{"id":"56161b5c-3558-42b5-bb89-e4fc3e562386","arxiv_id":"1909.01558","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Radon-transform and overlapping-window method locates the optic nerve head in fundus and fluorescein angiography images with 91 to 100 percent accuracy on four test sets, at lower computational cost than comparable methods.","lead":"This paper describes a fast method for finding the optic nerve head, the bright round region where the optic nerve leaves the retina, in eye photographs and angiography images. It uses Radon transforms and overlapping image windows to locate the nerve head more quickly than older methods while keeping detection accuracy high.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Step 3 validation cannot uniquely select ONH: any radially symmetric bright structure (exudates, illumination) also minimizes the MSE, and no tie-breaking rule is specified.","rationale":"The central claim requires both high accuracy and computational efficiency. The efficiency comparison is weakened by cross-hardware timing and the approximate complexity table, but the operation-count analysis makes the speed advantage plausible. The more serious gap is in the detection-validation step: the criterion 'round object gives identical Radon profiles' is satisfied by exudates and uneven illumination, exactly the failure modes the paper acknowledges. The reader's weakest assumption identified this. This review sharpens the concern: as written, the algorithm does not specify how to choose among multiple candidate sub-images with comparable or identical MSE, so the reported detection rates are not fully reproducible from the description. This supports a conditional verdict rather than full acceptance: the method is plausible and supported by reported results on public datasets, but it needs a clarified selection rule, code, or a sensitivity analysis of tie-breaking before the claimed robustness is established. The reader's CONDITIONAL verdict remains appropriate; no change to the verdict is needed, but the condition should explicitly include resolving this underspecification.","tokens_in":10078,"tokens_out":7913,"duration_ms":86761,"concrete_test":"Re-implement the algorithm strictly from Section III on all 40 DRIVE and 81 STARE images; in each image, record the set of candidate sub-images satisfying the 0.9 threshold and their Eq. (4) scores. If in any image two or more candidates attain the same minimal MSE (within, say, 0.1% relative), re-run detection under two tie-breaking rules (largest Radon peak; top-left-most sub-image). If the reported 100%/96.3% accuracy changes across tie-breaks, the missing ranking rule is load-bearing; if not, it is cosmetic.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III.B Step 3's validation criterion (Eq. (4)) is not discriminating. The property 'for a round object RT provides the same profile along all directions' holds for any radially symmetric bright pattern, including the exudates and uneven round-shaped illumination that the paper itself cites as failure cases in Figs. 13(h), 14(d), and 15(d). The text states that the sub-image minimizing the MSE among its projections is selected, but it never specifies how a final global choice is made when multiple candidate sub-images (e.g., one containing the ONH, another containing a bright round exudate) have comparable or identical MSE values. Eq. (4) is computed per reference column r within one sub-image; no cross-sub-image ranking, no tolerance, and no tie-breaking rule is given. Consequently, the reported detection rates cannot be reproduced from the algorithm description alone, and the claimed robustness to bright round lesions is not established. This is load-bearing because the entire method reduces to selecting the 'roundest bright' sub-image, and that criterion is not unique to the ONH.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a method for optic nerve head detection in retinal fundus images, covering both color fundus and fluorescein angiography images. The pipeline consists of: masking the fundus region; partitioning the image into overlapping n×n sub-images; applying the Radon transform to the blue channel of each masked sub-image; selecting candidate windows whose Radon-space peaks exceed a threshold (0.9 of the sub-image maximum); validating candidates by comparing Radon projections across angles via mean-square error; and taking the center of the validated sub-image as the ONH center. Reported results are 100% (40/40 DRIVE), 96.3% (78/81 STARE), 97.5% (117/120 MUMS color), and 91.3% (110/120 MUMS FA), with 4.1 seconds per STARE image. The authors claim the method is computationally more efficient than existing algorithms with comparable detection rates and does not require vessel segmentation.","tokens_in":10280,"tokens_out":6583,"duration_ms":68538,"significance":"If the claims hold, the paper offers a practically useful, lightweight ONH detector: it avoids vessel segmentation, uses public benchmarks (DRIVE, STARE) plus a clinical local database, reports a complexity model in Section III.C, and candidly identifies failure cases such as uneven round-shaped illumination and darker-than-surrounding ONH. The underlying idea of using Radon-projection consistency to capture disk-like bright structures is interesting and the reported processing times are attractive for screening applications. However, the validation step in Section III.B is specified too loosely to be reproduced or to uniquely identify the ONH, and the parameter/evaluation protocol does not rule out in-sample tuning. These issues do not necessarily invalidate the empirical results, but they are load-bearing for the paper's central claim of robust, reproducible detection.","major_comments":[{"comment":"The validation step is not specified sufficiently to support the claimed discrimination or to allow reproduction. Eq. (4) defines an MSE per reference column r within the Radon matrix of one sub-image, but the text never states how candidate sub-images are ranked against each other, how a global final choice is made when several sub-images have comparable MSE values, or what tie-breaking rule is used. This matters because the 'same profile along all directions' property is shared by any centered disk-like bright structure, and every sub-image is multiplied by a circular mask, so the mask alone produces identical projections along all directions. The paper's own failures (Figs. 13(h), 14(d), 15(d)) are attributed to uneven round-shaped illumination, which is precisely the non-uniqueness described. The authors should provide the complete selection rule—including threshold application, cross-sub-image ranking, and tie handling—and ideally add a quantitative experiment showing that bright round exudates and synthetic illumination artifacts do not pass the validation criterion.","section":"Section III.B Step 3, Eq. (4), Figs. 13(h), 14(d), 15(d)"},{"comment":"The four free parameters—window size n, overlap step s, Radon peak threshold 0.9, and number of projection angles a—are reported as chosen values, with n set per database, but the manuscript does not state how these values were selected or whether the evaluation datasets were also used for tuning. Since the central claim is robustness across databases and lesion types, the reported detection rates could be optimistic in-sample estimates. Please specify the parameter-selection procedure (e.g., fixed a priori, cross-validated, or tuned on a separate subset), report confidence intervals or bootstrap estimates for the detection rates, and state explicitly which, if any, parameter values were adjusted after looking at the test images.","section":"Section III.B Steps 1-3 and Section IV (Table I)"},{"comment":"The computational-efficiency claim is not established at the level of rigor used for the accuracy claims. The operation-count formula NM(1+9as^2)+3anc is asserted without a step-by-step derivation, and the runtime comparison in Section IV mixes different hardware generations (Core i5, Centrino, Pentium IV, Core2Duo), different implementations, and different image resolutions. This does not support the headline conclusion that the method is the fastest among methods with comparable detection rates. Please derive the operation count from the actual algorithm steps (windows, Radon calls per window, candidate validation), and, if possible, provide a runtime comparison on identical hardware or a normalized hardware-independent measure.","section":"Section III.C and Table II"}],"minor_comments":[{"comment":"The text refers to '2MHz Intel Core i5', '2MHz Intel Centrino 1.7', and '2MHz Intel Pentium IV'; these should be 2 GHz, and the '2.66 Intel Core2Duo' should read 2.66 GHz.","section":"Section IV"},{"comment":"The sentence reporting 'the average distance (for the 69 successful images) ... and for the 9 successful images' does not reconcile with the earlier statement of 78 correct detections out of 81 STARE images; please clarify the group sizes and what each average refers to.","section":"Section IV"},{"comment":"The statement that 'as step is increased, the computation time increases exponentially' is inconsistent with the quadratic dependence NM(1+9as^2); the word should be 'quadratically' unless the formula is corrected.","section":"Section III.B/III.C"},{"comment":"The symbols M and N are used both for the input image dimensions in Section III.C and for the dimensions of the Radon matrix in Eq. (4); please use distinct notation to avoid confusion.","section":"Section III.B Step 3 and Section III.C"},{"comment":"The expression '9MNs2a' is missing a superscript; it should be typeset as 9 M N s^2 a for consistency with the total NM(1+9as^2).","section":"Section III.C"},{"comment":"The integration variable z in Eq. (1) is not explicitly defined; please state the parametrization of the line integral more carefully.","section":"Section III.A.3, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a previously published journal article (Biomedical Signal Processing and Control, 2014) posted to arXiv; if this submission is being considered for a different venue, the editor should verify the novelty and overlap policies. In addition, the MUMS-DB dataset is not public, so independent verification of the reported 97.5% and 91.3% rates is not possible; reproducibility therefore rests on the DRIVE and STARE numbers. The main technical blocker is the underspecified validation rule in Section III.B Step 3, which should be addressed in a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a competent, clearly written algorithm for optic nerve head localization using the Radon transform and multi-overlapping windows. The reported detection rates on DRIVE and STARE are respectable, and the runtime analysis is a real strength. The main soft spot is the validation step: the MSE criterion in Eq. (4) does not uniquely separate the ONH from other radially symmetric bright structures, and the paper never specifies a tie-breaking rule when multiple candidate sub-images have comparable scores.\n\nThe combination of Radon transform with overlapping windows and projection-profile MSE validation is genuinely not in the cited literature. The algorithm description is detailed enough to reimplement, apart from the ambiguous final selection. Using DRIVE and STARE is good practice, and the failure cases are shown and honestly attributed to uneven round-shaped illumination. The paper also states its own limitation for darker-than-surrounding ONH. That transparency earns credit.\n\nThe stress-test concern is legitimate. Eq. (4) measures how much a sub-image's Radon projections differ from one reference column, but the text does not explain how the global choice is made across sub-images. Any round bright object—exudate, illumination artifact—will also produce similar projections. The paper's own failures confirm this: the method fails exactly when a bright round non-ONH structure appears. The claim of robustness to bright round lesions therefore rests on datasets where such structures may be rare, not on a discriminative test. This is a real gap, though not a fatal one: in practice the global MSE minimum often lands on the ONH because it is generally the largest and brightest round object.\n\nThe other soft spots are more minor. No confidence intervals or statistical tests are given. The window size, overlap step, and 0.9 threshold appear hand-set, with no sensitivity analysis. The timing comparison uses different CPUs and implementations, so the 4.1-second figure is only suggestive, not a clean benchmark. The MUMS-DB data are private and no code is provided, which limits reproducibility.\n\nOverall, this is a legitimate methods paper. The central idea is plausible, the presentation is honest, and the results are likely real. It is incremental rather than field-changing, but it deserves serious referee time. If I were handling it, I would send it to peer review and ask for a defined cross-sub-image selection rule, a sensitivity analysis, and a fairer timing comparison. For someone working on retinal landmark detection, it is a useful baseline and worth citing.\n\nMy recommendation: engage with it as a solid, modest contribution, but do not treat the efficiency or robustness claims as fully established until the validation step is tightened.","headline":"Solid, clearly described ONH detector with respectable accuracy, but the Step 3 validation criterion is not as discriminative as claimed and the empirical comparison has unaddressed confounds.","tokens_in":10784,"tokens_out":1956,"would_cite":false,"duration_ms":22502,"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":"Radon transform plus window voting finds the optic nerve head in about 4 seconds per image, with 96–100% localization accuracy on three retinal image sets.","keywords":["optic nerve head detection","Radon transform","fundus image","fluorescein angiography","diabetic retinopathy","multi-overlapping windows","computer-aided diagnosis","retinal image analysis"],"falsifier":"Run the algorithm on a set of fundus images in which a bright circular exudate or round illumination artifact lies away from the ONH and ask whether the reported center lands on the artifact; if even one such image is systematically selected over the ONH, the roundness-MSE criterion is not sufficient and the claimed accuracy depends on dataset-specific window sizes and thresholds.","tokens_in":9898,"feed_emoji":"👁️","tokens_out":5726,"duration_ms":54815,"temperature":0.7,"pith_summary":"The paper claims that a fast optic nerve head detector can be built from two simple ingredients: the Radon transform and overlapping image windows. Because the optic nerve head is a bright, roughly circular disk, its Radon projections are nearly identical in all directions; the detector slides a window over the fundus image, keeps windows whose Radon peaks are strong, and picks the window with the least disagreement among projection profiles. On four test sets the reported hit rates are 100 percent (40 images), 96.3 percent (81), 97.5 percent (120 color), and 91.3 percent (120 angiography), and a STARE image is processed in about 4.1 seconds. The point of the speed is practical: if true, the method could serve as a cheap first step in retinal screening pipelines.","feed_headline":"Optic nerve head found in 4 seconds by Radon profile match","feed_subtitle":"A roundness test on Radon projections hits 96–100% accuracy on three retinal image sets without tracing a single vessel.","key_machinery":"The load-bearing object is the Radon transform paired with an MSE-over-projections validation rule. The Radon transform is the set of line integrals of image intensity taken along parallel beams at many angles; a bright circular disk produces a prominent, nearly angle-independent peak in this space. The paper adds a circular mask to each sub-image to remove the artificial diagonal bias of square windows, then uses the mean square error among the projection profiles as a similarity score: the sub-image whose projections agree best is declared to contain the ONH. All of the reported speed comes from keeping this machinery local, with small windows and an overlap step of 4, and from avoiding vessel segmentation entirely.","core_discovery":"The central discovery is that optic nerve head localization can be posed as a roundness test on Radon projections rather than as vessel tracking or template matching. After masking out the non-fundus background and the window corners, the algorithm applies the Radon transform to each $n\\times n$ sliding window, with window size set by the expected ONH diameter ($n=79$ for DRIVE, $n=130$ for STARE, $n=313$ for the high-resolution angiography set). In color images it uses the blue channel because yellowish structures like the ONH stand out there. A sub-image is kept as a candidate when its peak Radon projection exceeds 0.9 of the largest projection in that sub-image, and the final choice is the candidate that minimizes the mean square error among its projection profiles at different angles, relying on the property that a round object gives the same profile in every direction. The center of that winning window is reported as the ONH center, and the paper reports this procedure lands within 60 pixels of a manual reference in all 40 DRIVE images, 78 of 81 STARE images, 117 of 120 color fundus images, and 110 of 120 fluorescein angiography images.","pith_inferences":["The same roundness-of-projection test may transfer to detecting other circular retinal structures, such as the foveal avascular zone, though its contrast behavior differs and would need separate validation.","If the method were inverted to look for windows whose projections disagree most, it might act as a detector of non-circular bright lesions such as exudates, a hypothesis the paper does not test.","Because the MSE criterion is purely geometric, a straightforward stress test is to synthesize images containing a bright disk that is not the ONH; the paper's claim implies the algorithm would still select it.","The failure mode on darker-than-surrounding ONH suggests a practical extension: run the same pipeline on inverted intensities and combine the two scores, which the paper does not address."],"forward_implications":["Because the method avoids vessel segmentation, it can serve as an ONH locator in images where vessel extraction is costly or unreliable.","At roughly 4.1 seconds per STARE-size image, the detector is fast enough to run as a pre-filter before slower segmentation or classification steps.","Halving the image resolution preserved accuracy (100% DRIVE, 96.3% STARE, 95.9% for the color set) while cutting runtime by more than 12 times, so the method scales to high-resolution cameras.","The accuracy standard used throughout is placement within 60 pixels of the manually marked center, so the reported percentages are for localization, not boundary segmentation.","The 0.9 peak threshold and the overlap step of 4 are fixed in the paper, while the window size $n$ is chosen from the expected ONH diameter, making the method portable across image resolutions."],"supporting_citations":[{"why":"Supplies the anatomical characterization of the ONH as a bright disk with rapid intensity variation, motivating the round-bright object model.","marker":"[3]"},{"why":"Defines the 60-pixel success criterion used for every reported accuracy and provides a comparison baseline with 97.5% detection.","marker":"[10]"},{"why":"Provides a high-accuracy baseline (98.8%) whose runtime of about 3.5 minutes the new method is compared against.","marker":"[11]"},{"why":"Is the faster but lower-accuracy baseline (0.46 seconds, 92.6%) used to position the speed-accuracy trade-off of the introduced method.","marker":"[33]"},{"why":"Supplies the DRIVE image set on which the 100% detection result is reported.","marker":"[35]"},{"why":"Supplies the STARE image set on which the 96.3% detection result and the 4.1-second processing time are reported.","marker":"[36]"}],"fun_headline_variants":["Radon trick finds optic nerve head without tracing vessels","Optic nerve head found by Radon roundness test, no vessels needed","Radon profile match finds optic nerve head faster than vessel tracing","Roundness on Radon projections locates optic nerve head quickly","Optic nerve head detection: roundness beats vessel tracking"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the ONH is a round, bright object whose Radon profiles match in all directions, so any equally round, bright structure, such as an exudate or uneven round-shaped illumination, can win the MSE vote; the paper attributes its failures to the latter.","fun_headline_variants_meta":{"raw":{"variants":["Radon trick finds optic nerve head without tracing vessels","Optic nerve head found by Radon roundness test, no vessels needed","Radon profile match finds optic nerve head faster than vessel tracing","Roundness on Radon projections locates optic nerve head quickly","Optic nerve head detection: roundness beats vessel tracking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00085,"raw_usage":{"total_tokens":3658,"prompt_tokens":868,"completion_tokens":2790,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":2704}},"tokens_in":484,"tokens_out":2790,"duration_ms":20990,"temperature":1.0,"reasoning_tokens":2704,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:13:30.542395+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the algorithm on a set of fundus images in which a bright circular exudate or round illumination artifact lies away from the ONH and ask whether the reported center lands on the artifact; if even one such image is systematically selected over the ONH, the roundness-MSE criterion is not sufficient and the claimed accuracy depends on dataset-specific window sizes and thresholds.","supporting_citations":[{"cited_title":"Automated localisation of the optic disc, fovea, and retinal blood vessels from digital colour fundus images,","cited_arxiv_id":null,"evidence_quote":"Supplies the anatomical characterization of the ONH as a bright disk with rapid intensity variation, motivating the round-bright object model."},{"cited_title":"Detection of optic disc in retinal images by means of a geometrical model of vessel structure,","cited_arxiv_id":null,"evidence_quote":"Defines the 60-pixel success criterion used for every reported accuracy and provides a comparison baseline with 97.5% detection."},{"cited_title":"Optic disc detection from normalized digital fundus images by means of a vessels’ direction matched filter,","cited_arxiv_id":null,"evidence_quote":"Provides a high-accuracy baseline (98.8%) whose runtime of about 3.5 minutes the new method is compared against."},{"cited_title":"Fast localization of the optic disc using projection of image features,","cited_arxiv_id":null,"evidence_quote":"Is the faster but lower-accuracy baseline (0.46 seconds, 92.6%) used to position the speed-accuracy trade-off of the introduced method."},{"cited_title":"Ridge -based vessel segmenta tion in color images of the retina,","cited_arxiv_id":null,"evidence_quote":"Supplies the DRIVE image set on which the 100% detection result is reported."},{"cited_title":"Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels,","cited_arxiv_id":null,"evidence_quote":"Supplies the STARE image set on which the 96.3% detection result and the 4.1-second processing time are reported."}],"review_version":1}