynamic Wild World","cited_arxiv_id":null,"evidence_quote":"Is a robustness-oriented test-time adaptation baseline compared across all ten datasets."},{"cited_title":"Do- mainAdaptor: A Novel Approach to Test-Time Adaptation","cited_arxiv_id":null,"evidence_quote":"Is a batch-normalization-based test-time adaptation baseline compared across all ten datasets."},{"cited_title":"MedBN: Robust Test-Time Adapta- tion Against Malicious Test Samples","cited_arxiv_id":null,"evidence_quote":"Is a robust batch-normalization baseline compared across all ten datasets."}],"review_version":1}