Three AI studies map retinal screening from lab to clinic

Three recent studies illustrate how artificial intelligence is evolving across the full spectrum of retinal disease detection—from specialist reading centers and biomarker research to point-of-care screening in resource-limited settings.

The research, first reported by Ophthalmology Times, covers a foundation model that automates 3D optical coherence tomography (OCT) workflows, a systematic review of OCT angiography biomarkers for early diabetic retinopathy, and an offline smartphone platform tested in India for multi-disease screening.

Automating the full OCT workflow

Researchers at Zhongshan Ophthalmic Center, Sun Yat-sen University, developed FOCUS (Full-process OCT-based Clinical Utility System), described in npj Digital Medicine. The system automates the entire OCT pipeline: image quality assessment, pathology-relevant slice selection, and diagnostic interpretation.

FOCUS operates in two stages. An image quality module screens unusable OCT slices, then a diagnostic stage built on a Vision Foundation Model extracts features from individual 2D B-scans. A component called the Unified Adaptive Aggregation Classifier fuses slice-level predictions into patient-level diagnoses, weighting each slice by diagnostic relevance rather than treating all equally.

The system detects nine categories including AMD, diabetic retinopathy, macular hole, and retinitis pigmentosa. Trained on 3,300 patients and validated on 1,345 patients across four Chinese clinical centers using different OCT devices, FOCUS achieved F1-scores of 99.01% for quality assessment and 94.39% for patient-level diagnosis on internal testing, with external validation scores between 90.22% and 95.24%.

In head-to-head comparison, FOCUS matched retinal specialists in multi-disease diagnosis (F1-score 93.49% versus 91.35%) while processing 842 images in 22.7 seconds compared with 32.1 minutes for specialists.

Peng Xiao, professor at Zhongshan Ophthalmic Center and corresponding author, told Ophthalmology Times that FOCUS "changes the fundamental economics of OCT screening" by enabling deployment in primary care settings where trained interpreters are scarce. He emphasized the system functions as "an intelligent triage and quality-assurance layer" rather than a specialist replacement.

Biomarker validation remains incomplete

A systematic review by Loukia Politi and colleagues, published in the European Journal of Ophthalmology, examined 21 studies evaluating OCTA-derived biomarkers for early diabetic retinopathy detection. The review found vessel density (VD) to be the most consistent marker—13 of 15 studies reported reduced VD in diabetic eyes compared with controls.

However, other commonly used measures showed mixed results. Enlargement of the foveal avascular zone was reported in only nine of 17 studies. Only three studies reported formal diagnostic accuracy metrics, with sensitivities ranging from 77.2% to 93.62%.

The authors concluded that while VD shows promise, inconsistent imaging protocols and limited diagnostic accuracy studies prevent routine clinical implementation of OCTA for early diabetic retinopathy detection.

Offline smartphone screening in India

Aditya Kelkar and colleagues evaluated Medios AI, an offline system integrated into a smartphone-based fundus camera, in a study published in the European Journal of Ophthalmology. The prospective study enrolled 193 adults (371 eyes) at a tertiary eye care center in Pune, India, between May and December 2024.

The system screened for diabetic retinopathy, glaucoma, and AMD simultaneously using a single AI module. For detecting any retinal disease, Medios AI achieved 99.3% sensitivity and 95.7% specificity. Disease-specific performance varied: glaucoma (98.2% sensitivity), AMD (88.9% sensitivity), and diabetic retinopathy (84.6% sensitivity, rising to 95% for referrable cases). The system generates annotated reports in under 15 seconds without internet connectivity.

Corresponding author Sabyasachi Sengupta, consultant vitreoretinal surgeon at Future Vision Eye Care in Mumbai, told Ophthalmology Times that some reduction in disease-specific sensitivity is expected when systems move from single-disease to multi-disease detection. He noted that community-based validation faces challenges including small pupils and media haze, which could increase ungradable-image rates.

Why it matters

These studies demonstrate AI's expanding role across different clinical settings, but each highlights the same critical gap: broader validation is needed. FOCUS was developed primarily on Chinese populations, the OCTA review found inconsistent protocols limiting biomarker reliability, and Medios AI was tested in a high-prevalence referral center. Multi-center studies in diverse populations and lower-prevalence community settings will determine whether these tools can scale beyond their development environments—a necessary step before widespread clinical deployment.

Details were first reported by Ophthalmology Times.