Table 2.
Description and notification of current AI systems
| AI system | Description and notification |
| IRIS (Intelligent Retinal Imaging Systems, Pensacola, FL, USA) | IRIS is an FDA class II cleared medical system that has a moderate risk to consumers and must demonstrate that it is “substantially equivalent” to similar products. The IRIS program is a cloud-based platform to screen for vision-threatening DR, with sensitivity and specificity of 66.4% and 72.8%, respectively [77] |
| ARDA (Google LLC, Mountain View, CA, USA) | ARDA is a DL algorithm developed by Google Health from >128 000 retinal photographs of patients from the United States and India and validated in >10 000 photographs from the UK to detect referrable and sight-threatening DR [7]. The validation study of ARDA was the first showing robust performance of DL to detect referrable DR with >95% of both sensitivity and specificity [7]. Later, ARDA was prospectively validated in India [78] and a nationwide screening program in Thailand [79] |
| SELENA+ (EyRIS Pte Ltd, Singapore) | The Singapore Eye Research Institute and Singapore National Eye Center has developed a DL-based algorithm, SELENA+, to screen for referable DR, vision-threatening DR, DR-related vascular risk factors, suspected glaucoma, and late-stage AMD. It is a multicenter collaborative research effort with half a million retinal images from people of different ethnicities such as Caucasians from Australia and the United States, and Singapore Chinese, Malayans, Indians, Chinese, individuals from Hong Kong, Mexicans, Hispanics, and African Americans. Real-world application and clinical translation of SELENA+ has been integrated into the Singapore Integrated Diabetic Retinopathy Programme in recent years. SELENA+ has significant diagnostic performance in DR, with sensitivity of 91%, specificity of 90%, and area under the curve of 0.93 [80] |
| IDx-DR (Digital Diagnostics Inc., Coralville, IA, USA) | IDx-DR was the first FDA-approved ophthalmic device to autonomously detect DR, including DME. It can analyze retinal images, detect vision-threatening DR, and provide referral recommendations [6,81]. The external validation to detect referable DR showed sensitivity and specificity of 91% and 84%, respectively [79] |
| Medios AI (Remidio Innovative Solutions Pvt Ltd., Karnataka, India) | Medios AI is an integrated offline system with a Remidio smartphone-based, nonmydriatic retinal camera to detect referable DR. The fundus images can be captured by minimally trained healthcare providers. The sensitivity and specificity of diagnosing referable DR were 100% and 88.4%, respectively [82] |
| RetCAD (Thirona Retina BV, Nijmegen, Netherlands) | This commercially available DL algorithm can determine referable DR and AMD based on a dataset of CFPs to reduce the workload of screening programs by up to 96%, with sensitivity of 90.53% and specificity of 97.13%. Patients’ CFPs can be captured by camera and then transferred to the Thirona server for analysis. The examination report will provide referable suggestions and visualization of heatmaps [83] |
| EyeArt (Eyenuk, Inc., Woodland Hills, CA, USA) | This cloud-based autonomous AI system can detect more-than-mild DR and vision-threatening DR by submitting fundus photography to the platform. It is designed to work with various types of retinal cameras. It assesses the quality of uploaded images and explains the reasons behind grading. This algorithm can provide the grading of DR and report the results for each eye based on the UK National Health Service diabetic eye screening program scale. The sensitivity and specificity showed 96% and 98%, respectively [84] |
| VUNO Med-Fundus AI (VUNO Inc., Seoul, Korea) | The AI-based VUNO Med-Fundus AI analyzes CFP to detect multiple retinal lesions (areas under receiver operating characteristic curves for all findings were at 96.2%) [85]. The area under the receiver operating characteristic curves for DR-related findings was 95%. It was approved as a class III medical device by the Ministry of Food and Drug Safety in Korea |
| THEIA (Toku Eyes, Auckland, New Zealand) | The New Zealand company Toku Eyes developed THEIA, an AI platform for cloud-based multimodal image analysis of referable DR and AMD. The THEIA system was developed from two of the largest screening data sets in Auckland, New Zealand: the Auckland District Health Board and the Counties Manukau District Health Board. It can analyze color fundus images, OCT, and OCT-A to provide results about referable DR (sensitivity of 93% and specificity of 63%) and intermediate dry AMD (accuracy of 96%) [86,87]. This AI system is considered to be useful in reducing the workload in the New Zealand National Diabetic Retinopathy Screening Program [88] |
| iPredict (iHealthScreen Inc., Richmond Hill, NY, USA) | The iPredict AI Eye Screening System offers fully automated diagnosis of referable DR (sensitivity of 97.0% and specificity of 96.3%) and AMD (sensitivity of 86.6% and specificity of 92.1%) by analyzing CFPs [89] |
| Notal Home OCT (Notal Vision, Inc., Manassas, VA, USA) | Notal Home OCT, the first FDA-cleared in-home OCT device, which includes an AI algorithm and monitoring center, is designed to detect AMD. The imaging quality showed great correlation with in-office OCT for detecting the presence of fluid in 95% agreement with human graders [90]. A patient's ability to use an in-home setting for self-imaging without training demonstrated good capacity with a 95% success rate [91] |
| OphtAI (Evolucare/ADCIS, Villers-Bretonneux, France) | OphtAI DR is a semiautomatic AI algorithm that assesses the pathologic lesions and grading of DR and detects AMD and glaucoma. In a multicenter, head-to-head, real-world validation study to compare different algorithms in detecting DR, the OphtAI DR algorithm provided better results (sensitivity of 80.47% and specificity of 81.28%) than an ophthalmologist [92]. It is also deemed clinically safe and economically efficient in reducing the costs by more than U.S. $15 per patient [92] |
| Retmarker (Retmarker, SA, Taveiro, Portugal) | This AI technology can provide screening for DR and AMD by annotating pathologic lesions, such as microaneurysms, drusen, hypopigmentation, hyperpigmentation, and geographic atrophy [93]. The sensitivity in classifying DR is 73.0% for any DR, 85.0% for referable DR, and 97.9% for proliferative DR [93]. The screening performance of Retmarker appeared to vary with patients’ age, ethnicity, and camera type. In economic analysis, the Retmarker was more cost effective than manual grading [93] |
| RetinaLyze (RetinaLyze System A/S, Hellerup, Denmark) | The RetinaLyze system is a screening software that can detect DR on nonmydriatic CFPs (sensitivity of 89.9% and specificity of 85.7%) [94]. It can detect DR lesions, including microaneurysms and minor hemorrhages (specificity of 71.4%) [95]. It can also evaluate biological aging [96] and hemoglobin on optic disc photographs [97] |
| RetinAI Discovery (RetinAI Medical AG, Bern, Switzerland) | The Discovery platform can analyze medical data and ophthalmic images such as OCT scans and CFP from a variety of devices. It can help automatically detect the location of the fovea (mean total location error of 0.101 mm), the quantification of pathologic fluid, and the segmentation of atrophic retina on OCT in patients with geographic atrophy [98–100]. It can detect and quantify fluid from DR, DME, AMD, and RVO. The performance of the AI system showed that the accuracy, specificity, and sensitivity for intraretinal fluid was 0.87, 0.88, 0.84 and 0.93, 0.95, 0.93 and for subretinal fluid was 0.93, 0.93, 0.93 and 0.95, 0.95, 0.95 in the AMD and DME cohorts, respectively [101▪,102] |
Note: All medical devices approved by the FDA or accredited by the CE mark from January 2015 to January 2023 were collected. These devices were searched for in the European Database on Medical Devices (EUDAMED) database [74], the FDA website (on the webpage of Artificial Intelligence and Machine Learning-Enabled Medical Devices) [75], and the FDA 510(k) Premarket Notification [76]. The approved devices were summarized and their performance in related trials was searched for in PubMed, with the data source cited as the reference.
ARDA, Automatic Retinal Disease Assessment; CFP, color fundus photograph; DME, diabetic macular edema; IRIS, Intelligent Retinal Imaging Systems; OCT, optical coherence tomography; OCT-A, optical coherence tomography angiography; RVO, retinal vein occlusion; SELENA, Singapore Eye LEsioN Analyzer.