{"id":"b7315805-30c3-4279-8c1a-ca6aee65883d","arxiv_id":"1909.01557","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"A Radon-transform and overlapping-window method detects retinal microaneurysms in fluorescein angiography, reporting 94% sensitivity and 75% specificity on a local database and 100% sensitivity and 70% specificity on a second.","lead":"This paper presents a computer vision method for detecting microaneurysms, the earliest visible sign of diabetic retinopathy, in fluorescein angiography images of the retina. It uses Radon transforms and overlapping image windows to find small bright circular spots after masking the optic nerve head and blood vessels.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported specificity (75%/70%) rests on 20 and 10 normal images and per-database empirical tuning; small-sample CIs and 5 FP/image make the image-level DR detection claim fragile.","rationale":"The strongest claim is the abstract's sensitivity/specificity. The paper's experimental design is better than many CAD papers: separate training/test partitions (Section 4.2), lesion-based and image-based analyses, and comparison against human readers. The main image-based confusion matrices in Tables 2 and 3 are internally consistent (61+4+15+5=85; 30+0+7+3=40). The method is a plausible extension of earlier Radon-transform vessel and ONH detection work, and the reported numbers are not absurd. However, the central accuracy claim is not yet secure for three reasons. First, the normal-sample sizes are small, so the specificity estimates are noisy and one of them is not statistically distinguishable from chance. Second, the many empirically tuned parameters are database-specific, and the paper does not supply code or public data to check whether the pipeline generalizes beyond the two local datasets; the paper even states its data distribution is biased (Section 3). Third, the ROC database results are internally inconsistent (text reports 122 TP, 13 FN, 91%; Table 6 reports 125 TP, 10 FN), which undercuts confidence in the care with which all metrics were computed. These concerns do not prove the claim wrong, but they mean the reported 94/75 and 100/70 values should be read as pilot results on two local datasets. The reader's CONDITIONAL verdict is appropriate; my stress-test confirms that the claim needs additional validation but does not move the verdict.","tokens_in":14577,"tokens_out":8842,"duration_ms":92863,"concrete_test":"Compute exact binomial (Clopper-Pearson) 95% confidence intervals for the specificity values reported in Tables 2 and 3 (15/20 and 7/10). If the 7/10 interval's lower bound is below 0.5, then the claimed 70% specificity on the Second Local Database is not statistically established at the 95% level, and an independent validation set with at least 50 normal images would be required to support the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—image-level DR detection at 94%/75% and 100%/70% sensitivity/specificity—depends on two weakly supported conditions. First, the specificity estimates come from only 20 normal MUMS images and 10 normal 2nd-DB images (Section 4.4.1). With 15/20 and 7/10 correct, the exact binomial 95% confidence intervals are roughly 53–89% and 39–90%; the 2nd-DB interval includes 50%, so 70% specificity is not statistically distinguishable from chance at the 95% level. Second, the reported numbers are produced after per-database empirical selection of at least 11 parameters (window sizes n in Sections 3.3, 3.5.1, and 3.5.2; step values of 4 and 5; alpha 0.75 and 0.6; top-hat disk d=25; averaging window 75). The distinguishing criterion for MAs versus vessel endpoints and bifurcations is qualitative—'minimum deviation' between Radon profiles (Section 3.5.2)—with no stated threshold or independent validation. The consequence is visible in the lesion-based results: FP per image is 5 (Section 4.4.2), an order of magnitude above the human readers (0.41 and 2.58 in Table 1), so a normal eye can cross the 'more than 5 MAs' threshold used for image-level DR diagnosis (Section 4.4.1). If the per-database tuning is removed or the normal sample is enlarged, the headline specificity values are unlikely to hold at the same level. This does not make the method unsound, but it makes the central accuracy claim conditional on unvalidated choices.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an automated pipeline for detecting microaneurysms (MAs) in fluorescein angiography (FA) fundus images, using a Radon transform combined with multi-overlapping windows. The pipeline first detects and masks the optic nerve head and the vascular tree, then identifies MAs as circular bright patterns whose Radon profiles are consistent across projection angles. The authors evaluate the method on two local databases (MUMS-DB and Second Local Database) and a subset of the public ROC database, reporting image-level DR detection with sensitivities/specificities of 94%/75% and 100%/70% for the two local databases, and lesion-level sensitivities of 92%, 95%, and 91% for MUMS, 2nd-DB, and ROC, respectively. The paper also compares results with human readers and prior methods, and reports processing times.","tokens_in":14978,"tokens_out":3778,"duration_ms":38861,"significance":"If the reported performance holds under independent validation, the work would demonstrate a useful complementary approach for MA detection in FA images, a clinically important task for diabetic retinopathy screening. The use of FA as a gold-standard modality, the inclusion of three databases with different cameras and resolutions, and the reporting of both image-level and lesion-level analyses are strengths. The Radon-transform-based detection of circular patterns combined with vessel/ONH masking is a reasonable and somewhat distinctive idea. However, the significance is currently limited by the small number of normal images used for specificity estimation, the per-database empirical tuning of many parameters, and unresolved numeric inconsistencies in the reported results.","major_comments":[{"comment":"The ROC database lesion-level results are internally inconsistent. The text states that among 135 MAs the method found 122 true positives and missed 13, yielding a sensitivity of 91%; Table 6 reports 125 true positives and 10 false negatives. These two sets of numbers cannot both be correct, and the reported sensitivity is consistent only with the 122/135 values. Additionally, the text says the false-positive count is 41 and the FP per image is 5, but 41 divided by 22 (or even 20) is not 5. These inconsistencies must be resolved because the lesion-level results are a central part of the evaluation.","section":"Section 4.4.2 and Table 6"},{"comment":"The image-level specificity estimates are based on very small numbers of normal images: 20 in MUMS-DB and 10 in 2nd-DB. With 15/20 true negatives the exact binomial 95% confidence interval is roughly 53-89%, and with 7/10 it is roughly 39-90%; the latter interval includes 50%, so the reported 70% specificity is not statistically distinguishable from chance at the 95% level. The paper should either provide confidence intervals for all reported sensitivities and specificities or enlarge the normal-image sample before claiming specific performance levels.","section":"Section 4.4.1, Tables 2 and 3"},{"comment":"Many key parameters are chosen empirically and per database: window sizes for ONH, vessel, and MA detection (n=309/38/114, n=30/15/17, n=18/10/12), sliding steps (4 or 5), the vessel width parameter alpha (0.75 or 0.6), top-hat disk diameter (d=25), averaging filter size (75), and the DR diagnosis threshold. Several of these are set using object-size statistics (e.g., 'maximum diameter of the biggest MA in pixel') that are computed from each database, including the test set. This means the reported performance is not an independent test of a fixed algorithm but reflects database-specific tuning. The authors should validate with cross-validation within each database or with parameters fixed across databases, and report how performance varies with parameter choices.","section":"Sections 3.3, 3.5.1, 3.5.2 and 4.2"},{"comment":"The average number of false positives per image is reported as 5 for both local databases, which is exactly the threshold used for image-level DR diagnosis ('more than 5 MAs' implies DR). This places the image-level classification very close to the operating point expected for a normal image, so the reported specificity of 75%/70% is fragile: any increase in the FP rate would rapidly convert normal images into false positives. The paper should report the distribution of FP counts across images and show how the diagnosis threshold is positioned with respect to that distribution.","section":"Sections 4.4.1 and 4.4.2"},{"comment":"The MA validation criterion is described qualitatively: profiles related to projections have 'minimum deviation' with each other, and a candidate is accepted if its Radon peaks are 'approximately similar' across projection angles. No explicit metric or threshold is defined for this deviation, and the intensity threshold is also only described as 'predefined.' This lack of a precise decision rule makes the method irreproducible and prevents independent verification of the reported lesion-level accuracy. The authors should state the exact formula and threshold used to accept or reject a candidate MA.","section":"Section 3.5.2"}],"minor_comments":[{"comment":"The equations defining the Radon transform appear to be missing from the manuscript; only the equation numbers are visible. This needs to be fixed so that the mathematical basis is complete.","section":"Section 3.2"},{"comment":"The number of ROC images is inconsistent: the introduction and Section 3 state 22 images, while Section 4.2 says the test set included 20 images from the ROC database. Please reconcile these counts.","section":"Section 4.2"},{"comment":"In the third paragraph, the text says 'the optimum of step was selected 5 pixels for both databases empirically,' but the method is applied to three databases. Please clarify which databases are meant.","section":"Section 3.5.1"},{"comment":"The paper alternates between 'pixel based' and 'lesion based' analysis; the latter term is used for the actual evaluation. Please define the terminology consistently, since lesion-based analysis is not the same as pixel-based analysis.","section":"Sections 4.1 and 4.4.2"},{"comment":"The comparison with previous work would be stronger if the evaluation protocol (e.g., whether the same images are used for training and testing, and how the 5-FP-per-image threshold was chosen) were explicitly matched across studies. Currently, the comparison with Cree et al. and Spencer et al. is qualitative and does not account for differences in image sets and ground-truth definitions.","section":"Section 4.5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a reprint of a paper originally published in Pattern Recognition in 2013, submitted to arXiv in 2019. If this is a resubmission of already-published work, the editor should consider whether the current version provides sufficient novelty or contextualization relative to the earlier publication. The technical concerns above—especially the numeric inconsistencies and the per-database parameter tuning—should be addressed before the paper is considered further."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is a competent classical image-processing paper for detecting microaneurysms in fluorescein angiography, and it does something slightly new by applying the Radon transform with multi-overlapping windows in that modality. The headline numbers (94%/75% and 100%/70% sensitivity/specificity at image level) are plausible but softer than they look.\n\nWhat is new: the specific pipeline—ONH masking, vessel removal, then MA detection via Radon profiles with circularity validation—is not in the cited literature. The earlier Radon work [29] used a different operator on color fundus images. The paper is also honest about its data limitations and does useful comparisons against expert readers and prior FA methods.\n\nThe soft spots are real. The specificity estimates come from 20 normal images in one database and 10 in the other; 70% specificity on the second database has a 95% confidence interval that includes 50%, so it is not statistically distinguishable from chance. There are eleven-plus free parameters selected empirically per database, and the MA validation criterion—'minimum deviation' between Radon profiles—is qualitative. The lesion-based false-positive rate is 5 per image, an order of magnitude above the human readers (0.41 and 2.58), which puts the image-level decision rule ('more than 5 MAs = DR') uncomfortably close to the FP floor. There is also a small inconsistency in the ROC database lesion results (text says 122 TP / 13 FN; Table 6 says 125 / 10). That does not sink the paper but needs a fix.\n\nThe stress-test note gets the central point right: the reported accuracy is conditional on per-database tuning. If the parameters transfer without retuning, the method may hold, but there is no evidence either way. The authors admit the data is biased, which is in their favor.\n\nWho is this for? Anyone working on FA image analysis or classical (pre-deep-learning) CAD pipelines. It deserves a serious referee because the method is coherent and evaluated on real data, but the specificity claims need confidence intervals or a larger normal sample, and the parameter selection should be handled as part of the training procedure or validated with fixed parameters across databases.\n\nRecommendation: send to peer review with a request for major revision on the statistical reporting and the consistency fix. Not a desk reject.\n\nBest,","headline":"Solid classical FA microaneurysm pipeline, but headline specificity is statistically fragile and needs a consistency correction.","tokens_in":15529,"tokens_out":2869,"would_cite":false,"duration_ms":27561,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92C55","68U10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Scanning overlapping windows in fluorescein angiography with the Radon transform can flag diabetic retinopathy by detecting microaneurysms, with 94–100% sensitivity and 70–75% specificity.","keywords":["diabetic retinopathy","microaneurysm detection","fluorescein angiography","Radon transform","fundus image analysis","computer aided diagnosis","image segmentation","screening"],"falsifier":"Take a new set of fluorescein angiography images with expert-annotated microaneurysm locations, implement the pipeline with the paper's stated window sizes, step values, and threshold constants, and compare detected candidates against the annotations; the central claim is weakened if many true microaneurysms located next to vessels are masked as vessels, or if laser scars and abnormal dilated capillaries are counted as microaneurysms at the reported false-positive rate.","tokens_in":1535,"feed_emoji":"👁️","tokens_out":1673,"duration_ms":97924,"temperature":0.7,"pith_summary":"The paper sets out to show that microaneurysms — the earliest visible retinal sign of diabetic retinopathy — can be detected automatically in fluorescein angiography images by applying the Radon transform to small overlapping windows and looking for circular patterns that give the same peak profile at every projection angle. The authors argue this is a useful complement to morphology-based detection because the same transform also supplies numerical data such as position, size, and orientation. In testing, image-level diagnosis of diabetic retinopathy reached a sensitivity and specificity of 94% and 75% on a 120-image local database and 100% and 70% on a 50-image second database, with lesion-level sensitivities of 92–95% on these two sets. The clinical motivation is that automated, deterministic pre-screening could catch the first sign of diabetic retinopathy while cutting the number of normal images a specialist has to read.","feed_headline":"Radon transform spots diabetic eye lesions at 94-100% sensitivity","feed_subtitle":"Sliding-window Radon scanning flags early diabetic retinopathy so specialists can focus on ambiguous scans.","key_machinery":"The Radon transform, defined as the integral of image intensity over straight lines, $\\\\int f(x,y)\\\\,\\\\delta(x\\\\cos\\\\theta+y\\\\sin\\\\theta-s)\\\\,dx\\\\,dy$, is the central object. Applied to each small overlapping sub-image, a round microaneurysm produces a peak whose profile is nearly identical in every projection angle, while elongated vessels produce peaks that vary strongly with angle; this contrast is the detection criterion. The multi-overlapping window strategy ensures lesions near sub-image borders are not missed. The same transform is reused earlier to find the optic nerve head (by roundness) and to segment the vascular tree (by line peaks), and those masks remove the two major sources of false positives before microaneurysm detection begins.","core_discovery":"The discovery is that a microaneurysm can be recognised in the Radon domain as a peak that appears in every projection column with nearly Gaussian, mutually similar profiles, because a round object yields essentially identical projections at all angles. After masking the optic nerve head and the vascular tree — both found with the same Radon-transform machinery on overlapping sub-images — the authors threshold remaining candidates by intensity, size under 125 μm, and roundness of the Radon profiles. On image-level detection of diabetic retinopathy they report sensitivity and specificity of 94% and 75% for the first local database and 100% and 70% for the second; on lesion-level detection the sensitivity is 92%, 95%, and 91% across the two local databases and the public challenge subset, indicating the method can also support microaneurysm counting for follow-up. The paper presents the system as a complementary diagnostic aid rather than a full screening package, and notes that adding image registration would make it useful for tracking disease progression.","pith_inferences":["If the Radon-profile consistency criterion transfers to other imaging hardware, the same circular-spot detector could be retrained for other round retinal lesions, such as hard exudates or drusen in color fundus photography; the paper only tests fluorescein angiography images.","The roughly 11-minute per-image runtime suggests that the pipeline, as published, is a diagnostic aid rather than a population-level screening tool; deployment would require optimising the vessel-detection step, which dominates the runtime.","Because specificity (70–75%) trails sensitivity (94–100%), an automated triage built on this method would still refer a substantial fraction of normal patients to human readers; the acceptable trade-off would depend on the cost of false alarms versus missed disease."],"forward_implications":["Automated fluorescein-angiography reading could flag the earliest stage of diabetic retinopathy with image-level sensitivity of 94–100%, meaning few diseased eyes would be sent back as healthy.","A deterministic pipeline of this kind gives repeatable readings that do not degrade with reader fatigue, so it could sit in front of a human specialist and separate clearly normal images.","The paper's ablation shows the system depends on vessel masking: without it, specificity drops from 75% to 61% on the first local database and from 70% to 56% on the second, so improvements in vessel segmentation should translate directly into better microaneurysm screening.","With an added image-registration step, the same microaneurysm detection could support longitudinal counting of microaneurysms for treatment follow-up."],"supporting_citations":[{"why":"Supplies the optic nerve head detection routine the paper masks before microaneurysm detection.","marker":"[30]"},{"why":"Describes the Radon-transform vessel detection method reused here to mask the vascular tree.","marker":"[31]"},{"why":"Supplies the public challenge dataset and the FROC evaluation convention used for lesion-level analysis.","marker":"[7]"},{"why":"Supplies the top-hat preprocessing and 2D-Gaussian shape model used to validate microaneurysm candidates.","marker":"[23]"},{"why":"Gives an earlier fully automated microaneurysm detection system used as a performance baseline in the discussion.","marker":"[24]"},{"why":"Provides the morphology-based microaneurysm segmentation strategy and its 82%/86% sensitivity/specificity used for comparison.","marker":"[21]"},{"why":"Links Radon-based detection to prior microaneurysm work on color fundus images.","marker":"[29]"}],"fun_headline_variants":["Radon-transform scans catch early diabetic eye lesions at 94-100%","Automated microaneurysm detection: Radon windows flag DR at 94% sensitivity","Complementary Radon method spots microaneurysms for diabetic retinopathy","Sliding-window Radon finds microaneurysms in fluorescein angiography","94-100% sensitivity: Radon-based detection of early diabetic retinopathy"],"cache_read_input_tokens":17536,"weakest_assumption_plain":"The method relies on the assumption that a true microaneurysm looks round at every angle in the Radon domain while vessel endings, branch points, noise, and laser scars do not, and that the empirically chosen window sizes, step values, and thresholds keep this separation working on new image collections.","fun_headline_variants_meta":{"raw":{"variants":["Radon-transform scans catch early diabetic eye lesions at 94-100%","Automated microaneurysm detection: Radon windows flag DR at 94% sensitivity","Complementary Radon method spots microaneurysms for diabetic retinopathy","Sliding-window Radon finds microaneurysms in fluorescein angiography","94-100% sensitivity: Radon-based detection of early diabetic retinopathy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1458,"prompt_tokens":1014,"completion_tokens":444,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":338}},"tokens_in":630,"tokens_out":444,"duration_ms":4055,"temperature":1.0,"reasoning_tokens":338,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:13:40.972096+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a new set of fluorescein angiography images with expert-annotated microaneurysm locations, implement the pipeline with the paper's stated window sizes, step values, and threshold constants, and compare detected candidates against the annotations; the central claim is weakened if many true microaneurysms located next to vessels are masked as vessels, or if laser scars and abnormal dilated capillaries are counted as microaneurysms at the reported false-positive rate.","supporting_citations":[{"cited_title":"Tavakoli, M.H","cited_arxiv_id":null,"evidence_quote":"Supplies the optic nerve head detection routine the paper masks before microaneurysm detection."},{"cited_title":"Tavakoli, A.R","cited_arxiv_id":null,"evidence_quote":"Describes the Radon-transform vessel detection method reused here to mask the vascular tree."},{"cited_title":"Niemeijer, B","cited_arxiv_id":null,"evidence_quote":"Supplies the public challenge dataset and the FROC evaluation convention used for lesion-level analysis."},{"cited_title":"Walter, J","cited_arxiv_id":null,"evidence_quote":"Supplies the top-hat preprocessing and 2D-Gaussian shape model used to validate microaneurysm candidates."},{"cited_title":"Cree, J.A","cited_arxiv_id":null,"evidence_quote":"Gives an earlier fully automated microaneurysm detection system used as a performance baseline in the discussion."},{"cited_title":"Spencer, J.A","cited_arxiv_id":null,"evidence_quote":"Provides the morphology-based microaneurysm segmentation strategy and its 82%/86% sensitivity/specificity used for comparison."},{"cited_title":"Giancardoa, F","cited_arxiv_id":null,"evidence_quote":"Links Radon-based detection to prior microaneurysm work on color fundus images."}],"review_version":1}