{"id":"c8722433-f1d2-44fc-b186-15b9f8b155fb","arxiv_id":"2607.23371","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"For vessel segmentation in microscopy and fundus images, CNNs lean on local pixel intensity and a ~20-pixel effective context; texture and global geometry play smaller roles.","lead":"Blood-vessel segmentation networks were tested to see which visual cues they truly rely on in microscopy and retinal images. The results: local intensity and a roughly 20-pixel context dominate, while texture and global shape contribute little — a quantitative basis for smaller, more auditable models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cue isolation is not clean: D−i,−t's >75% accuracy is likely explained by per-patch intensity histograms, so the intensity-vs-texture ranking is unestablished.","rationale":"The reader's weakest_assumption is exactly this: Section III.B assumes normalization removes 'intensity' and shuffling removes 'texture', but the D−i,−t result shows residual signal, so the boundary is not observationally closed. I agree. The single most load-bearing concern is that the central empirical claim — intensity is more relevant than texture — is not established because the perturbations are not selective. The paper is transparent about residual features being an open problem, and the ERF/patch-size experiments are independent and sound in outline, so the work retains value. A histogram-only baseline would settle whether the residual accuracy is just intensity statistics. Since the reader already flagged this and the verdict is CONDITIONAL, my read does not change the verdict; it reinforces it. No fatal flaw, but a concrete control is needed before the headline ranking can be accepted.","tokens_in":17490,"tokens_out":3464,"duration_ms":35625,"concrete_test":"Train a simple baseline on the same D, D−i, D−t, D−i,−t patches using only per-patch intensity histograms — e.g., a linear classifier on the sorted 81 pixel values, with no spatial structure. If this histogram-only baseline reproduces the CNN's accuracy profile (especially >75% on D−i,−t), then the perturbations did not remove intensity information, and the claimed intensity-vs-texture ranking is an artifact of the probe rather than a property of the segmentation task.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III.B's cue-isolation logic is the load-bearing point. Per-patch z-normalization (Eq. 1) removes each patch's mean and variance, but preserves relative intensity differences (contrast) and, after shuffling, the full intensity histogram. Per-patch shuffling removes spatial arrangement but preserves the intensity histogram exactly. Thus D−i removes only global brightness/scale, and D−t removes only spatial order, not 'texture' in any ordinary sense. The reported >75% accuracy on D−i,−t (Fig. 6) is therefore not evidence that networks operate without intensity and texture; it is what a histogram-based classifier would achieve. The D−i vs D−t comparison conflates the amount of perturbation with the semantic cue being removed. The headline 'pixel intensity is more relevant than texture' and the conclusion that CNNs are local-intensity machines rest on this contrast. If the residual accuracy is attributable to intensity statistics that survive normalization, the ranking is an artifact of the probe rather than a property of vessel segmentation. This is load-bearing because the architecture recommendation (lightweight models with ~20-pixel ERF) and the XAI roadmap assume this ranking is real.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper investigates which visual cues drive CNN-based vascular segmentation, using fluorescence microscopy (VessMAP) and fundus photography (DRIVE) as case studies. Three experiment blocks are reported: (i) 9x9-patch classification after per-patch intensity normalization, pixel shuffling, or both, in order to isolate intensity and texture; (ii) segmentation from contour-only (Dc) and centerline-only (Ds) inputs in order to isolate shape; and (iii) a sweep of kernel size, dilation, and downsampling across 160 U-Net variants, plus training on WxW input patches, to measure theoretical and effective receptive fields. The authors report that intensity is more important than texture, that more than 75% patch accuracy survives removal of both, that shape cues alone yield Dice scores of about 54-57% versus roughly 90% with full information, and that performance saturates with an effective receptive field near 20 pixels (32x32 context for VessMAP), while DRIVE continues to benefit from larger contexts.","tokens_in":17617,"tokens_out":9201,"duration_ms":98391,"significance":"If established, the findings would support practical design recommendations: lightweight U-Nets with modest receptive fields are sufficient for these modalities, and XAI/saliency methods should be made cue-aware rather than generic. The experimental design is generally transparent: repeated runs (5x patch classification, 3x RF sweep), confidence intervals, public code, and explicit discussion of limitations. The ERF is measured with an external method rather than inferred from architecture alone, and the patch-size manipulation offers a clean test of context need. However, the interpretation of the patch-perturbation experiment is not clean because the 'removal' operations preserve first-order intensity statistics; this undermines the intensity-vs-texture ranking. The conclusion about shape is partly supported by the non-ill-posed Dc experiment but is conflated with the ill-posed Ds experiment. The paper is therefore a valuable and reproducible case study, but its headline interpretive claims need revision.","major_comments":[{"comment":"The labels 'intensity removed' for D−i and 'texture removed' for D−t are not accurate as cue isolation. Eq. (1) subtracts the patch mean and divides by the patch standard deviation; it preserves within-patch relative intensity differences and the full shape of the normalized intensity histogram. Random shuffling preserves the intensity histogram exactly and removes only spatial order. Thus D−i,−t still contains complete first-order intensity statistics, and the reported >75% accuracy on D−i,−t is what a histogram-based classifier could achieve. The comparison between D−i and D−t conflates the type of perturbation (removing mean/variance vs. removing spatial order) with the semantic cue being removed. The headline 'pixel intensity is more relevant than texture' (Abstract, §V) rests on this comparison and is not established. I recommend adding a baseline classifier trained on per-patch int","section":"§III.B, Eq. (1), Fig. 6"},{"comment":"The centerline dataset Ds is ill-posed by construction: vessel caliber cannot be recovered from a centerline, so a Dice ceiling below 100% is guaranteed regardless of model capacity. The manuscript acknowledges this in §III.C, yet the Abstract and §IV.B treat Ds as part of the evidence that 'CNNs struggle to extrapolate full vessel geometry from shape cues alone.' The contour dataset Dc is not ill-posed in the same way and provides the actual test of shape extrapolation; its low Dice (54.4% for W-Net) supports the claim. I recommend presenting Dc as the primary shape-only experiment and Ds as an extreme auxiliary condition, and explicitly stating that the Ds result is an expected lower bound. This would remove the misleading impression that 57.0% Dice on Ds is a meaningful 'best' shape-only result.","section":"§III.C, Table I"},{"comment":"The intensity/texture conclusions are drawn from 9x9-patch classification on VessMAP only, but the Abstract and Conclusion state them as properties of vessel-segmentation networks in the evaluated modalities. Patch classification is not the same task as full-image segmentation, and the paper does not show that the residual features exploited on 9x9 patches (e.g., intensity histograms) transfer to segmentation. Because DRIVE was excluded from the patch experiments because its vessels are too thin, the conclusion 'for these modalities, pixel intensity is more critical than local texture' overstates the support. Please either restrict the claim to VessMAP patch classification or add a segmentation-level probe (e.g., apply the same perturbations to full images and measure Dice changes).","section":"§III.B/§IV.A"},{"comment":"The RF sweep varies kernel size, which changes the number of trainable parameters as well as the receptive field. The manuscript acknowledges this in §IV.C ('Increasing the kernel size also adds more trainable parameters and, in turn, increases capacity'), but the following discussion still attributes performance gains to larger receptive fields. The patch-size experiment (Fig. 9) is a cleaner test of context need and does support the saturation conclusion for VessMAP, so this issue is not fatal. However, the manuscript should either control for parameter count (e.g., by adjusting width/depth) or explicitly state that the kernel-size results measure a joint capacity-plus-RF effect, and should adjust the text accordingly.","section":"§IV.C, Fig. 7-9"}],"minor_comments":[{"comment":"DRIVE images are originally 584x565, but H is set to 512 in the patch experiment. The manuscript does not state whether images were resized or cropped before patchification. This should be specified, as it can affect the global-context conclusion for DRIVE.","section":"§III.D, Fig. 9"},{"comment":"The statement that the shape results for DRIVE were 'similar' to VessMAP is not supported by any table or figure. Please report at least the DRIVE Dice values (or include them in an appendix) to substantiate the generalization claim.","section":"§IV.B"},{"comment":"The text reports losses relative to the base dataset (e.g., 'accuracy loss of 7.1%') but does not give the absolute accuracies of D, D−i, D−t, and D−i,−t. Please include the absolute values in the text or figure caption for reproducibility.","section":"§IV.A, Fig. 6"},{"comment":"The related work is adequate, but the comparison with Loos et al. could note explicitly that the present RF sweep keeps depth fixed while varying kernel/dilation/pooling, which clarifies the difference in the manipulation.","section":"§II"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope and the methodology is reproducible, but the cue-isolation claim is not yet established. The ERF and Dc experiments are valuable and can likely be published after the interpretation is corrected and the requested controls are added. I would encourage a major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a transparent, well-scoped empirical study with one clean and useful result — the ERF saturation around 20 pixels / 32×32 patch size for these two vessel datasets — and one load-bearing claim that does not survive close reading: the \"pixel intensity > texture\" ranking.\n\nThe receptive-field experiment is the strongest part. Sweeping kernel sizes, dilations, and pooling across 160 configurations, measuring ERF with the Luo et al. method, and getting a consistent saturation plateau is solid work. The patch-size experiment corroborates it nicely on VessMAP (saturation at 32×32) and honestly reports that DRIVE keeps improving with context. That finding is independent of the cue-isolation problem and is worth building on.\n\nThe shape-cue ablation is informative but weaker. The contour condition (Dc) is a legitimate probe: best Dice 54.4% vs ~90% with full information, with the predicted oversegmentation pattern. The centerline condition (Ds) is, as the authors admit, ill-posed — no network can infer vessel caliber from a centerline — so it adds little beyond confirming the trivial.\n\nThe soft spot is the texture/intensity isolation. The stress-test note is right: per-patch z-normalization removes mean and variance but preserves the intensity histogram; pixel shuffling preserves the same multiset of values. So D−i,−t does not remove \"intensity\" in any complete sense. The >75% accuracy on that condition is likely the network exploiting standardized intensity distribution shape, not some mysterious \"other features.\" And the D−i vs D−t comparison is apples-to-oranges: normalization and shuffling destroy different kinds of information, so the observed accuracy drop does not license the claim that intensity matters more than texture. The paper's own framing (\"intensity normalization is relevant mostly for patches containing only vessel or background pixels\") hints at this, but the abstract and conclusion state the ranking unhedged.\n\nAll that said, the authors are unusually upfront about limitations, the code and hyperparameters are public, and the experiments are repeated with confidence intervals. The paper deserves a serious referee — a good reviewer would send it back for a revised cue-isolation methodology (e.g., replacing patch intensities with sampled noise matched only on low-order statistics) or for substantially toned-down claims. The ERF and shape results are publishable as a case study; the cue-ranking headline needs fixing first.","headline":"Solid ERF and shape-cue experiments, but the intensity-vs-texture ranking is not established because the perturbation procedure preserves the intensity histogram.","tokens_in":18236,"tokens_out":3570,"would_cite":false,"duration_ms":38424,"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":"This paper argues that CNNs segment blood vessels primarily from local pixel intensity, with an effective receptive field saturating near 20 pixels, and that shape cues alone are insufficient.","keywords":["blood vessel segmentation","visual cues","texture bias","intensity","effective receptive field","U-Net","fluorescence microscopy","fundus imaging"],"falsifier":"Train a U-Net on DRIVE with an architectural cap that forces the effective receptive field below 10 pixels, for example using all 1×1 convolutions with no pooling, and show it matches the base model's Dice; that would break the claim that roughly 20 pixels of context are needed. Alternatively, on VessMAP, show that 32×32 input patches yield a large Dice drop relative to full images once data augmentation or class balance is changed, revealing that the plateau is an artifact of the particular training setup rather than a true context budget.","tokens_in":1366,"feed_emoji":"🔬","tokens_out":1649,"duration_ms":52252,"temperature":0.7,"pith_summary":"This paper tries to establish which visual cues convolutional networks actually use when segmenting blood vessels in fluorescence microscopy and retinal fundus images, and to quantify how much spatial context they need. Using per-patch pixel shuffling and intensity normalization, it argues that pixel intensity is more important than local texture, yet networks still score above 75% accuracy when both cues are removed, implying they exploit other statistical regularities. Training networks on contours and centerlines alone, it finds shape cues are insufficient (best Dice 54–57% versus 90% with full images), so networks do not reconstruct vessel geometry from boundaries. The paper also claims the effective receptive field saturates around 20 pixels across both datasets, a 32×32 patch reaches peak performance in microscopy, and only fundus images show modest gains from larger context. If correct, this means lightweight local-feature models are enough for many vessel-segmentation settings and that failures are better explained by local appearance than missing global shape.","feed_headline":"Vessel networks work with just 20 pixels of context","feed_subtitle":"Local pixel intensity beats texture and shape; shape-only cues score ~57% Dice vs 90% with full images.","key_machinery":"The identifying machinery is the cue-isolation protocol: patch-level pixel shuffling to destroy texture, patch-level mean and standard-deviation normalization to destroy absolute intensity, homogenization (both together), contour-only and centerline-only inputs to strip internal vessel appearance, and a systematic sweep of kernel sizes, dilation ratios, and pooling across 160 U-Net configurations to map theoretical versus effective receptive fields, with the ERF measured by gradient magnitudes. This protocol lets the authors decompose performance into separate contributions from intensity, texture, shape, and spatial context, and to locate the context budget beyond which further architectura","core_discovery":"The central discovery is that these CNNs are local-intensity machines: segmentation decisions rest on a compact spatial window of about 20 pixels, and local contrast and pixel statistics dominate over texture as defined by spatial arrangement. The evidence: removing texture by shuffling pixels inside 9×9 patches costs only 0.1% classification accuracy, while removing intensity by normalization costs more; removing both still leaves above 75% accuracy; and supplying only contours or centerlines of vessels, without interior appearance, drops Dice from about 90% to 54–57%, with strong oversegmentation. The receptive-field experiments show performance plateaus near an average kernel size of 3 an","pith_inferences":["If the patch-level cue isolation transfers to whole-image segmentation, a practical shortcut is to audit new imaging domains with the same cheap probes—shuffle, normalize, and measure effective receptive field—before committing to large architecture searches.","The residual accuracy after homogenization suggests the network may be using per-patch variance, local contrast statistics, or edge orientation that survive shuffling; a testable next experiment is to explicitly ablate these statistics, for example by matching variances across classes, and check whether accuracy collapses.","The 20-pixel effective receptive field may be a property of the thin vessels in these datasets rather than of the task itself; applying the same protocol to high-resolution images with larger vessels should shift the saturation point, which would sharpen the claim.","The paper's proposed cue-specific attribution maps could be built directly from the measured performance deltas: saliency weights for intensity, texture, and shape are quantified by the three ablations, turning a qualitative bias diagnosis into a per-model score."],"forward_implications":["For microscopy datasets like VessMAP, a model with an effective field near 20 pixels and input crops of 32×32 reaches the same plateau as full-image models, so deeper or globally attentive architectures add little for this modality.","Dilated convolutions that enlarge the theoretical receptive field produce no consistent Dice gains, indicating the networks already operate with sufficient context at the base configuration.","Because shape cues alone yield around 54–57% Dice versus 90% with full information, vessel segmentation models should be expected to fail when local appearance is ambiguous, not because they lack global geometry.","For fundus images, global context helps modestly, so dataset-specific global structure such as the optic disk and retinal boundaries is part of what models use; architecture choices should account for this.","Networks retain above-chance patch classification after both intensity and texture perturbations, implying they rely on additional uncharacterized statistical features; identifying those features becomes a clear next step."],"fun_headline_variants":["Vessel AI relies on just 20 pixels of local intensity","Texture-free: CNNs still segment vessels with pixel intensity","Shape cues fail: vessel segmentation drops to ~57% Dice","Local intensity, not texture, drives vessel segmentation","Blood-vessel CNNs need only 20-pixel context"],"cache_read_input_tokens":19456,"weakest_assumption_plain":"The ranking of intensity over texture rests on the assumption that shuffling a 9×9 patch removes texture while leaving everything else intact, and that per-patch normalization removes intensity without destroying features the network uses; the paper's own result that accuracy stays above 75% even after both perturbations shows this separation is not clean.","fun_headline_variants_meta":{"raw":{"variants":["Vessel AI relies on just 20 pixels of local intensity","Texture-free: CNNs still segment vessels with pixel intensity","Shape cues fail: vessel segmentation drops to ~57% Dice","Local intensity, not texture, drives vessel segmentation","Blood-vessel CNNs need only 20-pixel context"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00082,"raw_usage":{"total_tokens":3423,"prompt_tokens":739,"completion_tokens":2684,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":2616}},"tokens_in":483,"tokens_out":2684,"duration_ms":16961,"temperature":1.0,"reasoning_tokens":2616,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T23:37:26.622769+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a U-Net on DRIVE with an architectural cap that forces the effective receptive field below 10 pixels, for example using all 1×1 convolutions with no pooling, and show it matches the base model's Dice; that would break the claim that roughly 20 pixels of context are needed. Alternatively, on VessMAP, show that 32×32 input patches yield a large Dice drop relative to full images once data augmentation or class balance is changed, revealing that the plateau is an artifact of the particular training setup rather than a true context budget.","supporting_citations":[],"review_version":1}