Foundation models in ophthalmology.

Foundation models represent a paradigm shift in artificial intelligence (AI), evolving from narrow models designed for specific tasks to versatile, generalisable models adaptable to a myriad of diverse applications. Ophthalmology as a specialty has the potential to act as an exemplar for other medical specialties, offering a blueprint for integrating foundation models broadly into clinical practice. This review hopes to serve as a roadmap for eyecare professionals seeking to better understand foundation models, while equipping readers with the tools to explore the use of foundation models in their own research and practice. We begin by outlining the key concepts and technological advances which have enabled the development of these models, providing an overview of novel training approaches and modern AI architectures. Next, we summarise existing literature on the topic of foundation models in ophthalmology, encompassing progress in vision foundation models, large language models and large multimodal models. Finally, we outline major challenges relating to privacy, bias and clinical validation, and propose key steps forward to maximise the benefit of this powerful technology.

Validation of a deep learning system for the detection of diabetic retinopathy in Indigenous Australians.

BACKGROUND/AIMS: Deep learning systems (DLSs) for diabetic retinopathy (DR) detection show promising results but can underperform in racial and ethnic minority groups, therefore external validation within these populations is critical for health equity. This study evaluates the performance of a DLS for DR detection among Indigenous Australians, an understudied ethnic group who suffer disproportionately from DR-related blindness. METHODS: We performed a retrospective external validation study comparing the performance of a DLS against a retinal specialist for the detection of more-than-mild DR (mtmDR), vision-threatening DR (vtDR) and all-cause referable DR. The validation set consisted of 1682 consecutive, single-field, macula-centred retinal photographs from 864 patients with diabetes (mean age 54.9 years, 52.4% women) at an Indigenous primary care service in Perth, Australia. Three-person adjudication by a panel of specialists served as the reference standard. RESULTS: For mtmDR detection, sensitivity of the DLS was superior to the retina specialist (98.0% (95% CI, 96.5 to 99.4) vs 87.1% (95% CI, 83.6 to 90.6), McNemar's test p<0.001) with a small reduction in specificity (95.1% (95% CI, 93.6 to 96.4) vs 97.0% (95% CI, 95.9 to 98.0), p=0.006). For vtDR, the DLS's sensitivity was again superior to the human grader (96.2% (95% CI, 93.4 to 98.6) vs 84.4% (95% CI, 79.7 to 89.2), p<0.001) with a slight drop in specificity (95.8% (95% CI, 94.6 to 96.9) vs 97.8% (95% CI, 96.9 to 98.6), p=0.002). For all-cause referable DR, there was a substantial increase in sensitivity (93.7% (95% CI, 91.8 to 95.5) vs 74.4% (95% CI, 71.1 to 77.5), p<0.001) and a smaller reduction in specificity (91.7% (95% CI, 90.0 to 93.3) vs 96.3% (95% CI, 95.2 to 97.4), p<0.001). CONCLUSION: The DLS showed improved sensitivity and similar specificity compared with a retina specialist for DR detection. This demonstrates its potential to support DR screening among Indigenous Australians, an underserved population with a high burden of diabetic eye disease.

Capabilities of GPT-4 in ophthalmology: an analysis of model entropy and progress towards human-level medical question answering.

BACKGROUND: Evidence on the performance of Generative Pre-trained Transformer 4 (GPT-4), a large language model (LLM), in the ophthalmology question-answering domain is needed. METHODS: We tested GPT-4 on two 260-question multiple choice question sets from the Basic and Clinical Science Course (BCSC) Self-Assessment Program and the OphthoQuestions question banks. We compared the accuracy of GPT-4 models with varying temperatures (creativity setting) and evaluated their responses in a subset of questions. We also compared the best-performing GPT-4 model to GPT-3.5 and to historical human performance. RESULTS: GPT-4-0.3 (GPT-4 with a temperature of 0.3) achieved the highest accuracy among GPT-4 models, with 75.8% on the BCSC set and 70.0% on the OphthoQuestions set. The combined accuracy was 72.9%, which represents an 18.3% raw improvement in accuracy compared with GPT-3.5 (p<0.001). Human graders preferred responses from models with a temperature higher than 0 (more creative). Exam section, question difficulty and cognitive level were all predictive of GPT-4-0.3 answer accuracy. GPT-4-0.3's performance was numerically superior to human performance on the BCSC (75.8% vs 73.3%) and OphthoQuestions (70.0% vs 63.0%), but the difference was not statistically significant (p=0.55 and p=0.09). CONCLUSION: GPT-4, an LLM trained on non-ophthalmology-specific data, performs significantly better than its predecessor on simulated ophthalmology board-style exams. Remarkably, its performance tended to be superior to historical human performance, but that difference was not statistically significant in our study.

A foundation model for generalizable disease detection from retinal images.

Medical artificial intelligence (AI) offers great potential for recognizing signs of health conditions in retinal images and expediting the diagnosis of eye diseases and systemic disorders1. However, the development of AI models requires substantial annotation and models are usually task-specific with limited generalizability to different clinical applications2. Here, we present RETFound, a foundation model for retinal images that learns generalizable representations from unlabelled retinal images and provides a basis for label-efficient model adaptation in several applications. Specifically, RETFound is trained on 1.6 million unlabelled retinal images by means of self-supervised learning and then adapted to disease detection tasks with explicit labels. We show that adapted RETFound consistently outperforms several comparison models in the diagnosis and prognosis of sight-threatening eye diseases, as well as incident prediction of complex systemic disorders such as heart failure and myocardial infarction with fewer labelled data. RETFound provides a generalizable solution to improve model performance and alleviate the annotation workload of experts to enable broad clinical AI applications from retinal imaging.

The Association of Physical Activity with Glaucoma and Related Traits in the UK Biobank.

PURPOSE: To examine the association of physical activity (PA) with glaucoma and related traits, to assess whether genetic predisposition to glaucoma modified these associations, and to probe causal relationships using Mendelian randomization (MR). DESIGN: Cross-sectional observational and gene-environment interaction analyses in the UK Biobank. Two-sample MR experiments using summary statistics from large genetic consortia. PARTICIPANTS: UK Biobank participants with data on self-reported or accelerometer-derived PA and intraocular pressure (IOP; n = 94 206 and n = 27 777, respectively), macular inner retinal OCT measurements (n = 36 274 and n = 9991, respectively), and glaucoma status (n = 86 803 and n = 23 556, respectively). METHODS: We evaluated multivariable-adjusted associations of self-reported (International Physical Activity Questionnaire) and accelerometer-derived PA with IOP and macular inner retinal OCT parameters using linear regression and with glaucoma status using logistic regression. For all outcomes, we examined gene-PA interactions using a polygenic risk score (PRS) that combined the effects of 2673 genetic variants associated with glaucoma. MAIN OUTCOME MEASURES: Intraocular pressure, macular retinal nerve fiber layer (mRNFL) thickness, macular ganglion cell-inner plexiform layer (mGCIPL) thickness, and glaucoma status. RESULTS: In multivariable-adjusted regression models, we found no association of PA level or time spent in PA with glaucoma status. Higher overall levels and greater time spent in higher levels of both self-reported and accelerometer-derived PA were associated positively with thicker mGCIPL (P < 0.001 for trend for each). Compared with the lowest quartile of PA, participants in the highest quartiles of accelerometer-derived moderate- and vigorous-intensity PA showed a thicker mGCIPL by +0.57 µm (P < 0.001) and +0.42 µm (P = 0.005). No association was found with mRNFL thickness. High overall level of self-reported PA was associated with a modestly higher IOP of +0.08 mmHg (P = 0.01), but this was not replicated in the accelerometry data. No associations were modified by a glaucoma PRS, and MR analyses did not support a causal relationship between PA and any glaucoma-related outcome. CONCLUSIONS: Higher overall PA level and greater time spent in moderate and vigorous PA were not associated with glaucoma status but were associated with thicker mGCIPL. Associations with IOP were modest and inconsistent. Despite the well-documented acute reduction in IOP after PA, we found no evidence that high levels of habitual PA are associated with glaucoma status or IOP in the general population. FINANCIAL DISCLOSURE(S): Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

The Association of Alcohol Consumption with Glaucoma and Related Traits: Findings from the UK Biobank.

PURPOSE: To examine the associations of alcohol consumption with glaucoma and related traits, to assess whether a genetic predisposition to glaucoma modified these associations, and to perform Mendelian randomization (MR) experiments to probe causal effects. DESIGN: Cross-sectional observational and gene-environment interaction analyses in the UK Biobank. Two-sample MR experiments using summary statistics from large genetic consortia. PARTICIPANTS: UK Biobank participants with data on intraocular pressure (IOP) (n = 109 097), OCT-derived macular inner retinal layer thickness measures (n = 46 236) and glaucoma status (n = 173 407). METHODS: Participants were categorized according to self-reported drinking behaviors. Quantitative estimates of alcohol intake were derived from touchscreen questionnaires and food composition tables. We performed a 2-step analysis, first comparing categories of alcohol consumption (never, infrequent, regular, and former drinkers) before assessing for a dose-response effect in regular drinkers only. Multivariable linear, logistic, and restricted cubic spline regression, adjusted for key sociodemographic, medical, anthropometric, and lifestyle factors, were used to examine associations. We assessed whether any association was modified by a multitrait glaucoma polygenic risk score. The inverse-variance weighted method was used for the main MR analyses. MAIN OUTCOME MEASURES: Intraocular pressure, macular retinal nerve fiber layer (mRNFL) thickness, macular ganglion cell-inner plexiform layer (mGCIPL) thickness, and prevalent glaucoma. RESULTS: Compared with infrequent drinkers, regular drinkers had higher IOP (+0.17 mmHg; P < 0.001) and thinner mGCIPL (-0.17 µm; P = 0.049), whereas former drinkers had a higher prevalence of glaucoma (odds ratio, 1.53; P = 0.002). In regular drinkers, alcohol intake was adversely associated with all outcomes in a dose-dependent manner (all P < 0.001). Restricted cubic spline regression analyses suggested nonlinear associations, with apparent threshold effects at approximately 50 g (∼6 UK or 4 US alcoholic units)/week for mRNFL and mGCIPL thickness. Significantly stronger alcohol-IOP associations were observed in participants at higher genetic susceptibility to glaucoma (Pinteraction < 0.001). Mendelian randomization analyses provided evidence for a causal association with mGCIPL thickness. CONCLUSIONS: Alcohol intake was consistently and adversely associated with glaucoma and related traits, and at levels below current United Kingdom (< 112 g/week) and United States (women, < 98 g/week; men, < 196 g/week) guidelines. Although we cannot infer causality definitively, these results will be of interest to people with or at risk of glaucoma and their advising physicians. FINANCIAL DISCLOSURE(S): Proprietary or commercial disclosure may be found after the references.

Utility of biomarkers and temporal artery biopsy length for investigating giant cell arteritis in Western Australia.

AIM: To explore demographic characteristics, biopsy length, and blood biomarker performance in an Australian cohort of patients who have undergone temporal artery biopsy (TAB) for giant cell arteritis (GCA). METHODS: We extracted data on biopsies performed for GCA between January 2016 and December 2020 at public hospitals in Perth. Sensitivity, specificity, and area under the curve (AUC) were calculated for blood results. We evaluated the proportion of biopsies with post-fixation length less than 15 mm and explored several length associations. RESULTS: We retrospectively reviewed biopsies of 360 patients (65.8% female, mean age 72.1 years). Biopsy-positive patients were older (6.0 years, P < 0.01), and had higher C-reactive protein (CRP) (44.5 mg/L, P < 0.01), erythrocyte sedimentation rate (ESR) (18.9 mm/h, P < 0.01), and platelets (86.8 × 103 /µL, P < 0.01) compared with biopsy-negative patients. CRP and platelets had the highest AUCs at 0.76 and 0.71, respectively. Sensitivities for CRP and ESR were 96.2% and 91.5%, respectively. Specificities were comparatively low at 41.3% for CRP and 37.4% for ESR. The proportion of biopsies with sub-optimal length was 55.9% and this varied significantly by site (P < 0.01). Smaller sites performed worse, with a sub-optimal biopsy rate of 87% amongst the three smallest sites. CONCLUSION: ESR and CRP are helpful preliminary investigations, especially in identifying low-risk patients, but their specificity is limited. Smaller centers had a higher proportion of biopsies with sub-optimal length. Considering the importance of biopsy length for TAB diagnostic value, reviewing biopsy data may assist services in developing improvement strategies.

Prevalence of Diabetic Retinopathy in Indigenous and Non-Indigenous Australians: A Systematic Review and Meta-analysis.

TOPIC: This systematic review and meta-analysis summarizes evidence relating to the prevalence of diabetic retinopathy (DR) among Indigenous and non-Indigenous Australians. CLINICAL RELEVANCE: Indigenous Australians suffer disproportionately from diabetes-related complications. Exploring ethnic variation in disease is important for equitable distribution of resources and may lead to identification of ethnic-specific modifiable risk factors. Existing DR prevalence studies comparing Indigenous and non-Indigenous Australians have shown conflicting results. METHODS: This study was conducted following Joanna Briggs Institute guidance on systematic reviews of prevalence studies (PROSPERO ID: CRD42022259048). We performed searches of Medline (Ovid), EMBASE, and Web of Science until October 2021, using a strategy designed by an information specialist. We included studies reporting DR prevalence among diabetic patients in Indigenous and non-Indigenous Australian populations. Two independent reviewers performed quality assessments using a 9-item appraisal tool. Meta-analysis and meta-regression were performed using double arcsine transformation and a random-effects model comparing Indigenous and non-Indigenous subgroups. RESULTS: Fifteen studies with 8219 participants met criteria for inclusion. The Indigenous subgroup scored lower on the appraisal tool than the non-Indigenous subgroup (mean score 50% vs. 72%, P = 0.04). In the unadjusted meta-analysis, DR prevalence in the Indigenous subgroup (30.2%; 95% confidence interval [CI], 24.9-35.7) did not differ significantly (P = 0.17) from the non-Indigenous subgroup (23.7%; 95% CI, 16.8-31.4). After adjusting for age and quality, DR prevalence was higher in the Indigenous subgroup (P < 0.01), with prevalence ratio point estimates ranging from 1.72 to 2.58, depending on the meta-regression model. For the secondary outcomes, prevalence estimates were higher in the Indigenous subgroup for diabetic macular edema (DME) (8.7% vs. 2.7%, P = 0.02) and vision-threatening DR (VTDR) (8.6% vs. 3.0%, P = 0.03) but not for proliferative DR (2.5% vs. 0.8%, P = 0.07). CONCLUSIONS: Indigenous studies scored lower for methodological quality, raising the possibility that systematic differences in research practices may be leading to underestimation of disease burden. After adjusting for age and quality, we found a higher DR prevalence in the Indigenous subgroup. This contrasts with a previous review that reported the opposite finding of lower DR prevalence using unadjusted pooled estimates. Future epidemiological work exploring DR burden in Indigenous communities should aim to address methodological weaknesses identified by this review.

AutoMorph: Automated Retinal Vascular Morphology Quantification Via a Deep Learning Pipeline.

Purpose: To externally validate a deep learning pipeline (AutoMorph) for automated analysis of retinal vascular morphology on fundus photographs. AutoMorph has been made publicly available, facilitating widespread research in ophthalmic and systemic diseases. Methods: AutoMorph consists of four functional modules: image preprocessing, image quality grading, anatomical segmentation (including binary vessel, artery/vein, and optic disc/cup segmentation), and vascular morphology feature measurement. Image quality grading and anatomical segmentation use the most recent deep learning techniques. We employ a model ensemble strategy to achieve robust results and analyze the prediction confidence to rectify false gradable cases in image quality grading. We externally validate the performance of each module on several independent publicly available datasets. Results: The EfficientNet-b4 architecture used in the image grading module achieves performance comparable to that of the state of the art for EyePACS-Q, with an F1-score of 0.86. The confidence analysis reduces the number of images incorrectly assessed as gradable by 76%. Binary vessel segmentation achieves an F1-score of 0.73 on AV-WIDE and 0.78 on DR HAGIS. Artery/vein scores are 0.66 on IOSTAR-AV, and disc segmentation achieves 0.94 in IDRID. Vascular morphology features measured from the AutoMorph segmentation map and expert annotation show good to excellent agreement. Conclusions: AutoMorph modules perform well even when external validation data show domain differences from training data (e.g., with different imaging devices). This fully automated pipeline can thus allow detailed, efficient, and comprehensive analysis of retinal vascular morphology on color fundus photographs. Translational Relevance: By making AutoMorph publicly available and open source, we hope to facilitate ophthalmic and systemic disease research, particularly in the emerging field of oculomics.

Benefits of Integrating Telemedicine and Artificial Intelligence Into Outreach Eye Care: Stepwise Approach and Future Directions.

Telemedicine has traditionally been applied within remote settings to overcome geographical barriers to healthcare access, providing an alternate means of connecting patients to specialist services. The coronavirus 2019 pandemic has rapidly expanded the use of telemedicine into metropolitan areas and enhanced global telemedicine capabilities. Through our experience of delivering real-time telemedicine over the past decade within a large outreach eye service, we have identified key themes for successful implementation which may be relevant to services facing common challenges. We present our journey toward establishing a comprehensive teleophthalmology model built on the principles of collaborative care, with a focus on delivering practical lessons for service design. Artificial intelligence is an emerging technology that has shown potential to further address resource limitations. We explore the applications of artificial intelligence and the need for targeted research within underserved settings in order to meet growing healthcare demands. Based on our rural telemedicine experience, we make the case that similar models may be adapted to urban settings with the aim of reducing surgical waitlists and improving efficiency.

Screening for Glaucomatous Visual Field Defects in Rural Australia with an iPad.

AIM AND OBJECTIVE: Developing improved methods for early detection of visual field defects is pivotal to reducing glaucoma-related vision loss. The Melbourne Rapid Fields screening module (MRF-S) is an iPad-based test, which allows suprathreshold screening with zone-based analysis to rapidly assess the risk of manifest glaucoma. The versatility of MRF-S has potential utility in rural areas and during infectious pandemics. This study evaluates the utility of MRF-S for detecting field defects in non-metropolitan settings. MATERIALS AND METHODS: This was a prospective, multicenter, cross-sectional validation study. Two hundred and fifty-two eyes of 142 participants were recruited from rural sites through two outreach eye services in Australia. Participants were tested using MRF-S and compared with a reference standard; either Zeiss Humphrey Field Analyzer or Haag-Streit Octopus performed at the same visit. Standardized questionnaires were used to assess user acceptability. Major outcome measures were the area under the curve (AUC) for detecting mild and moderate field defects defined by the reference tests, along with corresponding performance characteristics (sensitivity, specificity). RESULTS: The mean test duration for MRF-S was 1.88 minutes compared with 5.92 minutes for reference tests. The AUCs for mild and moderate field defects were 0.81 [95% confidence interval (CI): 0.75-0.87] and 0.87 (95% CI: 0.83-0.92), respectively, indicating very good diagnostic accuracy. Using a risk criterion of 55%, MRF-S identified moderate field defects with a sensitivity and specificity of 88.4 and 81.0%, respectively. CONCLUSION AND CLINICAL SIGNIFICANCE: The MRF-S iPad module can identify patients with mild and moderate field defects while delivering favorable user acceptability and short test duration. This has potential application within rural locations and amidst infectious pandemics. HOW TO CITE THIS ARTICLE: Chia MA, Trang E, Agar A, et al. Screening for Glaucomatous Visual Field Defects in Rural Australia with an iPad. J Curr Glaucoma Pract 2021;15(3):125-131.