# Clinical validation of artificial intelligence algorithms for the detection of different central-involved retinal pathologies and glaucoma from non-mydriatic images

## TL;DR

This study evaluates AI algorithms for detecting retinal diseases and glaucoma from non-mydriatic images in a real-world diabetes screening setting.

## Contribution

The study validates AI algorithms for multiple retinal pathologies and glaucoma detection in a real-world teleophthalmology program.

## Key findings

- AI algorithms achieved high sensitivity and specificity for detecting diabetic retinopathy, AMD, glaucomatous optic neuropathy, and other retinal conditions.
- The AI system accurately classified eye laterality and image quality, with perfect accuracy for eye laterality.
- The AUROC scores ranged from 0.91 to 0.98 across different pathologies, indicating strong diagnostic performance.

## Abstract

The use of Artificial intelligence (AI) algorithms for detecting different ophthalmic diseases, especially diabetic retinopathy (DR), has become increasingly popular. In this paper, we evaluate the screening performance of different AI algorithms based on convolutional neural networks (CNNs) in a real-world scenario. To that aim, we conducted an observational and cross-sectional study on patients aged ≥18 years with type-2 diabetes mellitus, who had undergone fundus examination for DR screening using a teleophthalmology program. We used the UPRETINA diagnostic system, which consists of 8 AI algorithms based on CNNs. A total of 1,652 eyes from 871 patients were analyzed. The AI algorithms had a sensitivity/specificity of 86.8%/95.6% for detecting DR; 94.9%/94.3% for detecting age-related macular degeneration (AMD); 82.7%/92.4% for detecting glaucomatous optic neuropathy (GON); 87.0%/87.5% for detecting epiretinal membrane; and 89.7%/98.0% for detecting nevus. Additionally, the sensitivity/specificity for correctly classifying images as right eye/left eye and to correctly classifying images gradeability (medium or high quality) were 100% /100 and 92.9%/90.5%, respectively. The AUROC of the AI algorithms ranged between 0.9777 (AMD) and 0.9122 (GON). UPRETINA system was capable of automatically and accurately classifying the screening retinographies, reducing workload and leading to a scenario of more efficient optimization of resources.

https://clinicaltrials.gov/study/NCT04132401 NCT04132401.