Skip to main content
eScholarship
Open Access Publications from the University of California

School of Medicine

Department of Ophthalmology - Open Access Policy Deposits bannerUC San Diego

This series is automatically populated with publications deposited by UC San Diego School of Medicine Department of Ophthalmology researchers in accordance with the University of California’s open access policies. For more information see Open Access Policy Deposits and the UC Publication Management System.

Cover page of Glaucoma Detection and Feature Identification via GPT-4V Fundus Image Analysis

Glaucoma Detection and Feature Identification via GPT-4V Fundus Image Analysis

(2025)

Purpose

The aim is to assess GPT-4V's (OpenAI) diagnostic accuracy and its capability to identify glaucoma-related features compared to expert evaluations.

Design

Evaluation of multimodal large language models for reviewing fundus images in glaucoma.

Subjects

A total of 300 fundus images from 3 public datasets (ACRIMA, ORIGA, and RIM-One v3) that included 139 glaucomatous and 161 nonglaucomatous cases were analyzed.

Methods

Preprocessing ensured each image was centered on the optic disc. GPT-4's vision-preview model (GPT-4V) assessed each image for various glaucoma-related criteria: image quality, image gradability, cup-to-disc ratio, peripapillary atrophy, disc hemorrhages, rim thinning (by quadrant and clock hour), glaucoma status, and estimated probability of glaucoma. Each image was analyzed twice by GPT-4V to evaluate consistency in its predictions. Two expert graders independently evaluated the same images using identical criteria. Comparisons between GPT-4V's assessments, expert evaluations, and dataset labels were made to determine accuracy, sensitivity, specificity, and Cohen kappa.

Main outcome measures

The main parameters measured were the accuracy, sensitivity, specificity, and Cohen kappa of GPT-4V in detecting glaucoma compared with expert evaluations.

Results

GPT-4V successfully provided glaucoma assessments for all 300 fundus images across the datasets, although approximately 35% required multiple prompt submissions. GPT-4V's overall accuracy in glaucoma detection was slightly lower (0.68, 0.70, and 0.81, respectively) than that of expert graders (0.78, 0.80, and 0.88, for expert grader 1 and 0.72, 0.78, and 0.87, for expert grader 2, respectively), across the ACRIMA, ORIGA, and RIM-ONE datasets. In Glaucoma detection, GPT-4V showed variable agreement by dataset and expert graders, with Cohen kappa values ranging from 0.08 to 0.72. In terms of feature detection, GPT-4V demonstrated high consistency (repeatability) in image gradability, with an agreement accuracy of ≥89% and substantial agreement in rim thinning and cup-to-disc ratio assessments, although kappas were generally lower than expert-to-expert agreement.

Conclusions

GPT-4V shows promise as a tool in glaucoma screening and detection through fundus image analysis, demonstrating generally high agreement with expert evaluations of key diagnostic features, although agreement did vary substantially across datasets.

Financial disclosures

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Cover page of Automated Quantitative Assessment of Retinal Vascular Tortuosity in Patients with Sickle Cell Disease

Automated Quantitative Assessment of Retinal Vascular Tortuosity in Patients with Sickle Cell Disease

(2025)

Objective

To quantitatively assess the retinal vascular tortuosity of patients with sickle cell disease (SCD) and retinopathy (SCR) using an automated deep learning (DL)-based pipeline.

Design

Cross-sectional study.

Subjects

Patients diagnosed with SCD and screened for SCR at an academic eye center between January 2015 and November 2022 were identified using electronic health records. Eyes of unaffected matched patients (i.e., no history of SCD, hypertension, diabetes mellitus, or retinal occlusive disorder) served as controls.

Methods

For each patient, demographic data, sickle cell diagnosis, types and total number of sickle cell crises, SCD medications used, ocular and systemic comorbidities, and history of intraocular treatment were extracted. A previously published DL algorithm was used to calculate retinal microvascular tortuosity using ultrawidefield pseudocolor fundus imaging among patients with SCD vs. controls.

Main outcome measures

Cumulative tortuosity index (CTI).

Results

Overall, 64 patients (119 eyes) with SCD and 57 age- and race-matched controls (106 eyes) were included. The majority of the patients with SCD were females (65.6%) and of Black or African descent (78.1%), with an average age of 35.1 ± 20.1 years. The mean number of crises per patient was 3.4 ± 5.2, and the patients took 0.7 ± 0.9 medications. The mean CTI for eyes with SCD was higher than controls (1.06 ± vs. 1.03 ± 0.02, P < 0.001). On subgroup analysis, hemoglobin S, hemoglobin C, and HbS/beta-thalassemia variants had significantly higher CTIs compared with controls (1.07 vs. 1.03, P < 0.001), but not with sickle cell trait variant (1.04 vs. 1.03 control, P = .2). Univariable analysis showed a higher CTI in patients diagnosed with proliferative SCR, most significantly among those with sea-fan neovascularization (1.06 ± 0.02 vs. 1.04 ± 0.01, P < 0.001) and those with >3 sickle cell crises (1.07 ± 0.02 vs. 1.05 ± 0.02, P < 0.001).

Conclusions

A DL-based metric of cumulative vascular tortuosity associates with and may be a potential biomarker for SCD and SCR disease severity.

Financial disclosures

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Cover page of Implementing a Common Data Model in Ophthalmology: Mapping Structured Electronic Health Record Ophthalmic Examination Data to Standard Vocabularies

Implementing a Common Data Model in Ophthalmology: Mapping Structured Electronic Health Record Ophthalmic Examination Data to Standard Vocabularies

(2025)

Objective

To identify and characterize concept coverage gaps of ophthalmology examination data elements within the Cerner Millennium electronic health record (EHR) implementations by the Observational Health Data Sciences and Informatics Observational Medical Outcomes Partnership (OMOP) common data model (CDM).

Design

Analysis of data elements in EHRs.

Subjects

Not applicable.

Methods

Source eye examination data elements from the default Cerner Model Experience EHR and a local implementation of the Cerner Millennium EHR were extracted, classified into one of 8 subject categories, and mapped to the semantically closest standard concept in the OMOP CDM. Mappings were categorized as exact, if the data element and OMOP concept represented equivalent information, wider, if the OMOP concept was missing conceptual granularity, narrower, if the OMOP concept introduced excess information, and unmatched, if no standard concept adequately represented the data element. Descriptive statistics and qualitative analysis were used to describe the concept coverage for each subject category.

Main outcome measures

Concept coverage gaps in 8 ophthalmology subject categories of data elements by the OMOP CDM.

Results

There were 409 and 947 ophthalmology data elements in the default and local Cerner modules, respectively. Of the 409 mappings in the default Cerner module, 25% (n = 102) were exact, 53% (n = 217) were wider, 3% (n = 11) were narrower, and 19% (n = 79) were unmatched. In the local Cerner module, 18% (n = 173) of mappings were exact, 54% (n = 514) were wider, 1% (n = 10) were narrower, and 26% (n = 250) were unmatched. The largest coverage gaps were seen in the local Cerner module under the visual acuity, sensorimotor testing, and refraction categories, with 95%, 95%, and 81% of data elements in each respective category having mappings that were not exact. Concept coverage gaps spanned all 8 categories in both EHR implementations.

Conclusions

Considerable coverage gaps by the OMOP CDM exist in all areas of the ophthalmology examination, which should be addressed to improve the OMOP CDM's effectiveness in ophthalmic research. We identify specific subject categories that may benefit from increased granularity in the OMOP CDM and provide suggestions for facilitating consistency of standard concepts, with the goal of improving data standards in ophthalmology.

Financial disclosures

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Cover page of A case of iatrogenic CNV following macular surgery.

A case of iatrogenic CNV following macular surgery.

(2025)

PURPOSE: To report a case of iatrogenic trauma related choroidal neovascularization (CNV). METHODS: A 66 year old female presented with complaints of distortion of vision in the right eye. A diagnosis of epiretinal membrane (ERM) foveoschisis was made and the patient was recommended surgery. During surgery an inadvertent touch of the retina occurred. Post operatively, an iatrogenic choroidal neovascular membrane was diagnosed. RESULTS: Post-operative optical coherence tomography (OCT) and optical coherence tomography angiography (OCTA) revealed a CNV which was treated with intravitreal Bevacizumab. The visual acuity improved remarkably. The intraretinal and subretinal fluid resolved and the macula was almost completely dry at final visit. CONCLUSION: ERM surgery is one of the common procedures in vitreoretinal surgery. Iatrogenic CNV is a reported but rare complication following inadvertent trauma during surgery. We report a successful outcome of iatrogenic CNV treated with intravitreal injections of Bevacizumab.

Cover page of Robotic Visible-Light Optical Coherence Tomography Visualizes Segmental Schlemm's Canal Anatomy and Segmental Pilocarpine Response

Robotic Visible-Light Optical Coherence Tomography Visualizes Segmental Schlemm's Canal Anatomy and Segmental Pilocarpine Response

(2025)

Purpose

To use robotic visible-light optical coherence tomography (vis-OCT) to study circumferential segmental Schlemm's canal (SC) anatomy in mice after topical pilocarpine administration.

Methods

Anterior segment imaging using a robotic vis-OCT to maintain perpendicular laser illumination aimed at SC was performed. Sixteen mice were studied for repeatability testing and to study aqueous humor outflow (AHO) pathway response to topical drug. Pharmaceutical-grade pilocarpine (1%; n = 5) or control artificial tears (n = 9) were given, and vis-OCT imaging was performed before and 15 minutes after drug application. SC areas and volumes were measured circumferentially.

Results

Circumferential vis-OCT provided high-resolution imaging of the AHO pathways. Segmental SC anatomy was visualized with the average cross-sectional area greatest temporal (3971 ± 328 µm2) and the least nasal (2727 ± 218 µm2; P = 0.018). After pilocarpine administration, the SC became larger (pilocarpine, 26.8 ± 5.0% vs. control, 8.9 ± 4.6% volume increase; P = 0.030). However, the pilocarpine alteration was also segmental, with a greater increase observed superior (pilocarpine, 31.6 ± 8.9% vs. control, 1.8 ± 5.7% volume increase; P = 0.023) and nasal (pilocarpine, 41.1 ± 15.3% vs. control, 13.9 ± 4.5% volume increase; P = 0.045).

Conclusions

Circumferential noninvasive imaging of the AHO pathways was done in vivo. Segmental SC anatomy was seen, consistent with the known segmental nature of trabecular AHO. Segmental SC anatomical response to a muscarinic agonist was also observed. Segmental glaucoma drug response around the circumference of AHO pathways is an observation that may influence patient response to glaucoma treatments.

Cover page of Ablation of Htra1 leads to sub-RPE deposits and photoreceptor abnormalities

Ablation of Htra1 leads to sub-RPE deposits and photoreceptor abnormalities

(2025)

The high-temperature requirement A1 (HTRA1), a serine protease, has been demonstrated to play a pivotal role in the extracellular matrix (ECM) and has been reported to be associated with the pathogenesis of age-related macular degeneration (AMD). To delineate its role in the retina, the phenotype of homozygous Htra1-KO (Htra1-/-) mice was characterized to examine the effect of Htra1 loss on the retina and retinal pigment epithelium (RPE) with age. The ablation of Htra1 led to a significant reduction in rod and cone photoreceptor function, primary cone abnormalities followed by rods, and atrophy in the RPE compared with WT mice. Ultrastructural analysis of Htra1-/- mice revealed RPE and Bruch's membrane (BM) abnormalities, including the presence of sub-RPE deposits at 5 months (m) that progressed with age accompanied by increased severity of pathology. Htra1-/- mice also displayed alterations in key markers for inflammation, autophagy, and lipid metabolism in the retina. These results highlight the crucial role of HTRA1 in the retina and RPE. Furthermore, this study allows for the Htra1-/- mouse model to be utilized for deciphering mechanisms that lead to sub-RPE deposit phenotypes including AMD.

Cover page of Cross-sectional design and protocol for Artificial Intelligence Ready and Equitable Atlas for Diabetes Insights (AI-READI)

Cross-sectional design and protocol for Artificial Intelligence Ready and Equitable Atlas for Diabetes Insights (AI-READI)

(2025)

Introduction

Artificial Intelligence Ready and Equitable for Diabetes Insights (AI-READI) is a data collection project on type 2 diabetes mellitus (T2DM) to facilitate the widespread use of artificial intelligence and machine learning (AI/ML) approaches to study salutogenesis (transitioning from T2DM to health resilience). The fundamental rationale for promoting health resilience in T2DM stems from its high prevalence of 10.5% of the world's adult population and its contribution to many adverse health events.

Methods

AI-READI is a cross-sectional study whose target enrollment is 4000 people aged 40 and older, triple-balanced by self-reported race/ethnicity (Asian, black, Hispanic, white), T2DM (no diabetes, pre-diabetes and lifestyle-controlled diabetes, diabetes treated with oral medications or non-insulin injections and insulin-controlled diabetes) and biological sex (male, female) (Clinicaltrials.org approval number STUDY00016228). Data are collected in a multivariable protocol containing over 10 domains, including vitals, retinal imaging, electrocardiogram, cognitive function, continuous glucose monitoring, physical activity, home air quality, blood and urine collection for laboratory testing and psychosocial variables including social determinants of health. There are three study sites: Birmingham, Alabama; San Diego, California; and Seattle, Washington.

Ethics and dissemination

AI-READI aims to establish standards, best practices and guidelines for collection, preparation and sharing of the data for the purposes of AI/ML, including guidance from bioethicists. Following Findable, Accessible, Interoperable, Reusable principles, AI-READI can be viewed as a model for future efforts to develop other medical/health data sets targeted for AI/ML. AI-READI opens the door for novel insights in understanding T2DM salutogenesis. The AI-READI Consortium are disseminating the principles and processes of designing and implementing the AI-READI data set through publications. Those who download and use AI-READI data are encouraged to publish their results in the scientific literature.

Cover page of Deep Learning Approach Predicts Longitudinal Retinal Nerve Fiber Layer Thickness Changes.

Deep Learning Approach Predicts Longitudinal Retinal Nerve Fiber Layer Thickness Changes.

(2025)

This study aims to develop deep learning (DL) models to predict the retinal nerve fiber layer (RNFL) thickness changes in glaucoma, facilitating the early diagnosis and monitoring of disease progression. Using the longitudinal data from two glaucoma studies (Diagnostic Innovations in Glaucoma Study (DIGS) and African Descent and Glaucoma Evaluation Study (ADAGES)), we constructed models using optical coherence tomography (OCT) scans from 251 participants (437 eyes). The models were trained to predict the RNFL thickness at a future visit based on previous scans. We evaluated four models: linear regression (LR), support vector regression (SVR), gradient boosting regression (GBR), and a custom 1D convolutional neural network (CNN). The GBR model achieved the best performance in predicting pointwise RNFL thickness changes (MAE = 5.2 μm, R2 = 0.91), while the custom 1D CNN excelled in predicting changes to average global and sectoral RNFL thickness, providing greater resolution and outperforming the traditional models (MAEs from 2.0-4.2 μm, R2 from 0.94-0.98). Our custom models used a novel approach that incorporated longitudinal OCT imaging to achieve consistent performance across different demographics and disease severities, offering potential clinical decision support for glaucoma diagnosis. Patient-level data splitting enhances the evaluation robustness, while predicting detailed RNFL thickness provides a comprehensive understanding of the structural changes over time.

Cover page of Comparison of a Novel Ultra-Widefield Three-Color Scanning Laser Ophthalmoscope to Other Retinal Imaging Modalities in Chorioretinal Lesion Imaging

Comparison of a Novel Ultra-Widefield Three-Color Scanning Laser Ophthalmoscope to Other Retinal Imaging Modalities in Chorioretinal Lesion Imaging

(2025)

Purpose

To compare the assessment of clinically relevant retinal and choroidal lesions as well as optic nerve pathologies using a novel three-wavelength ultra-widefield (UWF) scanning laser ophthalmoscope with established retinal imaging techniques for ophthalmoscopic imaging.

Methods

Eighty eyes with a variety of retinal and choroidal lesions were assessed on the same time point using Topcon color fundus photography (CFP) montage, Optos red/green (RG), Heidelberg SPECTRALIS MultiColor 55-color montage (MCI), and novel Optos red/green/blue (RGB). Paired images of the optic nerve, retinal, or choroidal lesions were initially diagnosed based on CFP imaging. The accuracy of the imaging was then evaluated in comparison to CFP using a grading scale ranging from -1 (losing imaging information) to +1 (gaining imaging information).

Results

Eighty eyes of 43 patients with 116 retinal or choroidal pathologies, as well as 59 eyes with optic nerve imaging using CFP, MCI, RG, and RGB, were included in this study. Across all subgroups, RGB provided significantly more accurate clinical imaging with CFP as ground truth and compared to other modalities. This was true comparing RGB to both RG (P = 0.0225) and MCI (P < 0.001) overall. Although RGB provided more accurate clinical information overall, it was inferior to RG for melanocytic choroidal lesions (P = 0.011).

Conclusions

RGB can be considered as a useful tool to detect characteristics of central, midperipheral, and peripheral retinal lesions. Regarding melanocytic choroidal lesions, RGB was inferior to RG, and MCI was inferior to both RG and RGB modalities due to color changes.

Translational relevance

Traditional retinal ultra-widefield imaging uses two wavelengths. Here, we evaluated three wavelengths for ultra-widefield imaging. We examined new optics (basic science) effect on patient imaging (clinical care).

Cover page of Horizontal Gaze Tolerance and Its Effects on Visual Sensitivity in Glaucoma.

Horizontal Gaze Tolerance and Its Effects on Visual Sensitivity in Glaucoma.

(2025)

PURPOSE: This study evaluates the effect of 6° horizontal gaze tolerance on visual field mean sensitivity (MS) in patients with glaucoma using a binocular head-mounted automated perimeter, following findings of structural changes in the posterior globe from magnetic resonance imaging and optical coherence tomography. METHODS: In this cross-sectional study, a total of 161 eyes (85 primary open-angle glaucoma [POAG] and 76 healthy) from 117 participants were included. Logistic regression and 1:1 matched analysis assessed the propensity score for glaucoma and healthy eyes, considering age, sex, and axial length as confounders. Visual field tests were performed with the imo perimeter (CREWT Medical Systems, Inc., Tokyo, Japan) at central gaze, 6° abduction, and 6° adduction positions as fixation points. A mixed-effects model was used to compare MS under all conditions. RESULTS: The analysis included a total of 82 eyes, with 41 POAG and 41 healthy after matching. The mean (standard deviation) age was 68.0 (11.0) years, with a mean deviation of -9.9 (6.6) dB for POAG and -1.0 (1.9) dB for healthy eyes using Humphrey field analysis 24-2. MS did not significantly differ among central gaze (27.0 [1.8] dB), abduction (27.1 [1.9] dB), and adduction (26.9 [2.2] dB) in healthy eyes (P = 0.650). However, MS was significantly lower for adduction (17.2 [5.9] dB) compared to central gaze (18.1 [5.9] dB) and abduction (17.9 [5.9] dB) in glaucoma eyes (P = 0.001 and P = 0.022, respectively). CONCLUSIONS: Horizontal gaze, especially in adduction, significantly reduces visual sensitivity in glaucoma, suggesting a specific vulnerability associated with eye movement. This finding highlights the importance of eye positioning in glaucoma, warranting further investigation of its clinical significance.