{"id":"93e14db3-6bbd-4ac6-b043-33e7106c82c4","arxiv_id":"1908.04123","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A top-hat transform, CLAHE, and multiscale 2D-Gabor wavelet filters followed by Otsu thresholding segment retinal vessels on DRIVE with 94.32% accuracy.","lead":"This paper describes an unsupervised method that maps retinal blood vessels in eye images by combining morphological filters and Gabor wavelets. On a standard public dataset it reports 94.32% pixel accuracy, which could be a low-cost building block for automated eye-disease screening.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's claim of superior accuracy is contradicted by the paper's own Table 2, undermining the stated central claim.","rationale":"The paper's central claim, as stated in the abstract and echoed in the reader's strongest claim, is that the proposed algorithm produces better performance on DRIVE in terms of accuracy, sensitivity, and kappa agreement. The load-bearing weakness is that the paper's own Table 2 contradicts the accuracy component of that claim. This is a concrete, internal inconsistency rather than a mere disagreement with the field's consensus. The reader's chosen weakest assumption was parameter tuning without a separate validation split, which is a legitimate reproducibility concern, but it is secondary to the fact that the advertised superiority over 'major algorithms' is not demonstrated even in the paper's own comparison table. The reader did mention this contradiction in the rationale, so there is partial agreement, but it was not identified as the weakest assumption. The appropriate response is to keep the conditional verdict: the underlying pipeline may be a reasonable engineering contribution, but the abstract's comparative claim must be corrected or precisely scoped, and the parameter-selection procedure must be documented and preferably validated on a hold-out set. A full rejection would be too strong because the reported absolute metrics are plausible and the method is described in enough detail to be reimplemented, but the current text cannot be accepted as written without revision.","tokens_in":10406,"tokens_out":6495,"duration_ms":63776,"concrete_test":"List all rows in Table 2 that count as 'major algorithms'; verify for each row whether Acc ≤ 0.9432, Se ≤ 0.7503, and kappa ≤ 0.7374. Since at least Soares (Acc 0.9466), Staal (Acc 0.9441), Ricci (Acc 0.9595), Lahiri (Acc 0.9530), and Zhao (Acc 0.9477) already violate the accuracy inequality, the comparison claim fails unless these rows are explicitly excluded. The test is simply to require the authors to specify the comparison set and recompute the claim; no new experiments are needed. If they intend to claim superiority only over a named set of unsupervised baselines, the abstract and Section 4 must name that set and show a statistical comparison, not just cite aggregate numbers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the abstract—'our algorithm produces better performance concerning the accuracy, sensitivity, and kappa agreement'—is not supported by the paper's own comparative results. In Table 2, the proposed method's accuracy (0.9432) is below six listed entries: Soares et al. (0.9466), Staal et al. (0.9441), Ricci et al. (0.9595), Lahiri et al. (0.9530), Zhao et al. (0.9477), and the 2nd observer (0.9473). This failure persists even inside the unsupervised block: Zhao et al. and the 2nd observer have higher accuracy, and Miri et al. (0.9458) also exceeds the reported value. Therefore the sentence 'better performance concerning accuracy' is either false or relies on an undefined subset of 'major algorithms' that excludes the best-performing rows. If the comparison set is the full table, the claim fails on accuracy; if the set is restricted, the restriction is never stated. This matters because the abstract is the primary evidence for novelty; once the superiority claim is removed, the contribution reduces to a plausible but unvalidated configuration of standard image-processing tools. The remaining metrics (Se = 0.7503, kappa = 0.7374) are competitive but not uniformly better, since the 2nd observer has Se = 0.7760 and Gou et al. have Se = 0.7526. The authors must either specify the comparison set precisely and provide a paired statistical test, or revise the claimed superiority.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents an unsupervised retinal vessel segmentation method. The pipeline is: take the green channel of the retinal image, invert it and apply the fundus mask, apply a white top-hat transform with a circular structuring element of diameter 11, apply CLAHE, filter with a bank of multiscale 2D-Gabor wavelets (radial frequencies 0.7 to 1.3 at intervals of 0.2, orientations 0 to 180 degrees at 20 degree intervals, K=2.2), take the maximum response per pixel, and threshold globally with Otsu. The method is evaluated on the DRIVE test set (20 images) using accuracy, sensitivity, specificity, area under the ROC curve, and kappa agreement. Reported averages are Acc=0.9432 (std=0.0049), Se=0.7503, Sp=0.9717, Az=0.9524, and kappa=0.7374. The authors claim that their algorithm produces better performance than major algorithms with respect to accuracy, sensitivity, and kappa agreement.","tokens_in":10734,"tokens_out":6065,"duration_ms":60284,"significance":"The paper is an empirical benchmark study. Its main potential value is that the proposed method is fully unsupervised, requires no classifier training, and is evaluated on the public DRIVE test set with standard metrics and detailed per-image results in Table 1. If the performance claims are supported, the method would be a useful simple baseline for retinal vessel segmentation. However, the strength of the paper depends on the validity of the superiority claim and on the robustness of the reported metrics to parameter selection; both currently require additional evidence. The paper does not provide code, and the comparison table is assembled from previously published reports, so the comparison is only as reliable as those sources.","major_comments":[{"comment":"The statement that the algorithm 'produces better performance concerning the accuracy, sensitivity, and kappa agreement' is not supported by the reported numbers. The proposed accuracy (0.9432) is lower than at least seven rows in Table 2: Soares et al. (0.9466), Staal et al. (0.9441), Ricci et al. (0.9595), Lahiri et al. (0.9530), Zhao et al. (0.9477), the 2nd observer (0.9473), and Miri et al. (0.9458). The sensitivity (0.7503) is also lower than the 2nd observer (0.7760) and Gou et al. (0.7526). If the comparison set is meant to be a subset of 'major algorithms', that subset is not defined. Please specify the comparison set and provide paired statistical tests (e.g., per-image McNemar or bootstrap tests) for the differences, or revise the claimed superiority.","section":"Abstract and Section 4 (Table 2)"},{"comment":"The parameter choices are described as selected by 'few experiments on retinal images and Gabor wavelet filters'; the structuring element diameter is set to 11 pixels based on the stated DRIVE vessel width range of 1-10 pixels. Because the evaluation is performed on the DRIVE test set and no separate validation split is described, the reported accuracy may be optimistically biased by test-set-informed parameter selection. Please describe a validation protocol (for example, tuning on the DRIVE training set and then evaluating on the test set) and evaluate on external databases such as STARE or CHASE DB1 to support the generalization claim. The final section lists testing on those databases as future work, but such testing is necessary to support the current claims.","section":"Section 3.3"},{"comment":"The sentence defining the ROC computation states 'false positive rate (XR = 1−Sp) and true positive rate (ZR = Sp)', which equates the true positive rate with specificity. This is inconsistent with Eq. (6), where sensitivity is Se = Z/(Z+Y), and with standard ROC analysis, where TPR = Se. If this definition was used to compute the reported Az = 0.9524, the value is not a valid area under the ROC curve. Please correct the notation and confirm whether all reported Az values were computed with TPR = Se and FPR = 1−Sp.","section":"Section 4 (ROC description)"}],"minor_comments":[{"comment":"The conclusion text is duplicated several times, and the figure captions repeat the same 'Fig. 7' caption and also label the same material as 'Fig. 5'. The manuscript should be cleaned up so that each section and figure appears exactly once with a unique caption.","section":"Section 5 and Figure captions"},{"comment":"The symbols XR and ZR are used without definition; please define them clearly as the false positive rate and the true positive rate, respectively, and ensure they are used consistently with Eq. (6).","section":"Section 4 (ROC description)"},{"comment":"In Eq. (3), the notation ψθ(a,x,y,x0,y0) uses the scale parameter a both as an argument of the function and inside the argument list, which is confusing. Please use distinct symbols for the scale parameter and the spatial coordinates.","section":"Equation (3)"},{"comment":"There are several typographical errors, including 'Lee at el.' instead of 'Lee et al.', 'we have been proposed' instead of 'we have proposed', and inconsistent use of 'Tw' versus 'Twav' for the wavelet transform. A careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The load-bearing issue is the unsupported superiority claim in the abstract and Section 4. If the authors revise that claim, add a validation protocol, and correct the ROC definition, the contribution is modest but within the scope of an applied image-processing journal. The paper would be strengthened by including statistical tests for the comparison and by reporting results on at least one additional public database."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's central assertion, from the abstract and results, is that the proposed unsupervised pipeline (top-hat transform, CLAHE, multiscale 2D-Gabor wavelets, Otsu thresholding) achieves 'better performance concerning accuracy, sensitivity, and kappa agreement' than major algorithms on DRIVE. That claim does not survive contact with Table 2. The proposed method's accuracy (0.9432) is below Soares (0.9466), Staal (0.9441), Ricci (0.9595), Lahiri (0.9530), Zhao (0.9477), and even the 2nd observer (0.9473). Sensitivity (0.7503) is not uniformly better either, since the 2nd observer (0.7760) and Gou et al. (0.7526) are higher. Kappa is indeed the best in the listed set, but the abstract's phrasing is too broad.\n\nWhat is genuinely new here is a specific ordering: top-hat morphological enhancement before CLAHE, then a Gabor filter bank with max response, then Otsu. Each component is standard, and similar combinations appear in prior work (e.g., CLAHE plus Gabor in Zhao et al.), but the particular configuration and its DRIVE metrics are not reported in the cited papers. The authors also deserve credit for reporting per-image results, including a pathological image, and for explicitly noting the dependence on the structuring element and the need to test on other databases. Those are honest limitations.\n\nThe main soft spots are the parameters and the comparison. Section 3.3 says parameters were selected by 'performing few experiments on retinal images,' with no validation split. That means the reported performance could have been tuned with knowledge of the test set. A separate validation set, or at least a sensitivity analysis, is needed before the numbers become predictive. The comparison itself is also informal: metrics are taken from other papers, with no paired statistical test. Even if the abstract is corrected, the claim would need to be scoped to specific metrics and specific comparators.\n\nThis is a modest engineering contribution. It does not solve an open problem or enable a new clinical capability, but it is a reasonable addition to the long list of unsupervised retinal vessel segmentation approaches. The paper is clearly written and the method is described well enough to reproduce in principle, though no code is released.\n\nA serious referee should not desk-reject it, but should require the authors to fix the abstract, add an appropriate validation procedure, and either release code or give complete parameter settings. I would not cite it in my own work in the next year, but I would bring it to a reading group to discuss how easily benchmark claims can overstate what a table actually shows.","headline":"The pipeline is a plausible configuration of standard tools, but the abstract's superiority claim is contradicted by the paper's own Table 2 and should be revised before publication.","tokens_in":11241,"tokens_out":2028,"would_cite":false,"duration_ms":22825,"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":"Retinal vessel segmentation without training labels reaches 94.32% accuracy on DRIVE.","keywords":["retinal vessel segmentation","unsupervised segmentation","2D-Gabor wavelet","top-hat transform","CLAHE","Otsu thresholding","DRIVE database","fundus imaging"],"falsifier":"Tune nothing: run the published pipeline with the stated parameters on the STARE and CHASE DB1 databases and compare against their manual ground truths; if accuracy falls materially below the DRIVE figure, or if the structuring element must be re-tuned per database to recover performance, the claim that the algorithm generalizes in its fixed form is refuted.","tokens_in":10210,"feed_emoji":"👁️","tokens_out":10556,"duration_ms":95159,"temperature":0.7,"pith_summary":"The paper attempts to establish that a fully unsupervised sequence of image-processing steps—white top-hat morphological transformation, contrast-limited adaptive histogram equalization (CLAHE), multiscale 2D-Gabor wavelet filtering, and Otsu thresholding—can segment retinal blood vessels in fundus photographs accurately enough to compete with trained classifiers. On the DRIVE benchmark, the authors report an average accuracy of 94.32% with a standard deviation of 0.0049, sensitivity 75.03%, specificity 97.12%, and a kappa agreement of 0.7374. They argue that the top-hat step suppresses broad background structures such as the optic disc, fovea, and exudates, while the Gabor filter bank captures thick and thin vessels at all orientations. If the claim holds, screening tools for diabetic retinopathy and similar ocular diseases could obtain vascular maps without any labeled training data or model training.","feed_headline":"Training-free retinal vessel segmentation hits 94.32% accuracy","feed_subtitle":"Morphological plus Gabor filtering maps retinal vessels without any training labels.","key_machinery":"The load-bearing mechanism is the ordering of three complementary operators. White top-hat transformation with a circular structuring element 11 pixels in diameter extracts image details smaller than that width, which by the paper's estimate covers all blood vessels (1–10 pixels) while rejecting larger structures such as optic disc, fovea, and exudates. CLAHE then raises the local contrast of those remaining vessel pixels. A bank of 2D-Gabor wavelets—orientations sampled every 20° from 0° to 180° and radial frequencies from 0.7 to 1.5, with bandwidth constant $K=2.2$—represents thick and thin vessels at different angles; taking the maximum wavelet response per pixel and applying Otsu's global threshold produces the final binary vessel map.","core_discovery":"The central claim is that adding a top-hat morphological transform before CLAHE and Gabor filtering is the decisive improvement: it removes large non-vessel structures from the inverted green channel before contrast enhancement, so the subsequent filter bank responds mainly to vessel pixels. The paper reports that on the DRIVE test set this pipeline yields an average accuracy of 94.32%, sensitivity of 75.03%, specificity of 97.12%, area under the ROC curve of 0.9524, and kappa agreement of 0.7374, with a small accuracy spread ($\\sigma=0.0049$). It further claims that these results compare favorably with established supervised and unsupervised algorithms in accuracy, sensitivity, and kappa agreement, and that the method handles a pathological image containing exudates better than a comparable level-set/region-growing approach.","pith_inferences":["An independent rerun on STARE or CHASE with the same fixed parameters would separate genuine generalizability from DRIVE-specific tuning; the paper does not provide that test.","The top-hat-plus-Gabor response could also serve as input features to a lightweight supervised classifier, potentially combining the preprocessing's robustness with learned thresholds.","Swapping Otsu for local adaptive thresholding should preferentially recover 1–2 pixel thin vessels, at the risk of extra false positives around lesions."],"forward_implications":["A training-free pipeline can reach accuracy within about two percentage points of the strongest supervised classifiers on DRIVE, so unsupervised filtering approaches remain a practical baseline for retinal screening.","Because the structuring element diameter is tied to the maximum vessel width (1–10 pixels on DRIVE), applying the same algorithm to other databases requires only rescaling that one parameter to the new image resolution, a transfer step the paper itself notes.","The small standard deviation of accuracy (0.0049) across the 20 DRIVE test images suggests the method performs consistently on images with different pigmentation, illumination, and pathology.","The authors' suggestion that local adaptive thresholding could replace Otsu implies that thin-vessel detection (1–2 pixel vessels) is the main remaining bottleneck, and a thresholding upgrade is a natural next step."],"supporting_citations":[{"why":"Provides the CLAHE-and-Gabor preprocessing approach this paper reorders, and the comparison result on the pathological DRIVE image.","marker":"[2]"},{"why":"Derives the specific 2D-Gabor wavelet family and the parameter range for the bandwidth constant K used to build the filter bank.","marker":"[24]"},{"why":"Supplies the DRIVE database, the 20 test images, and the first-observer gold standard against which all metrics are computed.","marker":"[9]"},{"why":"Gives the 1–10 pixel vessel-width estimate for DRIVE that sets the 11-pixel structuring element of the top-hat transform.","marker":"[23]"},{"why":"Defines the top-hat morphological operation as highlighting objects smaller than the structuring element, the mechanism that suppresses optic disc, fovea, and exudates.","marker":"[22]"},{"why":"Motivates the family of 2D-Gabor wavelets as a representation of simple-cell receptive fields, justifying the filter bank.","marker":"[25]"},{"why":"Classic oriented matched-filter detector that motivates orientation-tuned vessel filtering and serves as an unsupervised baseline in the comparison.","marker":"[3]"},{"why":"Source of the re-evaluated marker results for baseline methods in the comparison table.","marker":"[29]"}],"fun_headline_variants":["Morphology + Gabor filters map retinal vessels with 94.32% accuracy","Unsupervised vessel segmentation: 94.32% accuracy on DRIVE","Top-hat transform boosts vessel segmentation to 94.32%","Retinal vessel mapping without training hits 94.32% accuracy","Gabor wavelets plus morphology: 94.32% on DRIVE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported results rest on the assumption that the fixed filter settings (11-pixel structuring element, Gabor frequency range, and bandwidth constant) were chosen without peeking at the DRIVE test labels and will carry over to other fundus images unchanged.","fun_headline_variants_meta":{"raw":{"variants":["Morphology + Gabor filters map retinal vessels with 94.32% accuracy","Unsupervised vessel segmentation: 94.32% accuracy on DRIVE","Top-hat transform boosts vessel segmentation to 94.32%","Retinal vessel mapping without training hits 94.32% accuracy","Gabor wavelets plus morphology: 94.32% on DRIVE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1414,"prompt_tokens":877,"completion_tokens":537,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":439}},"tokens_in":493,"tokens_out":537,"duration_ms":5683,"temperature":1.0,"reasoning_tokens":439,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:50:21.466323+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Tune nothing: run the published pipeline with the stated parameters on the STARE and CHASE DB1 databases and compare against their manual ground truths; if accuracy falls materially below the DRIVE figure, or if the structuring element must be re-tuned per database to recover performance, the claim that the algorithm generalizes in its fixed form is refuted.","supporting_citations":[{"cited_title":"Pattern Recognition 47(7), 2437–2446 (2014)","cited_arxiv_id":null,"evidence_quote":"Provides the CLAHE-and-Gabor preprocessing approach this paper reorders, and the comparison result on the pathological DRIVE image."},{"cited_title":"IEEE Transactions on pattern analysis and machine intelligence 18(10), 959–971 (1996)","cited_arxiv_id":null,"evidence_quote":"Derives the specific 2D-Gabor wavelet family and the parameter range for the bandwidth constant K used to build the filter bank."},{"cited_title":"Biomedical Signal Processing and Con- trol 8(1), 71–80 (2013)","cited_arxiv_id":null,"evidence_quote":"Gives the 1–10 pixel vessel-width estimate for DRIVE that sets the 11-pixel structuring element of the top-hat transform."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the top-hat morphological operation as highlighting objects smaller than the structuring element, the mechanism that suppresses optic disc, fovea, and exudates."},{"cited_title":"IEEE Transactions on acoustics, speech, and signal processing 36(7), 1169–1179 (1988)","cited_arxiv_id":null,"evidence_quote":"Motivates the family of 2D-Gabor wavelets as a representation of simple-cell receptive fields, justifying the filter bank."},{"cited_title":"IEEE Transactions on medical imaging 8(3), 263–269 (1989)","cited_arxiv_id":null,"evidence_quote":"Classic oriented matched-filter detector that motivates orientation-tuned vessel filtering and serves as an unsupervised baseline in the comparison."},{"cited_title":"In: Medical Imaging 2004: Image Processing","cited_arxiv_id":null,"evidence_quote":"Source of the re-evaluated marker results for baseline methods in the comparison table."}],"review_version":1}