{"id":"4b1b1ec6-e15d-4c81-ae5b-ebac32dcfd8b","arxiv_id":"1909.00331","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An automated deep neural network pipeline can localize the scleral spur and segment anterior segment structures in OCT images with performance close to human experts.","lead":"This paper shows that a deep learning program can find and measure key structures in eye scans as accurately as an experienced eye doctor. The goal is to make angle-closure glaucoma screening faster and more objective, reducing human error in OCT scans.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central performance claims rest on agreement with the very human labels the network was trained to reproduce; without an independent-grader or consensus reference, 'as accurate as an ophthalmologist' may only mean 'matches the training annotators'.","rationale":"The reader's weakest assumption is the same as the primary load-bearing concern: human scleral-spur annotations and the single-annotator segmentation labels are treated as ground truth, but the paper explicitly acknowledges the absence of an absolute ground truth. I agree with the reader's conditional verdict rather than strengthening it, because the paper has real strengths: a patient-level train/test split, three human graders including an independent glaucoma expert, ICC and Bland-Altman analyses, and an explicit statement of the ground-truth limitation. The reviewer should not reject the paper, but the headline claim is not fully established until an independent-grader or consensus reference is supplied. The machine may simply be reproducing the annotation style of the two trainers, especially for the 95.7% Dice score, which rests on a single manual segmentation. The high inter-observer ICC values among A, B, and C are encouraging but do not rule out a shared systematic bias. A secondary, non-central error is that Section 3.2 defines specificity and sensitivity with identical formulas, so the reported specificity values are not valid; this does not affect the Dice-based central claim but should be corrected. The proposed external multi-grader test directly settles whether the central claim is about anatomical accuracy or merely about matching the training annotators.","tokens_in":11202,"tokens_out":6373,"duration_ms":59882,"concrete_test":"Select a random subset of at least 50 test ASOCT images not used in any prior experiment. Have three graders who were not involved in creating the training labels, ideally from a different institution, mark the scleral spur and segment the iris, corneo-sclera shell, and anterior chamber. Compute (a) machine-versus-each-independent-grader Dice and SSL distance, (b) inter-grader Dice and SSL distance, and (c) machine-versus-consensus (e.g., STAPLE or majority vote) agreement. If machine-versus-consensus falls within the inter-grader range, the central claim survives; if machine accuracy drops significantly relative to human-human agreement, the headline numbers were inflated by replication of the training annotators' style.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's strongest claim is that the DCNN detects the scleral spur 'as accurately as an experienced ophthalmologist' and segments anterior segment structures with a Dice coefficient of 95.7%. What would have to be true is that the human annotations used for evaluation are a valid reference. The paper itself concedes in Section 4 'the lack of an absolute ground truth in labelling.' More specifically, the SSL labels and manual segmentations used for training were prepared by observers A and B (Section 2.2), and the evaluation of unseen images was performed by A, B, and C. Thus the machine-to-human agreement (ICC in Table 2, Dice in Figure 11) is partly circular: for observers A and B, the comparison measures how well the network reproduces those annotators' own marking style, not anatomical accuracy. For segmentation the situation is stronger: Section 4 states 'we only had one trainer and observer,' so the 95.7% Dice is computed against a single manual segmentation that likely came from the training annotator. If that annotator is internally consistent, a network trained on that style can score high Dice without locating true anatomical boundaries. The claim would still be clinically useful if interpreted as 'matches expert consensus,' but the paper provides no consensus or independent-grader reference; it only shows pairwise agreement with the same lab's annotators and one expert. The high ICC among A, B, and C mitigates the concern, but does not eliminate it, because all three could share the same systematic bias from the same definitions and training environment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents a deep-learning pipeline for anterior-segment optical coherence tomography (ASOCT) images. A hybrid U-Net/FRRnet architecture (FRRUnet) is used to detect the scleral spur via a segmentation approach, an ensemble of U-Net, FRRnet, and FRRUnet segments the iris, corneo-sclera shell, and anterior chamber, and the outputs are used to compute eight clinically relevant ASOCT parameters with an automated quality check. The authors report that, with limited training data, the DCNN localizes the scleral spur on unseen ASOCT images as accurately as an experienced ophthalmologist and achieves a mean Dice coefficient of 95.7% for tissue segmentation. Validation includes inter- and intra-observer comparisons using ICC and Bland-Altman plots, training-size experiments, and parameter-level ICCs.","tokens_in":11459,"tokens_out":6476,"duration_ms":58991,"significance":"The clinical motivation is strong: objective, automated 360-degree quantification of ASOCT images could support angle-closure glaucoma screening and monitoring. The paper's strengths are the patient-level split between training and test sets, the explicit comparison of machine-human agreement with human-human agreement, the training-size ablation, and the use of Bland-Altman and ICC for reproducibility. However, the headline claims are about accuracy, whereas the study design can only establish agreement with human annotators, and the segmentation Dice is computed against a single annotator. The contribution is therefore better framed as a reproducible automated measurement pipeline whose accuracy relative to anatomical truth remains to be established. The reported specificity formula is also incorrect, and the intra-observer comparison for the machine is not apples-to-apples with the human intra-observer comparison.","major_comments":[{"comment":"The specificity and sensitivity are defined by identical formulas: Specificity = |D∩M|/|M| and Sensitivity = |D∩M|/|M|. The correct specificity is the true-negative rate, e.g., |D^c ∩ M^c|/|M^c|. Since the Discussion explicitly claims 'high sensitivity and specificity', the reported specificity values in Figure 11 cannot be interpreted. Please correct the formulas and verify the reported numbers.","section":"§3.2, Specificity/Sensitivity definitions"},{"comment":"The limitations paragraph concedes 'the lack of an absolute ground truth in labelling' and, for ASOCT segmentation, 'we only had one trainer and observer.' Yet the abstract states that the DCNN detects the scleral spur 'as accurately as an experienced ophthalmologist' and reports a 95.7% Dice coefficient. These claims are not supported: for SSL, the machine was trained on labels from observers A and B and then compared with those same observers plus observer C, so the comparison measures agreement with the training annotators, not accuracy against anatomical truth. For segmentation, the Dice score is agreement with a single annotator, not a measure of anatomical accuracy. In addition, §2.2 says manual segmentations were prepared by two trainers while §4 says there was only one trainer/observer for segmentation; this inconsistency must be resolved. Please reframe the claims as agreement/reproducibility, or provide a consensus or independent reference.","section":"§4 and §2.2"},{"comment":"The machine intra-observer test is performed by training a second model from scratch and comparing it with the first model. Since the network is deterministic after training, repeated application of the same model to the same image would yield zero intra-observer variability; comparing two separately trained models measures training stochasticity, not observer repeatability. The statement that the machine 'generally had lower variability than that of human grader' is therefore not established by this comparison. Please report repeated-inference consistency separately, or explicitly frame the second-model comparison as a training-stability analysis and avoid direct claims of lower intra-observer variability relative to humans.","section":"§3.1, Intra-observer test"},{"comment":"The SSL confidence threshold in the quality check is set to 0.8 after 'visually comparison of good and failed cases' on the data under study, and the test images are stated to be 'made sure to be of usable quality clinically.' This does not validate the claim that the quality check 'asserts the reliability' of the extracted parameters on poor-quality scans. Please validate the quality check on an independent set containing poor-quality images, or temper the claim accordingly.","section":"§3.4 and §2.6"}],"minor_comments":[{"comment":"'Nayoga, Japan' should be 'Nagoya, Japan'.","section":"§2.1"},{"comment":"The column headed 'A vs B vs C' is ambiguous; specify whether the ICC is computed across the three human graders jointly or as pairwise averages.","section":"Table 3"},{"comment":"The number of annotators for the manual segmentations is stated inconsistently; please clarify whether one or two trainers prepared the segmentation labels.","section":"§2.2 and §4"},{"comment":"In the Dice formula, consider using TP/FP/FN notation so that readers can verify the definition and compare it with the standard Dice formulation.","section":"§3.2"},{"comment":"The author name in reference [45] is corrupted ('zgn iek'); it should be corrected, e.g., Özgün Çiçek.","section":"Reference [45]"},{"comment":"Please clarify whether the two halves of a divided image can appear in both the training and testing sets; the current text only states that images were divided in half.","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for an applied imaging/ophthalmology venue. The main concern is overclaiming accuracy from agreement data. With corrected metric definitions, resolved annotator-count inconsistencies, and a reframing of the central claims from accuracy to agreement/reproducibility, the paper could be acceptable. No concerns about research integrity beyond the reporting errors noted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a genuine engineering contribution, not a mirage. The authors combine U-Net and FRRNet into a hybrid they call FRRUnet, train it on a modest dataset of swept-source ASOCT images, and validate scleral spur localization against three human observers with high ICC (0.98+ for coordinates) and reasonable Bland-Altman agreement. They also show that a segmentation approach beats a regression baseline and reaches human-level localization with about 100 training images. The automated quality check and 360-degree parameter extraction are useful practical additions.\n\nThe paper is honest about its biggest limitation: there is no absolute ground truth for the scleral spur. That means the headline claim should be read as \"as accurately as these experts mark the spur,\" not \"as accurately as the true anatomical landmark can be found.\" The high human inter-observer agreement helps, but all three graders share the same training definitions and environment, so shared bias is not ruled out.\n\nThe segmentation story is weaker than the localization story. The 95.7% Dice is against a single manual segmentation, and Section 4 admits there was only one trainer and observer. A network trained on one person's style can score high Dice without finding boundaries that another expert would agree on. There is no consensus reference for segmentation. Also, the specificity and sensitivity formulas in Section 3.2 are simply wrong: both are written as |D∩M|/|M|, which is not specificity or sensitivity by any standard definition. That is a concrete error that needs fixing before anyone relies on those numbers. No code or data is released, which limits reproducibility.\n\nNone of this kills the paper. The central claim, properly qualified, is plausible and clinically relevant. The ICC tables and training-size experiments are genuine evidence. What is missing is external validation with independent multi-rater labels, a corrected metrics section, and preferably public artifacts. I would send this to peer review rather than desk-reject. After revision, it would be a useful reference for anyone building automated ASOCT pipelines.","headline":"A useful ASOCT segmentation and scleral-spur localization pipeline whose headline accuracy is real but measured against the same lab's annotations, so it needs revision rather than rejection.","tokens_in":12054,"tokens_out":1959,"would_cite":true,"duration_ms":18051,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A deep convolutional neural network localizes the scleral spur on anterior-segment OCT images at expert level and segments anterior-chamber structures with a Dice coefficient of 95.7%.","keywords":["anterior segment OCT","scleral spur","deep convolutional neural network","image segmentation","landmark localization","angle closure glaucoma","glaucoma screening","U-Net"],"falsifier":"One test: gather a new ASOCT dataset in which multiple fellowship-trained ophthalmologists independently mark the scleral spur and segment the tissues on the same scans. If the machine's spur locations and Dice scores against the pooled expert consensus fall outside the human-human agreement range, or if the machine systematically tracks one annotator's bias rather than the consensus, the claim of expert-level accuracy fails. A second, harder test would compare spur locations implied by the OCT images against the same eyes' gonioscopic or histologic angle anatomy, the clinical reference for true closure.","tokens_in":11003,"feed_emoji":"👁","tokens_out":7257,"duration_ms":64869,"temperature":0.7,"pith_summary":"This paper reports a deep convolutional neural network that performs two tasks on anterior-segment optical coherence tomography (ASOCT) images of the eye: localizing the scleral spur, the landmark from which angle-closure glaucoma measurements are taken, and segmenting the iris, corneo-sclera shell, and anterior chamber. Using only 620 training images from 75 eyes, the network marked the scleral spur on unseen images as accurately as an experienced ophthalmologist, with inter-observer agreement statistically comparable to human grader agreement. It segmented the three structures with a Dice coefficient of 95.7%. From the predicted spur and segmentations, the system automatically computes eight clinical anterior-chamber parameters and attaches a quality score so that unreliable scans are flagged. If these results hold, a fully automated pipeline can replace manual grading and make 360-degree, whole-eye quantification of the anterior chamber feasible for glaucoma screening and monitoring.","feed_headline":"Deep network finds the eye's scleral spur as well as an expert","feed_subtitle":"It also segments iris, cornea, and anterior chamber at 95.7 percent Dice for automated glaucoma screening.","key_machinery":"The load-bearing mechanism is a hybrid network called FRRUnet, which combines the U-Net skip-connection encoder-decoder design (for local texture plus contextual arrangement) with the full-resolution residual pathway of the full-resolution residual network (for precise boundary localization). For landmark detection the paper turns localization into a three-class segmentation problem that predicts a focus region (a small square centered on the scleral spur), an attention region, and background; the spur is taken as the focus-region center, which avoids coordinate regression's need for many training examples. For tissue segmentation, predictions from the three base networks are blended by a tiny ensemble network with frozen base weights. A two-step quality check then filters scans whose spur confidence (intersection over union at least 0.80) or contour structure indicates poor image quality, so that only reliable parameter measurements are reported.","core_discovery":"The paper's central claim is that a segmentation-based deep-learning pipeline can replace the human grader for both scleral-spur localization and anterior-segment tissue segmentation in ASOCT images, despite limited training data. Rather than regressing coordinates directly, the network is trained to label a small focus region around the spur together with an attention region and background; this segmentation formulation is what lets a modest dataset reach expert-level accuracy. On unseen images the machine's spur placement agreed with a fellowship-trained glaucoma expert at intraclass correlation coefficients of 0.984 (X) and 0.993 (Y), and in parameter measurements the machine's ACW, TISA, and AOD values agreed with experts to a similar degree as experts agreed with each other, while the machine was more repeatable across repeat runs than human graders. The same pipeline segments the anterior-chamber structures at a mean Dice of 95.7%, then computes eight clinical parameters and passes or rejects each scan based on an automated confidence and contour-count quality check.","pith_inferences":["An extension the paper leaves implicit: averaging labels from several ophthalmologists during training should shrink the inherited human bias the authors discuss, and could make the machine's spur location closer to a consensus ground truth than any single expert.","The same \"detect a landmark by segmenting a small region around it\" trick may transfer to other small anatomical landmarks in OCT, ultrasound, or X-ray where coordinate regression needs excessive training data.","A testable consequence: retraining and testing the pipeline on scans from other devices and populations, or on eyes with more severe angle closure, would show whether the expert-level agreement and 95.7% Dice generalize beyond the single hospital cohort used here."],"forward_implications":["Because the spur and segmentations are produced automatically for every scan, all 128 radial sections of a swept-source ASOCT volume can be quantified, yielding a 360-degree map of angle metrics such as AOD and TISA instead of a few hand-picked sections.","The machine's spur localization agreed with human experts about as well as experts agreed with each other, while being more consistent across repeat runs, so automated measurements should improve reproducibility of angle-closure assessment.","The built-in quality check rejects scans with low spur confidence or implausible contour counts before parameters are reported, which is necessary for deploying automated quantification in a screening workflow.","The segmentation-based formulation reaches expert-level spur detection with roughly 100 images, so the approach can be adapted to other anterior-segment landmarks without a large annotation campaign."],"supporting_citations":[{"why":"It documents the reproducibility problem in manual scleral-spur identification for anterior-chamber angle measurements, motivating automated localization.","marker":"[4]"},{"why":"It supplies the three anatomical definitions of the scleral spur used to label the training images.","marker":"[32]"},{"why":"It provides the U-Net skip-connection architecture that the hybrid network builds on for local-plus-contextual segmentation.","marker":"[30]"},{"why":"It provides the full-resolution residual pathway that the hybrid FRRUnet uses for precise boundary localization.","marker":"[26]"},{"why":"It documents the coordinate-regression DCNN approach that requires large labelled datasets and is rejected in favor of segmentation-based localization.","marker":"[38]"},{"why":"It describes the semi-automated commercial program that the fully automatic pipeline is compared against as the clinical baseline.","marker":"[24]"},{"why":"It supplies the image-analysis software used to create the manual tissue segmentations that train and test the segmentation network.","marker":"[15]"},{"why":"It defines the AOD and TISA angle metrics that the automated parameter pipeline computes from the spur and segmentations.","marker":"[41]"}],"fun_headline_variants":["AI matches expert on scleral spur in OCT images","Deep net segments eye structures at 95.7% Dice","Automated glaucoma screening from OCT scans","Neural network identifies eye structures in OCT","Machine learning quantifies anterior segment dimensions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the human-annotated scleral spur and the single-annotator tissue labels used as training targets are reliable enough proxies for the true anatomy, since the paper concedes there is no absolute ground truth for labeling; if those labels carry systematic bias or error, the network simply reproduces that bias and \"as accurate as an ophthalmologist\" does not mean anatomically correct.","fun_headline_variants_meta":{"raw":{"variants":["AI matches expert on scleral spur in OCT images","Deep net segments eye structures at 95.7% Dice","Automated glaucoma screening from OCT scans","Neural network identifies eye structures in OCT","Machine learning quantifies anterior segment dimensions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000303,"raw_usage":{"total_tokens":1747,"prompt_tokens":950,"completion_tokens":797,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":726}},"tokens_in":566,"tokens_out":797,"duration_ms":7987,"temperature":1.0,"reasoning_tokens":726,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:54:49.232676+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One test: gather a new ASOCT dataset in which multiple fellowship-trained ophthalmologists independently mark the scleral spur and segment the tissues on the same scans. If the machine's spur locations and Dice scores against the pooled expert consensus fall outside the human-human agreement range, or if the machine systematically tracks one annotator's bias rather than the consensus, the claim of expert-level accuracy fails. A second, harder test would compare spur locations implied by the OCT images against the same eyes' gonioscopic or histologic angle anatomy, the clinical reference for true closure.","supporting_citations":[{"cited_title":"Cumba, Sunita Radhakrishnan, Nicholas P","cited_arxiv_id":null,"evidence_quote":"It documents the reproducibility problem in manual scleral-spur identification for anterior-chamber angle measurements, motivating automated localization."},{"cited_title":"Assessment of the scleral spur in anterior segment optical coherence tomography images","cited_arxiv_id":null,"evidence_quote":"It supplies the three anatomical definitions of the scleral spur used to label the training images."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation","cited_arxiv_id":null,"evidence_quote":"It provides the U-Net skip-connection architecture that the hybrid network builds on for local-plus-contextual segmentation."},{"cited_title":"Fullresolution residual net- works for semantic segmentation in street scenes","cited_arxiv_id":null,"evidence_quote":"It provides the full-resolution residual pathway that the hybrid FRRUnet uses for precise boundary localization."},{"cited_title":"Deep neural networks for object detection","cited_arxiv_id":null,"evidence_quote":"It documents the coordinate-regression DCNN approach that requires large labelled datasets and is rejected in favor of segmentation-based localization."},{"cited_title":"Okamoto, K","cited_arxiv_id":null,"evidence_quote":"It describes the semi-automated commercial program that the fully automatic pipeline is compared against as the clinical baseline."},{"cited_title":"Predicting the outcome of laser peripheral iridotomy for primary angle closure suspect eyes using anterior segment optical coherence tomography","cited_arxiv_id":null,"evidence_quote":"It supplies the image-analysis software used to create the manual tissue segmentations that train and test the segmentation network."},{"cited_title":"Ethnic difference of the anterior chamber area and volume and its association with angle width","cited_arxiv_id":null,"evidence_quote":"It defines the AOD and TISA angle metrics that the automated parameter pipeline computes from the spur and segmentations."}],"review_version":1}