{"id":"b2f55506-1f72-4922-a11d-cce854c137a6","arxiv_id":"2504.13553","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"HREFNet combines a high-resolution backbone, dynamic snake convolutions, an eight-direction Mamba scan, and multi-scale edge fusion to achieve reportedly state-of-the-art Dice scores on DRIVE, STARE, and CHASE DB1.","lead":"This paper presents HREFNet, a hybrid CNN and Mamba network for segmenting retinal blood vessels in eye images, and reports small accuracy improvements over existing methods on the DRIVE, STARE, and CHASE DB1 datasets. A generalist might read it because the architecture combines popular deep learning ideas and claims better handling of thin, low-contrast vessels relevant to eye disease screening.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported SOTA gains are within run-to-run noise; without protocol control or significance testing, Tables 1–3 cannot support the central claim.","rationale":"The reader's weakest assumption correctly identifies the strongest objection: the baseline comparison is not protocol-controlled and the claimed margins are tiny relative to the quoted standard deviations. I focused on this rather than on the unresolved Eq. (3) shape mismatch because the SOTA claim is the central claim and it rests entirely on the cross-table comparisons. The proposed check is decisive and feasible: if the small Dice advantages disappear when all methods are evaluated under one protocol, the headline claim fails; if they survive, the concern is resolved. This supports the CONDITIONAL verdict rather than changing it, since the architecture may be sound but the evidence in the current manuscript is insufficient to accept the claim as stated.","tokens_in":19152,"tokens_out":4828,"duration_ms":46129,"concrete_test":"Retrain HREFNet and the three closest competitors (Swin-UMamba, IMMF-Net, HRFormer) on DRIVE and STARE using identical preprocessing, splits, optimizer settings, and training schedule, with at least 5 seeds per method; compute paired per-image Dice and clDice differences and run a Wilcoxon signed-rank test. If HREFNet is not significantly better (p<0.05) under the shared protocol, Tables 1-3 do not support the claimed state-of-the-art performance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Tables 1–3 are the only evidence for the central 'state-of-the-art' claim, but the comparison is uncontrolled and the gains are small relative to reported variability. The paper never states whether U-Net, HRNet, HRFormer, DSCNet, U-Mamba, Swin-UMamba, and the other baselines were retrained under HREFNet's protocol (grayscale inputs, 224×224 random crops, batch size 1, AdamW with lr 1e-4, BCE loss, 300 epochs, same splits) or whether the numbers are quoted from papers using different preprocessing and splits. Under the quoted numbers, HREFNet's largest Dice advantage over the strongest competitor is 0.29 on DRIVE (82.14 vs 81.85), and on STARE it is 0.21 over only 5 test images with a reported standard deviation of ±3.94. Several metrics actually lose to baselines, including ACC on DRIVE to HRFormer, ACC/AUC on CHASE DB1 to Swin-UMamba, and AUC on STARE to U-Mamba. Without paired significance tests or protocol control, the reported gaps are indistinguishable from run-to-run noise, so the abstract's SOTA claim is not established by the tables as presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HREFNet, a hybrid CNN-Mamba architecture for retinal vessel segmentation built on a high-resolution (HRNet-like) backbone. Two components are introduced: the Dynamic Snake Visual State Space (DSVSS) block, which combines dynamic snake convolution with a Mamba-style selective scan extended to eight directions, and the Multi-scale Retina Edge Fusion (MREF) module, which fuses multi-scale dilated features with channel attention. The method is evaluated on DRIVE, STARE, and CHASE DB1 against CNN, Transformer, and Mamba baselines, reporting Dice, clDice, ACC, AUC, and HD95. The paper claims state-of-the-art performance and includes component-level ablations plus a stage-depth study. The code is made publicly available.","tokens_in":19444,"tokens_out":3242,"duration_ms":30035,"significance":"If the performance claims were robust, the paper would make a useful contribution to retinal vessel segmentation by demonstrating that combining curvature-adaptive convolution with state-space long-range modeling and multi-scale edge fusion can improve vessel continuity and boundary quality. The manuscript is honest in reporting several metrics where HREFNet does not win (e.g., ACC on DRIVE, ACC/AUC on CHASE DB1, AUC on STARE), and it ships code, which aids reproducibility. The component ablations on DRIVE are a positive feature. However, the central state-of-the-art claim is not supported by the evidence as presented: the reported margins over the second-best method are in several cases smaller than the reported standard deviations, no significance tests are provided, and the baseline comparison protocol is not described. The contribution is therefore credible as an engineering proposal but not yet established as a benchmark-leading method.","major_comments":[{"comment":"A concrete fix is to re-run all baselines under the same preprocessing, training schedule, and evaluation code, and to report paired significance tests (e.g., Wilcoxon signed-rank or bootstrap CIs) for the primary comparisons.","section":"§5.1.2, Tables 1–3"},{"comment":"This point is load-bearing because the final HREFNet configuration is not justified independently of the test-set results.","section":"§5.2.3, Table 6"},{"comment":"This is a load-bearing technical point for the proposed architecture and should be corrected in a revision.","section":"§3.1, Eq. (3)"},{"comment":"If the weights are indeed static, the authors should revise the terminology and the motivation accordingly; if the weights are input-dependent, the mechanism must be specified precisely.","section":"§3.2, Eqs. (7)–(8)"}],"minor_comments":[{"comment":"The clDice definitions reuse the symbols SP and VL for both the binary masks and their skeletons, which makes Eq. (14) confusing: Tprec is written as |SP ∩ VL| / |SP| without clearly indicating that one of these sets is a skeleton. Please use distinct symbols, e.g., S_P and S_L for skeletons, and correct the definitions accordingly.","section":"§4.2.2, Eq. (14)"},{"comment":"The citation for CHASE DB1 is incorrect: reference [42] is a paper on coronary angiography segmentation, not the CHASE dataset. Please cite the original CHASE DB1 source (Owen et al.) or the standard dataset reference.","section":"References, [42]"},{"comment":"The text contains a typo, 'Rencently', and the term 'Clustering Dice' is nonstandard; the topology-aware metric is conventionally called clDice, not 'Clustering Dice'. Please correct the spelling and terminology.","section":"§2.1"},{"comment":"The evaluation procedure is not fully specified: the paper does not state whether inference is performed on full-resolution images or on overlapping patches, whether test-time augmentation is used, or how the final segmentation is thresholded (fixed 0.5, Otsu, etc.). Description of these details would improve reproducibility of Tables 1–3.","section":"§5.1.2 and §4.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable engineering contribution with a reproducible codebase and component-level ablations. The main weakness is that the state-of-the-art claim is not supported by the current experimental evidence because of uncontrolled baselines and the lack of significance testing. The test-set model selection in Table 6 is another concern that should be addressed. If the authors can provide a controlled comparison with significance tests and fix the technical specification of Eq. (3), the paper could become acceptable. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent engineering paper, and the modules are reasonable, but the central SOTA claim is not established by the tables as presented. The genuinely new pieces are the eight-direction version of VMamba's SS2D scan with per-direction learned weights and the MREF edge-fusion module. Both are sensible for tubular structures, and the DRIVE ablation study is internally consistent: each component adds a small positive increment and the full model beats the baseline. The code link is a plus.\n\nI think the stress-test note is right. On DRIVE the best Dice margin over the runner-up is 0.29 (82.14 vs 81.85), and both numbers have standard deviations around 1.5. On STARE the margin is 0.21 over a baseline with a ±2.87 spread, and the test set has only five images. No significance tests are reported, and the paper never says whether the baselines were retrained under HREFNet's protocol or quoted from their original papers. Some metrics actually lose to baselines: ACC on DRIVE to HRFormer, AUC on CHASE DB1 to Swin-UMamba. So the cross-table comparisons cannot carry the abstract's claim.\n\nTwo other issues are worth naming. Equation (3) multiplies Q1 through Q4 element-wise, but those features come from different resolutions and the paper does not explain the upsampling. That is either a typo or an under-specified part of the architecture. And Table 6 selects HREFNet-large using DRIVE test-set results, which is a mild but real form of test-set model selection.\n\nNone of this is fatal. The architecture is a plausible combination, the ablations support the design choices, and the problems are mostly missing details rather than wrong results. What the paper needs is protocol transparency, paired significance tests or confidence intervals, a corrected Eq. (3), and a more careful treatment of model selection. If those are fixed, this becomes a solid incremental contribution.\n\nI would send it to peer review rather than desk reject it. There is enough substance, the experiments are broad, and the code is available. My own verdict would be conditional: ask for the clarifications above before accept. The paper is for readers tracking Mamba-based medical segmentation architectures; a specialist will want the protocol fixed, and the generic medical-imaging reader can probably skip it.","headline":"A competent incremental architecture whose SOTA claim is not backed by the tables as presented; the paper needs protocol transparency and significance testing more than it needs new modules.","tokens_in":19978,"tokens_out":2723,"would_cite":false,"duration_ms":26498,"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 hybrid CNN-Mamba network with curvature-aware scanning and multi-scale edge fusion reports the best retinal vessel segmentation on three public benchmarks.","keywords":["retinal vessel segmentation","Mamba","state space model","multi-scale edge fusion","Dynamic Snake Convolution","fundus imaging","medical image segmentation","low-contrast vessels"],"falsifier":"Re-run every baseline in Tables 1 to 3 under the exact protocol used for HREFNet: grayscale images, 224-by-224 random crops, AdamW with a learning rate that decays on plateau, 300 epochs, and the same train/test splits. If, under this matched protocol, HREFNet no longer holds the top Dice and clDice on all three datasets, then the central state-of-the-art claim is false.","tokens_in":18952,"feed_emoji":"👁️","tokens_out":11113,"duration_ms":93469,"temperature":0.7,"pith_summary":"Retinal vessel segmentation breaks down in two ways: thin vessels get fragmented and low-contrast boundaries become ambiguous. The paper proposes HREFNet, a hybrid network that addresses both failure modes by pairing a high-resolution-preserving backbone with a curvature-adaptive Mamba block and a multi-scale edge-fusion module. Results on DRIVE, STARE, and CHASE DB1 are reported as the best among the compared CNN, Transformer, and Mamba methods, with the strongest improvements in vessel continuity and edge clarity. The core claim a sympathetic reader should take away is architectural: high-resolution feature preservation, curvature-aware state-space scanning, and multi-scale edge fusion reinforce each other, and ablations show each component contributes to the final gain.","feed_headline":"New hybrid network tops retinal vessel segmentation on three datasets","feed_subtitle":"It reports top Dice and clDice on all three benchmarks, with clear gains on thin and low-contrast vessels.","key_machinery":"The load-bearing mechanism is the pair of new modules inserted into a high-resolution-preserving backbone. The DSVSS block couples Dynamic Snake Convolution, whose receptive field grows along a curved path so that it follows vessel geometry, with a Mamba state-space scan extended from four to eight directions and aggregated by learned softmax weights; this is what lets the model trace long, thin, winding vessels without losing continuity. The MREF module runs dilated convolutions at rates 1, 3, and 5, sums their outputs with a residual connection, and applies Squeeze-and-Excitation channel attention so that boundary features are emphasized and background noise suppressed. The argument rests on these modules because the ablations show each one improves the high-resolution baseline, and the combined model scores highest.","core_discovery":"The paper's central claim is that retinal vessel segmentation can be improved by fusing two mechanisms that address different failure modes: a selective-scan state-space block whose receptive field follows vessel curvature, and a multi-scale edge-fusion module that sharpens boundaries. HREFNet places these modules on an HRNet-style high-resolution backbone, replaces the standard blocks with the Dynamic Snake Visual State Space (DSVSS) block, and inserts the Multi-scale Retina Edge Fusion (MREF) module after each stage. The DSVSS block combines Dynamic Snake Convolution with a Mamba state-space scan extended from four to eight directions and aggregates the directional features with learned softmax weights; MREF extracts edge features with dilated convolutions at rates 1, 3, and 5, recalibrates them with channel attention, and fuses them back into the stage output. On this architecture the paper reports the best Dice and clDice scores on DRIVE, STARE, and CHASE DB1 among the compared methods, and the ablation study attributes the improvement to the two modules working together.","pith_inferences":["A natural extension is to test whether the same two modules help on other elongated or tubular structures, such as corneal nerves or coronary arteries, where thin, curved, low-contrast targets produce the same fragmentation and blur failure modes.","The eight-directional snake scan is architecture-agnostic; dropping it into a Transformer-based segmentation head would tell whether the gain comes from directional scanning per se or from the Mamba state-space backbone.","Because the paper reports results from only 20 to 28 training images per dataset, an open question is whether HREFNet's advantage holds on larger or multi-center fundus collections with more domain shift."],"forward_implications":["On the small training sets typical of retinal datasets, the architecture reports competitive results without pretraining or external data, which matters for clinical settings where annotated images are scarce.","Because the ablation shows DSVSS and MREF each add measurable gains and the largest gain comes from their combination, future designs can treat curvature-aware scanning and edge fusion as complementary rather than competing strategies.","The eight-directional scan with per-direction softmax weighting is defined as a general mechanism and can be lifted into other Mamba-based segmentation networks to test whether the continuity gain transfers beyond retinal vessels.","The reported HD95 improvements imply that the model not only finds more vessel pixels but also places their boundaries more accurately, which is the property most relevant for downstream clinical measurements."],"supporting_citations":[{"why":"Defines the U-Net baseline architecture that HREFNet is compared against on all three datasets.","marker":"[6]"},{"why":"Provides the Swin-UMamba baseline that is HREFNet's closest competitor on the CHASE DB1 benchmark.","marker":"[13]"},{"why":"Supplies the high-resolution multi-stage backbone that HREFNet preserves and extends.","marker":"[14]"},{"why":"Supplies Dynamic Snake Convolution, the curvature-adaptive receptive field mechanism the DSVSS block builds on.","marker":"[15]"},{"why":"Supplies the VMamba 2D Selective Scan mechanism that HREFNet extends from four to eight directions.","marker":"[16]"},{"why":"Supplies the Squeeze-and-Excitation channel attention used by MREF to recalibrate edge features.","marker":"[39]"},{"why":"Identifies the DRIVE dataset used as the first benchmark for training and evaluation.","marker":"[40]"},{"why":"Identifies the STARE dataset used as the second benchmark for training and evaluation.","marker":"[41]"},{"why":"Cited as the source of the CHASE DB1 dataset used as the third benchmark for training and evaluation.","marker":"[42]"},{"why":"Defines the clDice metric used to evaluate vascular topology and continuity, the paper's main claimed improvement.","marker":"[44]"}],"fun_headline_variants":["Hybrid CNN-Mamba improves retinal vessel segmentation on three benchmarks","Dynamic snake convolution and edge fusion sharpen vessel boundaries","Curvature-aware Mamba scan boosts retinal vessel continuity","Multi-scale edge fusion enhances low-contrast vessel detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The tables assume that every comparison method was measured under the same training and evaluation conditions as HREFNet, but the paper never states whether those baselines were re-trained in-house, which preprocessing was applied, or which splits produced the quoted numbers; if those numbers come from different protocols, the reported margins may reflect protocol differences rather than architectural gains.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid CNN-Mamba improves retinal vessel segmentation on three benchmarks","Dynamic snake convolution and edge fusion sharpen vessel boundaries","Curvature-aware Mamba scan boosts retinal vessel continuity","Multi-scale edge fusion enhances low-contrast vessel detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000341,"raw_usage":{"total_tokens":1932,"prompt_tokens":1053,"completion_tokens":879,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":813}},"tokens_in":669,"tokens_out":879,"duration_ms":8401,"temperature":1.0,"reasoning_tokens":813,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:06:05.791046+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run every baseline in Tables 1 to 3 under the exact protocol used for HREFNet: grayscale images, 224-by-224 random crops, AdamW with a learning rate that decays on plateau, 300 epochs, and the same train/test splits. If, under this matched protocol, HREFNet no longer holds the top Dice and clDice on all three datasets, then the central state-of-the-art claim is false.","supporting_citations":[{"cited_title":": Swin-umamba: Mamba-based unet with imagenet- based pretraining","cited_arxiv_id":null,"evidence_quote":"Provides the Swin-UMamba baseline that is HREFNet's closest competitor on the CHASE DB1 benchmark."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the high-resolution multi-stage backbone that HREFNet preserves and extends."},{"cited_title":"In: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision, pp","cited_arxiv_id":null,"evidence_quote":"Supplies Dynamic Snake Convolution, the curvature-adaptive receptive field mechanism the DSVSS block builds on."},{"cited_title":"In: The Thirty-eighth Annual Conference on Neural Information Processing Systems (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the VMamba 2D Selective Scan mechanism that HREFNet extends from four to eight directions."},{"cited_title":"IEEE transactions on medical imaging 23(4), 501–509 (2004)","cited_arxiv_id":null,"evidence_quote":"Identifies the DRIVE dataset used as the first benchmark for training and evaluation."},{"cited_title":"IEEE Transactions on Medical imaging 19(3), 203–210 (2000)","cited_arxiv_id":null,"evidence_quote":"Identifies the STARE dataset used as the second benchmark for training and evaluation."},{"cited_title":"Biomedical Signal Processing and Control 46, 1–9 (2018)","cited_arxiv_id":null,"evidence_quote":"Cited as the source of the CHASE DB1 dataset used as the third benchmark for training and evaluation."}],"review_version":1}