Pub Date : 2025-12-29DOI: 10.1016/j.xops.2025.101040
Nianjia Wang , Xindi Liu , Liang Yao
{"title":"Re: Hiraoka et al: Quantitative Mapping of Posterior Eye Curvature in Children Using Distortion-Corrected OCT: Insights into Temporal Region Morphology","authors":"Nianjia Wang , Xindi Liu , Liang Yao","doi":"10.1016/j.xops.2025.101040","DOIUrl":"10.1016/j.xops.2025.101040","url":null,"abstract":"","PeriodicalId":74363,"journal":{"name":"Ophthalmology science","volume":"6 3","pages":"Article 101040"},"PeriodicalIF":4.6,"publicationDate":"2025-12-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145969431","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-26DOI: 10.1016/j.xops.2025.101050
Sahana Srinivasan BEng , Xuguang Ai MSc , Thaddaeus Wai Soon Lo BEng , Aidan Gilson MD , Minjie Zou MD, MMed , Ke Zou PhD , Hyunjae Kim PhD , Mingjia Yang BEng , Krithi Pushpanathan MSc , Samantha Min Er Yew MSc , Wan Ting Loke MSc , Jocelyn Hui Lin Goh BEng , Yibing Chen BEng , Yiming Kong , Emily Yuelei Fu MSc , Michelle Ong BEng , Kristen Nwanyanwu MD, MHS , Amisha Dave MD , Kelvin Zhenghao Li MBBS, MMed , Chen-Hsin Sun MD. MMed , Yih-Chung Tham PhD
<div><h3>Purpose</h3><div>Current benchmarks evaluating large language models (LLMs) in ophthalmology are narrow and disproportionately prioritize accuracy. We introduce BEnchmarking LLMs for Ophthalmology (BELO), a standardized evaluation benchmark developed through multiple rounds of expert checking by 13 ophthalmologists. BEnchmarking LLMs for Ophthalmology assesses ophthalmology-related knowledge and reasoning quality.</div></div><div><h3>Subjects</h3><div>This study did not involve human participation.</div></div><div><h3>Design</h3><div>Cross-sectional study.</div></div><div><h3>Methods</h3><div>Using keyword matching and a fine-tuned PubMed Bidirectional Encoder Representations from Transformers model, we curated ophthalmology-specific multiple-choice questions (MCQs) from diverse medical data sets (Basic and Clinical Science Course [BCSC], Multi-Subject Multi-Choice Dataset for Medical domain [MedMCQA], Medical Question Answering [MedQA], Biomedical Semantic Indexing and Question Answering [BioASQ], and PubMed Question Answering [PubMedQA]). The data set underwent multiple rounds of expert checking. Duplicate and substandard questions were systematically removed. Ten ophthalmologists refined the explanations of each MCQ's correct answer. This was further adjudicated by 3 senior ophthalmologists. To illustrate BELO's utility, we evaluated 8 LLMs (OpenAI o1, o3-mini, GPT-5, GPT-4o, DeepSeek-R1, MedGemma-4B, Llama-3-8B, and Gemini 1.5 Pro).</div></div><div><h3>Main Outcome Measures</h3><div>The 8 LLMs were evaluated in terms using accuracy, macro-F1, and 5 text-generation metrics (Recall-Oriented Understudy for Gisting Evaluation, BERTScore, BARTScore, Metric for Evaluation of Translation with Explicit Ordering, and AlignScore). In a further evaluation involving human experts, 2 ophthalmologists qualitatively reviewed 50 randomly selected outputs for accuracy, comprehensiveness, and completeness.</div></div><div><h3>Results</h3><div>BEnchmarking LLMs for Ophthalmology consists of 900 high-quality, expert-reviewed questions aggregated from 5 sources: BCSC (260), BioASQ (10), MedMCQA (572), MedQA (40), and PubMedQA (18). To demonstrate BELO's utility, we conducted a series of benchmarking exercises. In the quantitative evaluation, GPT-5 achieved the highest accuracy (0.90, 95% confidence interval [CI]: 0.89–0.92) and macro-F1 score (0.91, 95% CI: 0.89–0.93). On the other hand, the models' performance on text-generation metrics varied and were generally suboptimal, with scores ranging from 20.4 to 72.0 (out of 100, excluding the BARTScore metric), indicating room for improvement in clinical reasoning. In expert evaluations, GPT-4o was rated highest for accuracy and readability, while Gemini 1.5 Pro scored highest for completeness. A public leaderboard has been established to promote transparent evaluation and reporting. Importantly, the BELO data set will remain a hold-out, evaluation-only benchmark to ensure fair and reproducible comparisons o
{"title":"BEnchmarking Large Language Models for Ophthalmology (BELO): An Expert-Curated Data Set and Evaluation Framework for Knowledge and Reasoning","authors":"Sahana Srinivasan BEng , Xuguang Ai MSc , Thaddaeus Wai Soon Lo BEng , Aidan Gilson MD , Minjie Zou MD, MMed , Ke Zou PhD , Hyunjae Kim PhD , Mingjia Yang BEng , Krithi Pushpanathan MSc , Samantha Min Er Yew MSc , Wan Ting Loke MSc , Jocelyn Hui Lin Goh BEng , Yibing Chen BEng , Yiming Kong , Emily Yuelei Fu MSc , Michelle Ong BEng , Kristen Nwanyanwu MD, MHS , Amisha Dave MD , Kelvin Zhenghao Li MBBS, MMed , Chen-Hsin Sun MD. MMed , Yih-Chung Tham PhD","doi":"10.1016/j.xops.2025.101050","DOIUrl":"10.1016/j.xops.2025.101050","url":null,"abstract":"<div><h3>Purpose</h3><div>Current benchmarks evaluating large language models (LLMs) in ophthalmology are narrow and disproportionately prioritize accuracy. We introduce BEnchmarking LLMs for Ophthalmology (BELO), a standardized evaluation benchmark developed through multiple rounds of expert checking by 13 ophthalmologists. BEnchmarking LLMs for Ophthalmology assesses ophthalmology-related knowledge and reasoning quality.</div></div><div><h3>Subjects</h3><div>This study did not involve human participation.</div></div><div><h3>Design</h3><div>Cross-sectional study.</div></div><div><h3>Methods</h3><div>Using keyword matching and a fine-tuned PubMed Bidirectional Encoder Representations from Transformers model, we curated ophthalmology-specific multiple-choice questions (MCQs) from diverse medical data sets (Basic and Clinical Science Course [BCSC], Multi-Subject Multi-Choice Dataset for Medical domain [MedMCQA], Medical Question Answering [MedQA], Biomedical Semantic Indexing and Question Answering [BioASQ], and PubMed Question Answering [PubMedQA]). The data set underwent multiple rounds of expert checking. Duplicate and substandard questions were systematically removed. Ten ophthalmologists refined the explanations of each MCQ's correct answer. This was further adjudicated by 3 senior ophthalmologists. To illustrate BELO's utility, we evaluated 8 LLMs (OpenAI o1, o3-mini, GPT-5, GPT-4o, DeepSeek-R1, MedGemma-4B, Llama-3-8B, and Gemini 1.5 Pro).</div></div><div><h3>Main Outcome Measures</h3><div>The 8 LLMs were evaluated in terms using accuracy, macro-F1, and 5 text-generation metrics (Recall-Oriented Understudy for Gisting Evaluation, BERTScore, BARTScore, Metric for Evaluation of Translation with Explicit Ordering, and AlignScore). In a further evaluation involving human experts, 2 ophthalmologists qualitatively reviewed 50 randomly selected outputs for accuracy, comprehensiveness, and completeness.</div></div><div><h3>Results</h3><div>BEnchmarking LLMs for Ophthalmology consists of 900 high-quality, expert-reviewed questions aggregated from 5 sources: BCSC (260), BioASQ (10), MedMCQA (572), MedQA (40), and PubMedQA (18). To demonstrate BELO's utility, we conducted a series of benchmarking exercises. In the quantitative evaluation, GPT-5 achieved the highest accuracy (0.90, 95% confidence interval [CI]: 0.89–0.92) and macro-F1 score (0.91, 95% CI: 0.89–0.93). On the other hand, the models' performance on text-generation metrics varied and were generally suboptimal, with scores ranging from 20.4 to 72.0 (out of 100, excluding the BARTScore metric), indicating room for improvement in clinical reasoning. In expert evaluations, GPT-4o was rated highest for accuracy and readability, while Gemini 1.5 Pro scored highest for completeness. A public leaderboard has been established to promote transparent evaluation and reporting. Importantly, the BELO data set will remain a hold-out, evaluation-only benchmark to ensure fair and reproducible comparisons o","PeriodicalId":74363,"journal":{"name":"Ophthalmology science","volume":"6 3","pages":"Article 101050"},"PeriodicalIF":4.6,"publicationDate":"2025-12-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146189246","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-24DOI: 10.1016/j.xops.2025.101043
Zhen Ji Chen PhD , Jun Yu MMed , Mary Ho MSc , Danny S.C. Ng MPH , Marten E. Brelen PhD , Alvin L. Young MMedSc , Jason C.S. Yam MD , Clement C. Tham BM, BCh , Chi Pui Pang DPhil , Li Jia Chen PhD
<div><h3>Purpose</h3><div>To evaluate the effects of haplotype-tagging single nucleotide polymorphisms (SNPs) in the complement factor H–complement factor H related 5 (<em>CFH</em>–<em>CFHR5</em>) locus on neovascular age-related macular degeneration (nAMD), polypoidal choroidal vasculopathy (PCV), and chronic central serous chorioretinopathy (cCSCR) in Chinese patients.</div></div><div><h3>Design</h3><div>Case-control genetic association study.</div></div><div><h3>Participants</h3><div>A total of 846 patients (341 nAMD, 288 PCV, and 217 cCSCR including 43 with secondary macular neovascularization [MNV]) and 632 healthy Chinese controls.</div></div><div><h3>Methods</h3><div>A total of 17 candidate SNPs were initially selected from the <em>CFH–CFHR5</em> region; after excluding 5 SNPs that deviated from Hardy–Weinberg equilibrium, 12 SNPs were retained for the final analysis. Association analyses included logistic regression adjusted for age and sex and haplotype-based analysis using Haploview. Study-wide significance threshold was set at <em>P</em> < 0.0042 for allelic tests (Bonferroni-corrected for 12 SNPs) and at <em>P</em> < 0.05 for haplotype tests (adjusted using 10 000 permutations).</div></div><div><h3>Main Outcome Measures</h3><div>Associations between individual SNPs and haplotypes in the <em>CFH</em>–<em>CFHR5</em> locus with nAMD, PCV, and cCSCR (with or without MNV), respectively.</div></div><div><h3>Results</h3><div>The tagging SNP, rs12144939, for the <em>CFHR3/1</em> deletion was significantly associated with nAMD (odds ratio [OR] = 0.37, <em>P</em> = 0.0031). Notably, we identified 3 candidate variants showing novel associations with PCV, including rs12144939 (OR = 0.29, <em>P</em> = 6.29 × 10<sup>–4</sup>), rs423641 in <em>CFHR1</em> (OR = 0.74, <em>P</em> = 0.0038), and rs10922152 in <em>CFHR5</em> (OR = 1.55, <em>P</em> = 0.0031). No SNP in this locus was associated with cCSCR without MNV, whereas <em>CFH</em> rs529825 was nominally associated with cCSCR with MNV (OR = 0.47, <em>P</em> = 0.0047). Similar patterns of haplotype associations were observed across the 3 maculopathies. Notably, the haplotype A-T-C-G spanning <em>CFHR4</em>, <em>CFHR2,</em> and <em>CFHR5</em> (OR = 1.81, permutation <em>P</em> = 0.0099) and haplotype G-A-G within <em>CFHR5</em> (OR = 1.56, permutation <em>P</em> = 0.025) were specifically associated with PCV.</div></div><div><h3>Conclusions</h3><div>This study validates the association of the <em>CFHR3/1</em> deletion (tagged by rs12144939) with nAMD. Furthermore, we reveal a novel genetic architecture for PCV within the <em>CFH</em>–<em>CFHR5</em> locus, characterized by associations at rs12144939, rs423641 (<em>CFHR1</em>), and rs10922152 (<em>CFHR5</em>), as well as risk haplotypes unique to PCV. These findings underscore the critical role of <em>CFH</em>-related genes in PCV and provide new insights into its genetic mechanisms.</div></div><div><h3>Financial Disclosures</h3><div>The author h
{"title":"The CFH–CFHR5 Locus in Wet Age-Related Macular Degeneration, Polypoidal Choroidal Vasculopathy, and Central Serous Chorioretinopathy","authors":"Zhen Ji Chen PhD , Jun Yu MMed , Mary Ho MSc , Danny S.C. Ng MPH , Marten E. Brelen PhD , Alvin L. Young MMedSc , Jason C.S. Yam MD , Clement C. Tham BM, BCh , Chi Pui Pang DPhil , Li Jia Chen PhD","doi":"10.1016/j.xops.2025.101043","DOIUrl":"10.1016/j.xops.2025.101043","url":null,"abstract":"<div><h3>Purpose</h3><div>To evaluate the effects of haplotype-tagging single nucleotide polymorphisms (SNPs) in the complement factor H–complement factor H related 5 (<em>CFH</em>–<em>CFHR5</em>) locus on neovascular age-related macular degeneration (nAMD), polypoidal choroidal vasculopathy (PCV), and chronic central serous chorioretinopathy (cCSCR) in Chinese patients.</div></div><div><h3>Design</h3><div>Case-control genetic association study.</div></div><div><h3>Participants</h3><div>A total of 846 patients (341 nAMD, 288 PCV, and 217 cCSCR including 43 with secondary macular neovascularization [MNV]) and 632 healthy Chinese controls.</div></div><div><h3>Methods</h3><div>A total of 17 candidate SNPs were initially selected from the <em>CFH–CFHR5</em> region; after excluding 5 SNPs that deviated from Hardy–Weinberg equilibrium, 12 SNPs were retained for the final analysis. Association analyses included logistic regression adjusted for age and sex and haplotype-based analysis using Haploview. Study-wide significance threshold was set at <em>P</em> < 0.0042 for allelic tests (Bonferroni-corrected for 12 SNPs) and at <em>P</em> < 0.05 for haplotype tests (adjusted using 10 000 permutations).</div></div><div><h3>Main Outcome Measures</h3><div>Associations between individual SNPs and haplotypes in the <em>CFH</em>–<em>CFHR5</em> locus with nAMD, PCV, and cCSCR (with or without MNV), respectively.</div></div><div><h3>Results</h3><div>The tagging SNP, rs12144939, for the <em>CFHR3/1</em> deletion was significantly associated with nAMD (odds ratio [OR] = 0.37, <em>P</em> = 0.0031). Notably, we identified 3 candidate variants showing novel associations with PCV, including rs12144939 (OR = 0.29, <em>P</em> = 6.29 × 10<sup>–4</sup>), rs423641 in <em>CFHR1</em> (OR = 0.74, <em>P</em> = 0.0038), and rs10922152 in <em>CFHR5</em> (OR = 1.55, <em>P</em> = 0.0031). No SNP in this locus was associated with cCSCR without MNV, whereas <em>CFH</em> rs529825 was nominally associated with cCSCR with MNV (OR = 0.47, <em>P</em> = 0.0047). Similar patterns of haplotype associations were observed across the 3 maculopathies. Notably, the haplotype A-T-C-G spanning <em>CFHR4</em>, <em>CFHR2,</em> and <em>CFHR5</em> (OR = 1.81, permutation <em>P</em> = 0.0099) and haplotype G-A-G within <em>CFHR5</em> (OR = 1.56, permutation <em>P</em> = 0.025) were specifically associated with PCV.</div></div><div><h3>Conclusions</h3><div>This study validates the association of the <em>CFHR3/1</em> deletion (tagged by rs12144939) with nAMD. Furthermore, we reveal a novel genetic architecture for PCV within the <em>CFH</em>–<em>CFHR5</em> locus, characterized by associations at rs12144939, rs423641 (<em>CFHR1</em>), and rs10922152 (<em>CFHR5</em>), as well as risk haplotypes unique to PCV. These findings underscore the critical role of <em>CFH</em>-related genes in PCV and provide new insights into its genetic mechanisms.</div></div><div><h3>Financial Disclosures</h3><div>The author h","PeriodicalId":74363,"journal":{"name":"Ophthalmology science","volume":"6 2","pages":"Article 101043"},"PeriodicalIF":4.6,"publicationDate":"2025-12-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146037950","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-24DOI: 10.1016/j.xops.2025.101042
Xiangtian Ling PhD , Yu Peng MMed , Yuzhou Zhang PhD , Charlene C. Yim FCOphthHK, FRCOphth , Hei-Nga Chan PhD , Yating Yang MMed , Qihang Sun MMed , Xiu-Juan Zhang PhD , Ka Wai Kam FCOphthHK, MSc , Wai Kit Chu DPhil , Patrick Ip MD , Alvin L. Young FRCSI, MMedSc , Christopher J. Hammond MD, FRCOphth , Stephen Kwok Wing Tsui PhD , Clement C. Tham FCOphthHK, FRCOphth , Chi Pui Pang DPhil , Li Jia Chen PhD, FCOphthHK , Jason C. Yam MD, FCOphthHK
Purpose
To identify the compositional and functional alterations in the ocular surface microbiome (OSM) which are associated with myopia in children and adolescents.
Design
A population-based, cross-sectional study.
Participants
Eight hundred forty-seven children and adolescents aged 3 to 17 years were included.
Methods
Conjunctival swab samples were collected from the participants and processed via 16S ribosomal RNA gene sequencing.
Main Outcome Measures
Microbial profiles of participants were processed with QIIME2. Alpha (species diversity) and beta diversity (community structure) metrics were calculated. Microbial functional profile was predicted using PICRUSt2.
Results
Shannon (P < 0.001) and observed (P = 0.010) indexes were different among samples from myopic eyes (n = 432), as compared with those from emmetropic (n = 214) and hyperopic (n = 201) eyes. They were correlated with spherical equivalent (Shannon P = 0.0036, observed P = 0.0129) and axial length (Shannon P = 0.0057, observed P = 0.012). Beta diversity with distinct microbial signatures was unique (P < 0.05) among the eyes with myopia (Haemophilus, Aquabacterium, Anaerococcus), emmetropia (Sphingobium, Clostridium sensu stricto 1, and Fusobacterium) and hyperopia (Streptococcus, Kocuria, and Gemella). Functional profiling found enrichment of several Kyoto Encyclopedia of Genes and Genomes pathways, including oxidative phosphorylation, in the myopic ocular surface, suggesting a distinct energy utilization pattern in the myopic microbiome.
Conclusions
This study reveals distinct compositional and functional profiles in the OSM of myopic children and adolescents. These findings demonstrate an association between refractive status and the OSM; however, causality has not been established, highlighting the need for further research.
Financial Disclosures
The author has no/the authors have no proprietary or commercial interest in any materials discussed in this article.
{"title":"Associations between Ocular Surface Microbiome and Refractive Status in Children and Adolescents","authors":"Xiangtian Ling PhD , Yu Peng MMed , Yuzhou Zhang PhD , Charlene C. Yim FCOphthHK, FRCOphth , Hei-Nga Chan PhD , Yating Yang MMed , Qihang Sun MMed , Xiu-Juan Zhang PhD , Ka Wai Kam FCOphthHK, MSc , Wai Kit Chu DPhil , Patrick Ip MD , Alvin L. Young FRCSI, MMedSc , Christopher J. Hammond MD, FRCOphth , Stephen Kwok Wing Tsui PhD , Clement C. Tham FCOphthHK, FRCOphth , Chi Pui Pang DPhil , Li Jia Chen PhD, FCOphthHK , Jason C. Yam MD, FCOphthHK","doi":"10.1016/j.xops.2025.101042","DOIUrl":"10.1016/j.xops.2025.101042","url":null,"abstract":"<div><h3>Purpose</h3><div>To identify the compositional and functional alterations in the ocular surface microbiome (OSM) which are associated with myopia in children and adolescents.</div></div><div><h3>Design</h3><div>A population-based, cross-sectional study.</div></div><div><h3>Participants</h3><div>Eight hundred forty-seven children and adolescents aged 3 to 17 years were included.</div></div><div><h3>Methods</h3><div>Conjunctival swab samples were collected from the participants and processed via 16S ribosomal RNA gene sequencing.</div></div><div><h3>Main Outcome Measures</h3><div>Microbial profiles of participants were processed with QIIME2. Alpha (species diversity) and beta diversity (community structure) metrics were calculated. Microbial functional profile was predicted using PICRUSt2.</div></div><div><h3>Results</h3><div>Shannon (<em>P</em> < 0.001) and observed (<em>P</em> = 0.010) indexes were different among samples from myopic eyes (n = 432), as compared with those from emmetropic (n = 214) and hyperopic (n = 201) eyes. They were correlated with spherical equivalent (Shannon <em>P</em> = 0.0036, observed <em>P</em> = 0.0129) and axial length (Shannon <em>P</em> = 0.0057, observed <em>P</em> = 0.012). Beta diversity with distinct microbial signatures was unique (<em>P</em> < 0.05) among the eyes with myopia (<em>Haemophilus</em>, <em>Aquabacterium</em>, <em>Anaerococcus</em>), emmetropia (<em>Sphingobium</em>, <em>Clostridium sensu stricto 1</em>, and <em>Fusobacterium</em>) and hyperopia (<em>Streptococcus</em>, <em>Kocuria</em>, and <em>Gemella</em>). Functional profiling found enrichment of several Kyoto Encyclopedia of Genes and Genomes pathways, including oxidative phosphorylation, in the myopic ocular surface, suggesting a distinct energy utilization pattern in the myopic microbiome.</div></div><div><h3>Conclusions</h3><div>This study reveals distinct compositional and functional profiles in the OSM of myopic children and adolescents. These findings demonstrate an association between refractive status and the OSM; however, causality has not been established, highlighting the need for further research.</div></div><div><h3>Financial Disclosures</h3><div>The author has no/the authors have no proprietary or commercial interest in any materials discussed in this article.</div></div>","PeriodicalId":74363,"journal":{"name":"Ophthalmology science","volume":"6 3","pages":"Article 101042"},"PeriodicalIF":4.6,"publicationDate":"2025-12-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039783","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-18DOI: 10.1016/j.xops.2025.101041
Douglas R. da Costa MD , Dániel Unyi , Rafael Scherer MD, PhD , Rohit Muralidhar , Mario Luiz Ribeiro Monteiro MD, PhD , Felipe A. Medeiros MD, PhD
Purpose
To evaluate the performance of a deep learning (DL) model based on graph isomorphism networks (GINs) for detecting glaucomatous visual field defects on 24-2 standard automated perimetry (SAP) and to compare it against traditional diagnostic criteria, a dense neural network (NN) model, and a convolutional neural network (CNN) model.
Design
A cross-sectional retrospective study. Participants: 1874 reliable SAP tests (Humphrey Field Analyzer, Carl-Zeiss Meditec Inc.) from 1009 eyes of 676 patients.
Methods
Standard automated perimetry tests were classified as normal or abnormal due to glaucomatous damage by two glaucoma specialists, with adjudication by a third. A GIN architecture was developed to classify tests using full 54-point spatial SAP data modeled as graphs, with node features comprising sensitivity, total deviation, and pattern deviation values. The dataset was split at the patient level (60% training/validation, 40% testing). The GIN model’s diagnostic performance was compared to the Anderson criteria, the glaucoma hemifield test/pattern standard deviation (GHT/PSD) criteria, a fully connected dense NN, and a CNN model.
Main Outcome Measures
Area under the receiver operating characteristic curve (AUC), precision–recall curve, sensitivity at 95% specificity, F1-score, repeatability, and model explainability.
Results
Among the 1874 SAP tests, 70.0% were graded as abnormal. The GIN model achieved an AUC of 0.982, significantly outperforming the Anderson criteria (AUC: 0.906, P < 0.001), GHT/PSD (AUC: 0.936, P = 0.006), the NN model (AUC: 0.941, P = 0.007), and the CNN model (AUC: 0.941, P = 0.027). At 95% specificity, the GIN model reached the highest sensitivity of 94.1%, surpassing the NN model (88.3%), CNN model (92.0%), GHT/PSD (90.1%), and Anderson criteria (85.1%). The GIN model also achieved the highest average precision (0.952) among evaluated criteria. Explainability analysis using GraphNOSE demonstrated that the GIN model emphasized clinically relevant regions associated with glaucomatous loss, offering interpretability advantages over conventional DL approaches.
Conclusions
By modeling SAP as a graph and incorporating spatial relationships among test points, the GIN model provided superior diagnostic performance and interpretability relative to traditional criteria and standard NNs. This graph-based approach offers a promising tool for accurate and explainable detection of glaucomatous visual field defects in clinical practice.
Financial Disclosures
Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
目的评价基于图同态网络（GINs）的深度学习（DL）模型在24-2标准自动视野测量（SAP）上检测青光眼视野缺陷的性能，并将其与传统诊断标准、密集神经网络（NN）模型和卷积神经网络（CNN）模型进行比较。设计横断面回顾性研究。参与者：1874个可靠的SAP测试（Humphrey Field Analyzer, Carl-Zeiss Meditec Inc.），来自676名患者的1009只眼睛。方法由两名青光眼专家将标准的自动视距检查分为青光眼损伤的正常或异常，并由第三名青光眼专家判定。开发了一个GIN架构，使用完整的54点空间SAP数据建模为图来对测试进行分类，节点特征包括灵敏度、总偏差和模式偏差值。数据集在患者层面被分割（60%用于训练/验证，40%用于测试）。将GIN模型的诊断性能与Anderson标准、青光眼半场测试/模式标准差（GHT/PSD）标准、全连接密集神经网络和CNN模型进行比较。主要结果测量：受试者工作特征曲线（AUC）下的面积、精密度-召回率曲线、95%特异性的灵敏度、f1评分、可重复性和模型可解释性。结果1874例SAP试验中，70.0%为异常。GIN模型的AUC为0.982，显著优于Anderson标准（AUC: 0.906, P < 0.001）、GHT/PSD （AUC: 0.936, P = 0.006）、NN模型（AUC: 0.941, P = 0.007）和CNN模型（AUC: 0.941, P = 0.027）。在95%的特异性下，GIN模型达到了94.1%的最高灵敏度，超过了NN模型（88.3%）、CNN模型（92.0%）、GHT/PSD（90.1%）和Anderson标准（85.1%）。GIN模型在评价标准中平均精度最高（0.952）。使用GraphNOSE进行的可解释性分析表明，GIN模型强调与青光眼丧失相关的临床相关区域，与传统DL方法相比具有可解释性优势。通过将SAP建模为图形，并结合测试点之间的空间关系，GIN模型相对于传统标准和标准神经网络具有更好的诊断性能和可解释性。这种基于图的方法为临床实践中青光眼视野缺陷的准确和可解释的检测提供了一种有前途的工具。财务披露专有或商业披露可在本文末尾的脚注和披露中找到。
{"title":"Convolutional Graph Isomorphism Network to Detect Glaucomatous Visual Field Defects","authors":"Douglas R. da Costa MD , Dániel Unyi , Rafael Scherer MD, PhD , Rohit Muralidhar , Mario Luiz Ribeiro Monteiro MD, PhD , Felipe A. Medeiros MD, PhD","doi":"10.1016/j.xops.2025.101041","DOIUrl":"10.1016/j.xops.2025.101041","url":null,"abstract":"<div><h3>Purpose</h3><div>To evaluate the performance of a deep learning (DL) model based on graph isomorphism networks (GINs) for detecting glaucomatous visual field defects on 24-2 standard automated perimetry (SAP) and to compare it against traditional diagnostic criteria, a dense neural network (NN) model, and a convolutional neural network (CNN) model.</div></div><div><h3>Design</h3><div>A cross-sectional retrospective study. Participants: 1874 reliable SAP tests (Humphrey Field Analyzer, Carl-Zeiss Meditec Inc.) from 1009 eyes of 676 patients.</div></div><div><h3>Methods</h3><div>Standard automated perimetry tests were classified as normal or abnormal due to glaucomatous damage by two glaucoma specialists, with adjudication by a third. A GIN architecture was developed to classify tests using full 54-point spatial SAP data modeled as graphs, with node features comprising sensitivity, total deviation, and pattern deviation values. The dataset was split at the patient level (60% training/validation, 40% testing). The GIN model’s diagnostic performance was compared to the Anderson criteria, the glaucoma hemifield test/pattern standard deviation (GHT/PSD) criteria, a fully connected dense NN, and a CNN model.</div></div><div><h3>Main Outcome Measures</h3><div>Area under the receiver operating characteristic curve (AUC), precision–recall curve, sensitivity at 95% specificity, F1-score, repeatability, and model explainability.</div></div><div><h3>Results</h3><div>Among the 1874 SAP tests, 70.0% were graded as abnormal. The GIN model achieved an AUC of 0.982, significantly outperforming the Anderson criteria (AUC: 0.906, <em>P</em> < 0.001), GHT/PSD (AUC: 0.936, <em>P</em> = 0.006), the NN model (AUC: 0.941, <em>P</em> = 0.007), and the CNN model (AUC: 0.941, <em>P</em> = 0.027). At 95% specificity, the GIN model reached the highest sensitivity of 94.1%, surpassing the NN model (88.3%), CNN model (92.0%), GHT/PSD (90.1%), and Anderson criteria (85.1%). The GIN model also achieved the highest average precision (0.952) among evaluated criteria. Explainability analysis using GraphNOSE demonstrated that the GIN model emphasized clinically relevant regions associated with glaucomatous loss, offering interpretability advantages over conventional DL approaches.</div></div><div><h3>Conclusions</h3><div>By modeling SAP as a graph and incorporating spatial relationships among test points, the GIN model provided superior diagnostic performance and interpretability relative to traditional criteria and standard NNs. This graph-based approach offers a promising tool for accurate and explainable detection of glaucomatous visual field defects in clinical practice.</div></div><div><h3>Financial Disclosures</h3><div>Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.</div></div>","PeriodicalId":74363,"journal":{"name":"Ophthalmology science","volume":"6 3","pages":"Article 101041"},"PeriodicalIF":4.6,"publicationDate":"2025-12-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146079753","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-18DOI: 10.1016/j.xops.2025.101030
Ahammed Sakir Nabil , Sina Gholami MS , Theodore Leng MD, MS , Jennifer I. Lim MD , Minhaj Nur Alam PhD
<div><h3>Purpose</h3><div>To conduct a comprehensive systematic evaluation of federated learning (FL) strategies for multi-disease retinal classification using OCT angiography (OCTA), implementing a 2-part experimental framework to establish foundational feasibility and optimize performance under realistic heterogeneous conditions while ensuring privacy preservation.</div></div><div><h3>Design</h3><div>Retrospective multi-center FL study using a systematic 2-part experimental design: (1) foundational feasibility evaluation under controlled homogeneous conditions, and (2) comprehensive optimization under realistic heterogeneous conditions using Dirichlet distribution partitioning (<em>α</em> = 0.5).</div></div><div><h3>Participants</h3><div>A total of 456 OCTA images from patients with 7 retinal pathologies, with diabetic retinopathy (31.1%) and normal cases (25.2%) comprising the majority, sourced from the public OCTA-500 data set (n = 300) and a private collection from the University of Illinois Chicago (n = 156).</div></div><div><h3>Methods</h3><div>Five FL aggregation strategies (federated averaging [FedAvg], federated proximal [FedProx], federated magnetic resonance imaging [FedMRI], federated Adagrad, and federated Yogi) were systematically evaluated across multiple optimization dimensions: 7 architecture configurations spanning vision transformers, established convolutional neural networks, and hybrid models; 5 transfer learning freezing strategies; 3 local epoch configurations (2, 5, and 10); and scalability analysis across 2, 3, and 5-client federations. Security mechanisms including differential privacy (ε = 1.0–8.0) and secure aggregation were integrated and evaluated. Performance was assessed across 3 classification scenarios: 7-class, 4-class modified, and 4-class streamlined.</div></div><div><h3>Main Outcome Measures</h3><div>Classification accuracy, receiver-operating-characteristic area under the curve (ROC-AUC), and macro-averaged F1-score with comprehensive privacy-utility analysis and computational efficiency metrics.</div></div><div><h3>Results</h3><div>Under controlled conditions, FL achieved superior performance in simplified classifications, with FedAvg, FedProx, and FedMRI reaching 72.09% accuracy versus 69.77% centralized training. Comprehensive optimization identified DenseNet121 as optimal architecture (79.55% accuracy, 89.68% ROC-AUC), with “most” freezing strategy (75% frozen layers) providing 60% training time reduction while maintaining superior performance. Federated proximal demonstrated exceptional resilience to heterogeneity (–11.7% degradation). Bonawitz secure aggregation achieved optimal privacy-utility balance (63.64% accuracy with cryptographic guarantees), whereas differential privacy maintained clinical utility under moderate constraints (ε ≈ 4–6).</div></div><div><h3>Conclusions</h3><div>This systematic evaluation establishes FL as a comprehensive solution for privacy-preserving multi-institutional OCTA-b
{"title":"Federated Learning for Multi-Disease Ophthalmic Diagnostics Using OCT Angiography","authors":"Ahammed Sakir Nabil , Sina Gholami MS , Theodore Leng MD, MS , Jennifer I. Lim MD , Minhaj Nur Alam PhD","doi":"10.1016/j.xops.2025.101030","DOIUrl":"10.1016/j.xops.2025.101030","url":null,"abstract":"<div><h3>Purpose</h3><div>To conduct a comprehensive systematic evaluation of federated learning (FL) strategies for multi-disease retinal classification using OCT angiography (OCTA), implementing a 2-part experimental framework to establish foundational feasibility and optimize performance under realistic heterogeneous conditions while ensuring privacy preservation.</div></div><div><h3>Design</h3><div>Retrospective multi-center FL study using a systematic 2-part experimental design: (1) foundational feasibility evaluation under controlled homogeneous conditions, and (2) comprehensive optimization under realistic heterogeneous conditions using Dirichlet distribution partitioning (<em>α</em> = 0.5).</div></div><div><h3>Participants</h3><div>A total of 456 OCTA images from patients with 7 retinal pathologies, with diabetic retinopathy (31.1%) and normal cases (25.2%) comprising the majority, sourced from the public OCTA-500 data set (n = 300) and a private collection from the University of Illinois Chicago (n = 156).</div></div><div><h3>Methods</h3><div>Five FL aggregation strategies (federated averaging [FedAvg], federated proximal [FedProx], federated magnetic resonance imaging [FedMRI], federated Adagrad, and federated Yogi) were systematically evaluated across multiple optimization dimensions: 7 architecture configurations spanning vision transformers, established convolutional neural networks, and hybrid models; 5 transfer learning freezing strategies; 3 local epoch configurations (2, 5, and 10); and scalability analysis across 2, 3, and 5-client federations. Security mechanisms including differential privacy (ε = 1.0–8.0) and secure aggregation were integrated and evaluated. Performance was assessed across 3 classification scenarios: 7-class, 4-class modified, and 4-class streamlined.</div></div><div><h3>Main Outcome Measures</h3><div>Classification accuracy, receiver-operating-characteristic area under the curve (ROC-AUC), and macro-averaged F1-score with comprehensive privacy-utility analysis and computational efficiency metrics.</div></div><div><h3>Results</h3><div>Under controlled conditions, FL achieved superior performance in simplified classifications, with FedAvg, FedProx, and FedMRI reaching 72.09% accuracy versus 69.77% centralized training. Comprehensive optimization identified DenseNet121 as optimal architecture (79.55% accuracy, 89.68% ROC-AUC), with “most” freezing strategy (75% frozen layers) providing 60% training time reduction while maintaining superior performance. Federated proximal demonstrated exceptional resilience to heterogeneity (–11.7% degradation). Bonawitz secure aggregation achieved optimal privacy-utility balance (63.64% accuracy with cryptographic guarantees), whereas differential privacy maintained clinical utility under moderate constraints (ε ≈ 4–6).</div></div><div><h3>Conclusions</h3><div>This systematic evaluation establishes FL as a comprehensive solution for privacy-preserving multi-institutional OCTA-b","PeriodicalId":74363,"journal":{"name":"Ophthalmology science","volume":"6 3","pages":"Article 101030"},"PeriodicalIF":4.6,"publicationDate":"2025-12-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146189214","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-17DOI: 10.1016/j.xops.2025.101039
Clara Martinez-Perez PhD , Ana Paula Oliveira PhD
Topic
This systematic review and meta-analysis evaluated whether myopia control interventions produce measurable changes in choroidal thickness (ChT) in children and adolescents with myopia compared with single-vision lenses or placebo. The primary aim was to describe patterns of ChT modulation.
Clinical Relevance
Myopia is the most common ocular disorder worldwide and is projected to affect 50% of the global population by 2050. High myopia increases the risk of complications such as myopic maculopathy and retinal detachment. Early biomarkers of treatment efficacy are critical, and ChT has emerged as a promising candidate given its rapid and bidirectional response to visual and pharmacological stimuli.
Methods
The protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO; CRD420251144689). Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines and A Measurement Tool to Assess Systematic Reviews 2 (AMSTAR-2) standards, randomized controlled trials (RCTs) were included if they assessed ChT changes after myopia control interventions in pediatric populations. Searches of PubMed, Web of Science, and Scopus were completed on August 5, 2025. Two reviewers independently screened, extracted, and assessed risk of bias using the Cochrane tool. Pooled mean differences with 95% confidence intervals (CIs) were calculated, and certainty of evidence was rated with Grading of Recommendations, Assessment, Development and Evaluation.
Results
Eleven RCTs including 2190 eyes were analyzed. Repeated low-level red-light therapy induced the largest thickening (mean difference = 24.1 μm, 95% CI: 19.8–28.5; I2 = 77%). Atropine produced modest but significant effects (mean difference = 10.6 μm, 95% CI: 6.7–14.5) with high heterogeneity (I2 = 97%). Orthokeratology yielded consistent increases (mean difference = 13.3 μm, 95% CI: 9.5–17.1; I2 = 6%), while lenslet spectacles showed moderate effects (mean difference = 13.2 μm, 95% CI: 5.7–20.7; I2 = 0%). Evidence certainty was rated high for most interventions and moderate for atropine.
Conclusions
Myopia control interventions produce early, measurable increases in ChT. These findings characterize patterns of choroidal modulation, while their clinical relevance remains uncertain. Further studies integrating ChT with efficacy measures are needed.
Financial Disclosure(s)
Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
{"title":"Impact of Myopia Control Interventions on Choroidal Thickness in Children: A Systematic Review and Meta-Analysis of Randomized Controlled Trials","authors":"Clara Martinez-Perez PhD , Ana Paula Oliveira PhD","doi":"10.1016/j.xops.2025.101039","DOIUrl":"10.1016/j.xops.2025.101039","url":null,"abstract":"<div><h3>Topic</h3><div>This systematic review and meta-analysis evaluated whether myopia control interventions produce measurable changes in choroidal thickness (ChT) in children and adolescents with myopia compared with single-vision lenses or placebo. The primary aim was to describe patterns of ChT modulation.</div></div><div><h3>Clinical Relevance</h3><div>Myopia is the most common ocular disorder worldwide and is projected to affect 50% of the global population by 2050. High myopia increases the risk of complications such as myopic maculopathy and retinal detachment. Early biomarkers of treatment efficacy are critical, and ChT has emerged as a promising candidate given its rapid and bidirectional response to visual and pharmacological stimuli.</div></div><div><h3>Methods</h3><div>The protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO; CRD420251144689). Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines and A Measurement Tool to Assess Systematic Reviews 2 (AMSTAR-2) standards, randomized controlled trials (RCTs) were included if they assessed ChT changes after myopia control interventions in pediatric populations. Searches of PubMed, Web of Science, and Scopus were completed on August 5, 2025. Two reviewers independently screened, extracted, and assessed risk of bias using the Cochrane tool. Pooled mean differences with 95% confidence intervals (CIs) were calculated, and certainty of evidence was rated with Grading of Recommendations, Assessment, Development and Evaluation.</div></div><div><h3>Results</h3><div>Eleven RCTs including 2190 eyes were analyzed. Repeated low-level red-light therapy induced the largest thickening (mean difference = 24.1 μm, 95% CI: 19.8–28.5; I<sup>2</sup> = 77%). Atropine produced modest but significant effects (mean difference = 10.6 μm, 95% CI: 6.7–14.5) with high heterogeneity (I<sup>2</sup> = 97%). Orthokeratology yielded consistent increases (mean difference = 13.3 μm, 95% CI: 9.5–17.1; I<sup>2</sup> = 6%), while lenslet spectacles showed moderate effects (mean difference = 13.2 μm, 95% CI: 5.7–20.7; I<sup>2</sup> = 0%). Evidence certainty was rated high for most interventions and moderate for atropine.</div></div><div><h3>Conclusions</h3><div>Myopia control interventions produce early, measurable increases in ChT. These findings characterize patterns of choroidal modulation, while their clinical relevance remains uncertain. Further studies integrating ChT with efficacy measures are needed.</div></div><div><h3>Financial Disclosure(s)</h3><div>Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.</div></div>","PeriodicalId":74363,"journal":{"name":"Ophthalmology science","volume":"6 2","pages":"Article 101039"},"PeriodicalIF":4.6,"publicationDate":"2025-12-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145976575","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-17DOI: 10.1016/j.xops.2025.101038
Souvick Mukherjee PhD , Dylan Wu , Leon von der Emde MD , Emily Vance MPH , Marco Ji MD , Mehdi Emamverdi MD , Tharindu De Silva PhD , Alisa T. Thavikulwat MD , Jayashree Kalpathy-Cramer PhD , Amitha Domalpally MD, PhD , Catherine A. Cukras MD, PhD , Tiarnán D.L. Keenan BM BCh, PhD
Objective
Reticular pseudodrusen (RPD) represent an important biomarker in age-related macular degeneration (AMD) but are difficult to grade and often assessed only for presence or absence, without quantitative or spatial analysis of RPD burden. The objective was to develop and validate a deep learning model for pixel-level RPD grading on near-infrared reflectance (NIR) images, which are commonly acquired in clinical practice and the most accurate en face detection modality.
Design
Deep learning model development study.
Participants
Five hundred eight images of 117 eyes (70 participants) with or without RPD, over a wide range of AMD severities.
Methods
The ground truth grading pipeline comprised reading center multimodal grading for RPD presence and NIR annotation with RPD contours, followed by pixel-level NIR annotation of all individual RPD lesions. The data set was split 80:20 into training and test sets. A DeepLabv3-ResNet-18 segmentation deep learning model (“ReticularNet”) was trained to perform pixel-level grading of RPD on NIR images. Its performance was compared with that of 4 ophthalmologists.
Main Outcome Measures
Dice similarity coefficient (DSC); intraclass correlation coefficient (ICC) for RPD lesion number, pixel area, and contour area.
Results
For pixel-level grading, ReticularNet achieved a mean DSC of 0.36 (standard deviation 0.16). This was significantly higher than the mean DSC of each ophthalmologist (0.03, 0.13, 0.19, and 0.23; P ≤ 0.02 for each) and of all ophthalmologists together (P < 0.0001). ReticularNet had ICCs of 0.44 (lesion number), 0.56 (pixel area), and 0.61 (contour area), with no significant underestimation or overestimation (P ≥ 0.24). These values were numerically higher than the ICCs of each ophthalmologist, who had ICC ranges of –0.08 to 0.23, –0.05 to 0.40, and –0.09 to 0.58, respectively, and significant underestimation in almost all cases. For all 3 parameters, ReticularNet’s ICC was significantly higher than that of all specialists considered together (P ≤ 0.02).
Conclusions
ReticularNet achieved automated pixel-level grading of RPD on NIR images. Its grading was superior to that of 4 ophthalmologists, across a variety of metrics. We are making the code/models available for research use. Improved access to quantitative and spatial RPD grading should lead to improved understanding of these lesions as important biomarkers of retinal disease.
Financial Disclosure(s)
Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
{"title":"ReticularNet: Automated Pixel-Level Segmentation of Reticular Pseudodrusen on Near-Infrared Reflectance Images by Deep Learning","authors":"Souvick Mukherjee PhD , Dylan Wu , Leon von der Emde MD , Emily Vance MPH , Marco Ji MD , Mehdi Emamverdi MD , Tharindu De Silva PhD , Alisa T. Thavikulwat MD , Jayashree Kalpathy-Cramer PhD , Amitha Domalpally MD, PhD , Catherine A. Cukras MD, PhD , Tiarnán D.L. Keenan BM BCh, PhD","doi":"10.1016/j.xops.2025.101038","DOIUrl":"10.1016/j.xops.2025.101038","url":null,"abstract":"<div><h3>Objective</h3><div>Reticular pseudodrusen (RPD) represent an important biomarker in age-related macular degeneration (AMD) but are difficult to grade and often assessed only for presence or absence, without quantitative or spatial analysis of RPD burden. The objective was to develop and validate a deep learning model for pixel-level RPD grading on near-infrared reflectance (NIR) images, which are commonly acquired in clinical practice and the most accurate en face detection modality.</div></div><div><h3>Design</h3><div>Deep learning model development study.</div></div><div><h3>Participants</h3><div>Five hundred eight images of 117 eyes (70 participants) with or without RPD, over a wide range of AMD severities.</div></div><div><h3>Methods</h3><div>The ground truth grading pipeline comprised reading center multimodal grading for RPD presence and NIR annotation with RPD contours, followed by pixel-level NIR annotation of all individual RPD lesions. The data set was split 80:20 into training and test sets. A DeepLabv3-ResNet-18 segmentation deep learning model (“ReticularNet”) was trained to perform pixel-level grading of RPD on NIR images. Its performance was compared with that of 4 ophthalmologists.</div></div><div><h3>Main Outcome Measures</h3><div>Dice similarity coefficient (DSC); intraclass correlation coefficient (ICC) for RPD lesion number, pixel area, and contour area.</div></div><div><h3>Results</h3><div>For pixel-level grading, ReticularNet achieved a mean DSC of 0.36 (standard deviation 0.16). This was significantly higher than the mean DSC of each ophthalmologist (0.03, 0.13, 0.19, and 0.23; <em>P</em> ≤ 0.02 for each) and of all ophthalmologists together (<em>P</em> < 0.0001). ReticularNet had ICCs of 0.44 (lesion number), 0.56 (pixel area), and 0.61 (contour area), with no significant underestimation or overestimation (<em>P</em> ≥ 0.24). These values were numerically higher than the ICCs of each ophthalmologist, who had ICC ranges of –0.08 to 0.23, –0.05 to 0.40, and –0.09 to 0.58, respectively, and significant underestimation in almost all cases. For all 3 parameters, ReticularNet’s ICC was significantly higher than that of all specialists considered together (<em>P</em> ≤ 0.02).</div></div><div><h3>Conclusions</h3><div>ReticularNet achieved automated pixel-level grading of RPD on NIR images. Its grading was superior to that of 4 ophthalmologists, across a variety of metrics. We are making the code/models available for research use. Improved access to quantitative and spatial RPD grading should lead to improved understanding of these lesions as important biomarkers of retinal disease.</div></div><div><h3>Financial Disclosure(s)</h3><div>Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.</div></div>","PeriodicalId":74363,"journal":{"name":"Ophthalmology science","volume":"6 2","pages":"Article 101038"},"PeriodicalIF":4.6,"publicationDate":"2025-12-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146037948","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-15DOI: 10.1016/j.xops.2025.101037
Jiaying Li MD, PhD , Ye Zhang MD, PhD , Zhen Cheng MD, PhD , Chunyan Qiao MD, PhD , Kai Cao MD, PhD , Minguang He MD, PhD , Ningli Wang MD, PhD
Objective
Primary angle closure disease (PACD) has long been the focus of screening, yet most cases are nonprogressive. In contrast, primary angle closure with or without glaucoma (PAC/G) poses greater risk and may deserve more attention. We aimed to compare screening indicator profiles for PACD and PAC/G, highlighting potential differences that may inform a shift in screening priorities.
Design
A population-based cross-sectional study.
Participants
Adults aged ≥35 years who completed standardized eye examinations. Only right eyes were analyzed; eyes with prior laser peripheral iridotomy were excluded.
Methods
Participants underwent gonioscopy, anterior-segment OCT (AS-OCT), and A-scan ultrasound biometry. Univariable logistic regression and multivariable elastic-net models assessed the association and discriminative performance of AS-OCT parameters for detecting PACD and PAC/G.
Main Outcome Measures
Odds ratios (ORs) and area under the receiver operator characteristic curve (AUROC) values of AS-OCT parameters for detecting PACD and PAC/G.
Results
Of the 4546 eyes analyzed, 3831 had open angles and 715 had PACD, of which 53 had PAC/G. In the univariable logistic regression analysis, the angle opening distance (AOD) at 250 μm (AOD250) and the trabecular–iris space area (TISA) at 500 μm (TISA500) were more strongly associated with PAC/G than with PACD (P < 0.05, Wald test), whereas AOD500/750 and TISA750 were not significantly different in association. Additionally, TISA750 exhibited the highest AUROC value (0.88) for detecting PACD, whereas TISA500 had the highest AUROC value (0.91) for detecting PAC/G. Higher iris curvature was significantly associated with increased odds of PACD (OR: 1250.26; 95% confidence interval [CI]: 234.52–2265.99) but not with PAC/G (P < 0.05, Wald test). In the multivariable models for detecting PAC/G, a profile characterized by lower iris curvature (OR: 0.06; 95% CI: 0.05–0.08), narrower angle width (AOD500, TISA500, and TISA750), smaller anterior chamber area and volume, and smaller pupil diameter achieved an AUROC of 0.908 (95% CI: 0.858–0.959).
Conclusions
Angle parameters closer to the scleral spur were underestimated, and curved irises were overestimated when screening for PAC/G versus PACD. Anterior-segment OCT–derived models for PAC/G screening have shown promising feasibility in population-based settings.
Financial Disclosure(s)
The authors have no proprietary or commercial interest in any materials discussed in this article.
{"title":"Differences in Screening Indicator Performance for Primary Angle Closure and Primary Angle Closure Diseases: The Handan Eye Study","authors":"Jiaying Li MD, PhD , Ye Zhang MD, PhD , Zhen Cheng MD, PhD , Chunyan Qiao MD, PhD , Kai Cao MD, PhD , Minguang He MD, PhD , Ningli Wang MD, PhD","doi":"10.1016/j.xops.2025.101037","DOIUrl":"10.1016/j.xops.2025.101037","url":null,"abstract":"<div><h3>Objective</h3><div>Primary angle closure disease (PACD) has long been the focus of screening, yet most cases are nonprogressive. In contrast, primary angle closure with or without glaucoma (PAC/G) poses greater risk and may deserve more attention. We aimed to compare screening indicator profiles for PACD and PAC/G, highlighting potential differences that may inform a shift in screening priorities.</div></div><div><h3>Design</h3><div>A population-based cross-sectional study.</div></div><div><h3>Participants</h3><div>Adults aged ≥35 years who completed standardized eye examinations. Only right eyes were analyzed; eyes with prior laser peripheral iridotomy were excluded.</div></div><div><h3>Methods</h3><div>Participants underwent gonioscopy, anterior-segment OCT (AS-OCT), and A-scan ultrasound biometry. Univariable logistic regression and multivariable elastic-net models assessed the association and discriminative performance of AS-OCT parameters for detecting PACD and PAC/G.</div></div><div><h3>Main Outcome Measures</h3><div>Odds ratios (ORs) and area under the receiver operator characteristic curve (AUROC) values of AS-OCT parameters for detecting PACD and PAC/G.</div></div><div><h3>Results</h3><div>Of the 4546 eyes analyzed, 3831 had open angles and 715 had PACD, of which 53 had PAC/G. In the univariable logistic regression analysis, the angle opening distance (AOD) at 250 μm (AOD250) and the trabecular–iris space area (TISA) at 500 μm (TISA500) were more strongly associated with PAC/G than with PACD (<em>P</em> < 0.05, Wald test), whereas AOD500/750 and TISA750 were not significantly different in association. Additionally, TISA750 exhibited the highest AUROC value (0.88) for detecting PACD, whereas TISA500 had the highest AUROC value (0.91) for detecting PAC/G. Higher iris curvature was significantly associated with increased odds of PACD (OR: 1250.26; 95% confidence interval [CI]: 234.52–2265.99) but not with PAC/G (<em>P</em> < 0.05, Wald test). In the multivariable models for detecting PAC/G, a profile characterized by lower iris curvature (OR: 0.06; 95% CI: 0.05–0.08), narrower angle width (AOD500, TISA500, and TISA750), smaller anterior chamber area and volume, and smaller pupil diameter achieved an AUROC of 0.908 (95% CI: 0.858–0.959).</div></div><div><h3>Conclusions</h3><div>Angle parameters closer to the scleral spur were underestimated, and curved irises were overestimated when screening for PAC/G versus PACD. Anterior-segment OCT–derived models for PAC/G screening have shown promising feasibility in population-based settings.</div></div><div><h3>Financial Disclosure(s)</h3><div>The authors have no proprietary or commercial interest in any materials discussed in this article.</div></div>","PeriodicalId":74363,"journal":{"name":"Ophthalmology science","volume":"6 3","pages":"Article 101037"},"PeriodicalIF":4.6,"publicationDate":"2025-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146189242","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-12-13DOI: 10.1016/j.xops.2025.101036
Yu Jer Hsiao , Hengtong Li MSc , Can Can Xue PhD , Crystal Chun Yuen Chong , Enwen Zhu PhD , Qiang Yuan , Marco Yu PhD , Chui Ming Gemmy Cheung MD , Qiao Fan PhD , Charumathi Sabanayagam PhD , Yih-Chung Tham PhD , Ching-Yu Cheng MD, PhD
Purpose
To evaluate shared genetic influences and investigate the association of chronic kidney disease (CKD) with the risk for advanced age-related macular degeneration (AMD).
Design
Prospective cohort study and 2-sample Mendelian randomization (MR) analyses.
Participants
Data from 430 016 participants in the UK Biobank cohort and summary statistics from the largest publicly available genome-wide association studies on estimated glomerular filtration rate (eGFR) (n = 1 004 040) and advanced AMD (n = 33 976; 16 144 cases) were analyzed.
Methods
Cox regression models were used to assess the association between CKD and incident AMD, adjusting for demographic, lifestyle, and clinical covariates. For MR analyses, we used the random-effects inverse-variance weighted model as the primary model, supported by 5 additional MR models for sensitivity analyses. A causal relationship was considered significant if P < 0.05 in the primary model and in ≥2 sensitivity models, with all MR models showing a consistent effect direction. Colocalization analysis was performed to further identify shared genetic loci linking CKD and AMD.
Main Outcome Measures
Causal associations between eGFR and advanced AMD.
Results
In the UK Biobank, baseline CKD was significantly associated with an increased risk of incident AMD (hazard ratio, 1.12; 95% confidence interval [CI], 1.01–1.25; P = 0.035) over a 10-year follow-up. Mendelian randomization analyses also demonstrated causality between lower eGFR and higher risk of advanced AMD (odds ratio, 2.03; 95% CI, 1.01–4.08; P = 0.048). Colocalization analysis indicated that the apolipoprotein E gene may contribute to this causality (rs56131196; colocalization posterior probability = 1.00, P = 2.29 x 10-33 for AMD; P = 2.29 x 10-13 for eGFR).
Conclusions
Both prospective cohort and MR analyses support causality between CKD and AMD, highlighting the need for AMD screening among patients with CKD.
Financial Disclosure(s)
Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
目的探讨慢性肾脏疾病（CKD）与晚期老年性黄斑变性（AMD）风险的共同遗传影响和相关性。设计前瞻性队列研究和两样本孟德尔随机化（MR）分析。研究人员分析了来自英国生物银行队列43016名参与者的数据，以及来自最大的公开全基因组关联研究的汇总统计数据，这些研究涉及肾小球滤过率（eGFR） （n = 1 004 040）和晚期AMD （n = 33 976； 16 144例）。方法采用scox回归模型评估CKD与AMD发生率之间的关系，调整人口统计学、生活方式和临床协变量。在MR分析中，我们使用随机效应反方差加权模型作为主要模型，并使用另外5个MR模型进行敏感性分析。在主要模型和≥2个敏感性模型中，因果关系为被认为是显著的P <； 0.05，所有MR模型都显示出一致的影响方向。进行共定位分析以进一步确定连接CKD和AMD的共享遗传位点。eGFR与晚期AMD之间的因果关系。结果在UK Biobank中，基线CKD与AMD发生风险增加显著相关（风险比1.12；95%可信区间[CI]， 1.01-1.25; P = 0.035）。孟德尔随机化分析也显示eGFR较低与晚期AMD风险较高之间存在因果关系（优势比2.03;95% CI, 1.01-4.08; P = 0.048）。共定位分析表明载脂蛋白E基因可能与这种因果关系有关（rs56131196；共定位后验概率= 1.00，AMD的P = 2.29 x 10-33； eGFR的P = 2.29 x 10-13）。结论：前瞻性队列和MR分析均支持CKD和AMD之间的因果关系，强调CKD患者中AMD筛查的必要性。财务披露专有或商业披露可在本文末尾的脚注和披露中找到。
{"title":"Chronic Kidney Disease as a Risk Factor for Age-Related Macular Degeneration: A Prospective Cohort and Mendelian Randomization Analyses","authors":"Yu Jer Hsiao , Hengtong Li MSc , Can Can Xue PhD , Crystal Chun Yuen Chong , Enwen Zhu PhD , Qiang Yuan , Marco Yu PhD , Chui Ming Gemmy Cheung MD , Qiao Fan PhD , Charumathi Sabanayagam PhD , Yih-Chung Tham PhD , Ching-Yu Cheng MD, PhD","doi":"10.1016/j.xops.2025.101036","DOIUrl":"10.1016/j.xops.2025.101036","url":null,"abstract":"<div><h3>Purpose</h3><div>To evaluate shared genetic influences and investigate the association of chronic kidney disease (CKD) with the risk for advanced age-related macular degeneration (AMD).</div></div><div><h3>Design</h3><div>Prospective cohort study and 2-sample Mendelian randomization (MR) analyses.</div></div><div><h3>Participants</h3><div>Data from 430 016 participants in the UK Biobank cohort and summary statistics from the largest publicly available genome-wide association studies on estimated glomerular filtration rate (eGFR) (n = 1 004 040) and advanced AMD (n = 33 976; 16 144 cases) were analyzed.</div></div><div><h3>Methods</h3><div>Cox regression models were used to assess the association between CKD and incident AMD, adjusting for demographic, lifestyle, and clinical covariates. For MR analyses, we used the random-effects inverse-variance weighted model as the primary model, supported by 5 additional MR models for sensitivity analyses. A causal relationship was considered significant if <em>P</em> < 0.05 in the primary model and in ≥2 sensitivity models, with all MR models showing a consistent effect direction. Colocalization analysis was performed to further identify shared genetic loci linking CKD and AMD.</div></div><div><h3>Main Outcome Measures</h3><div>Causal associations between eGFR and advanced AMD.</div></div><div><h3>Results</h3><div>In the UK Biobank, baseline CKD was significantly associated with an increased risk of incident AMD (hazard ratio, 1.12; 95% confidence interval [CI], 1.01–1.25; <em>P</em> = 0.035) over a 10-year follow-up. Mendelian randomization analyses also demonstrated causality between lower eGFR and higher risk of advanced AMD (odds ratio, 2.03; 95% CI, 1.01–4.08; <em>P</em> = 0.048). Colocalization analysis indicated that the apolipoprotein E gene may contribute to this causality (rs56131196; colocalization posterior probability = 1.00, <em>P</em> = 2.29 x 10<sup>-33</sup> for AMD; <em>P</em> = 2.29 x 10<sup>-13</sup> for eGFR).</div></div><div><h3>Conclusions</h3><div>Both prospective cohort and MR analyses support causality between CKD and AMD, highlighting the need for AMD screening among patients with CKD.</div></div><div><h3>Financial Disclosure(s)</h3><div>Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.</div></div>","PeriodicalId":74363,"journal":{"name":"Ophthalmology science","volume":"6 2","pages":"Article 101036"},"PeriodicalIF":4.6,"publicationDate":"2025-12-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146037949","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：