Artificial intelligence (AI) can impact healthcare in many ways, from customizable treatment to routine screening and reducing barriers to care. In the field of ophthalmology, many diseases are diagnosed, monitored, and treated based on images, which may be enhanced by using AI.1
This article will discuss the use of AI in diabetic retinopathy and neovascular age-related macular degeneration care.
An overview of AI in diabetic retinopathy
Over 38 million people in the United States are living with type 2 diabetes mellitus (T2DM), and it is estimated that 9.6 million are living with diabetic retinopathy (DR).2,3 For those ages 20 to 74, diabetic retinopathy is the leading cause of new onset blindness. Both the prevalence of T2DM and DR are expected to increase dramatically by 2045.4
While early detection is critical in preventing vision loss, many of those living with diabetes do not receive the recommended frequency of eye examinations.5 To improve the screening burden placed on patients and healthcare providers, AI algorithms have been and continue to be developed to assist in screening.
Latest technologies for DR screening
EyeArt
EyeArt, which was FDA approved in 2020, utilizes two 45° field of view fundus photos to detect referable DR in 60 seconds. Images are classified as having no signs of DR, more than mild, vision-threatening, or ungradable.
If the images show:
- No signs of DR: Follow-up instructions direct them to get re-imaging done within 12 months.
- Ungradable: Follow-up instructions are to be referred to an eyecare specialist for a complete evaluation.
This technology is designed to be used by healthcare providers to screen patients in places such as diabetic centers or primary care offices.6 Numerous studies have shown a high sensitivity and negative predictive value of the EyeArt system, indicating it as a good screening tool for DR.7,8
EyeArt also has some limitations. One limitation is if one eye is classified as ungradable, the other eye will also be classified as ungradable despite the image quality. In validation studies, ungradable images were more likely if the patient was older.7 It also has a lower specificity and a high false positive rate, which may lead to over-referrals.
Many patients and primary care providers who have used EyeArt found it a positive experience. Patients stated it was time efficient, easy to use and would be interested in more screening. Primary care physicians reported the tool helped improve access to care, in particular to those who have barriers to receiving care, but stated more training was needed to use the system efficiently.9
IDx-DR
IDx-DR, approved by the FDA in 2018, is designed for the use of healthcare providers to screen for DR. IDx-DR uses 45° field of view photos of the macula and optic disc to detect DR. This technology classifies the images as more than mild DR, clinically significant diabetic macular edema (DME), or no signs of DR.
If the patient’s results are classified as more than mild DR or DME, they are referred to an ophthalmologist for further evaluation. If the screening shows no signs of DR, they must be re-screened in 1 year. When compared to retina specialist grading of fundus photos, IDx-DR has a higher sensitivity and lower specificity.
When utilizing IDx-DR as a screening tool, with referable images being overread by retina specialists, sensitivity remained at 95.5%, and specificity and gradability of images improved significantly. Utilizing a two-step AI-human model could reduce over-referrals and patient confusion.5
However, those whose images were screened solely by IDx-DR were found to have increased follow-up rates compared to those whose images were read by human-based teleophthalmology and the two-step AI-human model. Results generated by the AI model were given within 48 hours, while the other models took 7 days to return the results to the patients.
These results indicate that differences in time to results may impact follow-up care. Patients who did not follow up with the ophthalmology referral reported knowledge gaps, scheduling, and language as their biggest barriers. For those who did follow up with ophthalmology, when asked why they chose a specific location for follow-up, they were referred there by their primary care physician.
This emphasizes primary care physicians' impact on follow-up care and the importance of patient education to improve follow-up care rates. It also offers insights on improving follow-up care, such as contacting patients to schedule appointments in their preferred language.
AEYE: TOPCON and Aurora
The AEYE system was FDA-approved in 2022 and has two versions, with TOPCON, a desktop version, or with Aurora, a portable version. The desktop version has a sensitivity of 93% and a specificity of 91.4%. The portable version has a sensitivity of 91.9% and a specificity of 93.6%. Both versions had an imageability of 99%.10
SELENA for DR screening
In 2010, the Singapore Integrated Diabetic Retinopathy Retinopathy Program was developed to screen for retinal diseases, including DR and AMD. In 2019, this program began to utilize a deep learning tool, SELENA, for DR screening. SELENA utilizes fundus photos and has shown high sensitivity and specificity in multi-ethnic cohorts.11
Google AI tool
Google has developed an AI tool for screening DR. Studies conducted in Thailand found the tool to have a specificity of 95.4% and a sensitivity of 91.4%. However, a challenge with this tool was low-quality images, which resulted in more ungradable images. The strength of these studies was the multi-ethnic cohort, indicating the generalizability of the results.12,13
Remidio non-mydriatic gundus
Remidio non-mydriatic fundus on the phone is currently FDA 510k approved and features a smartphone-based retinal camera that utilizes fundus photos, which are then analyzed offline. Images are classified as referable diabetic retinopathy (more severe than mild, with/without macular edema) or diabetic retinopathy.
In a study conducted in Mumbai, India, sensitivity was 100%, and specificity was 88.4% for detecting referable diabetic retinopathy. For detecting diabetic retinopathy, sensitivity was 85.2%, and specificity was 92%. This study shows the feasibility of using this tool in rural areas, which can help address the decreased access to specialists in rural communities.14
List of AI tools for detecting DR
Table 1: Comparison of currently available AI imaging systems for diabetic retinopathy.
| Modality | Image Type | FDA Approval Year |
|---|---|---|
| EyeArt | Fundus photo | 2020 |
| IDx-DR | Fundus photo | 2018 |
| AEYE | Fundus photo | 2022 |
| Singapore Integrated Diabetic Retinopathy Program (SELENA) | Fundus photo | 2022 |
| Remidio Non-Mydriatic Fundus on Phone (Remidio Innovative Solutions) | Fundus photo | N/A |
| Google AI-based tool | Fundus photo | N/A |
Table 1: Courtesy of authors.
AI imaging for neovascular AMD
Nearly 20 million Americans are living with age-related macular degeneration (AMD).15 Current treatment modalities include anti-vascular endothelial growth factor (VEGF) injections with a treatment and extended protocol.
Artificial intelligence offers an opportunity to customize treatment. Currently, at-home optical coherence tomography (OCT) scans are being developed for patients with nAMD.
Notal Vision Home OCT
Notal Vision Home OCT is designed for patients' at-home use and detects and volumetrically quantifies retinal fluid.16 OCT captures a 3mm x 3mm, 10° x 10° field of view image.17 A pilot study of daily images in four patients showed the feasibility of daily imaging and showed a high agreement of intraretinal fluid amounts between the machine and human graders.17
When expanded to 15 patients, agreement of intraretinal fluid levels between the machine and human grading was 83%.16 The Notal Vision Home OCT is time efficient, with scans taking around 40 seconds, and was reported to be easy to use by elderly patients with impaired vision.18
To validate home OCT, the Notal Vision Home OCT version 2.5 and version 3 were compared to commercial OCT models. This study utilized a larger cohort than the previous home OCT studies. Both models had a high percentage of successful images taken with positive and negative predictive values.
When comparing the home OCT to the commercial OCT, there was not a statistically significant difference in the fluid detection levels, indicating a high correlation of the fluid levels with the different models of OCT detection.18
Comparing the benefits of home and traditional OCT imaging
Overall, home OCT can be used to customize treatment regimens for patients, which may reduce the treatment burden patients face, long-term vision outcomes, and healthcare productivity.16 The Notal Vision Home OCT was FDA-approved in May 2024 and will be branded as SCANLY. This tool estimates the volume of hypo-reflective spaces in the image, and physicians can be notified if it is over the set threshold.
A 2-year large cohort study of 600 participants is being conducted as part of DRCRNet Protocol AO to evaluate this tool further.19 The DRCRNet Protocol AO is a randomized control trial comparing Home OCT to the traditional treat and extend protocol.
Those in the home OCT group will perform daily OCT scans of both eyes. Participants will need to return to the clinic for evaluation if their last treatment was 3 weeks or more significant and if they had three consecutive scans where the total retinal fluid exceeded their threshold, or if they had one scan that exceeded 10nL of total retinal fluid.
To ensure patient compliance, a specialist will call the participant within 3 days if they have not completed a scan for 2 consecutive days. If participants do not wish to continue with the Home OCT protocol, they are added to the treat and extend cohort.20
iPredict
iPredict, developed by iHealthScreen, utilizes a color fundus photo in a primary care physician’s office. The report classifies the image as referable or non-referable AMD, provides follow-up instructions, and gives a percentage chance of developing late AMD over the next 1 to 2 years.21 iPredict is currently seeking FDA approval.
RetInSight
RetInSight was FDA approved in 2022. This tool measures intraretinal, subretinal, and pigment epithelial detachment fluid levels in nanoliters. This AI algorithm visualizes and tracks disease progression, allowing customizable treatment approaches and disease management.22
A validation study utilizing OCT images at various time points over 5 years of 1,127 eyes found that RetInSight can precisely measure fluid volumes of OCT images and that results were comparable to human graders.23
Another validation study consisting of internal validation images and external validation images from the FILLY trial found that RetInSight accurately detected geographic atrophy in both sets of data and performed within the integrability of human graders.24 Both studies show the feasibility of using RetInSight for patient care.
List of AI tools for detecting neovascular AMD
Table 2: Comparison of currently available AI imaging systems for neovascular AMD.
| Modality | Image Type | FDA Approval Year |
|---|---|---|
| SCANLY | Spectral domain OCT (SD-OCT) image | 2024 |
| iPredict | Fundus photo | N/A |
| RetInSight | OCT image | 2022 |
Table 2: Courtesy of authors.
Conclusions
Artificial intelligence in the vitreoretinal space can help improve screening rates for many diseases, decrease treatment burden, and offer customized treatment.
Patients can receive early and tailored healthcare by integrating AI into community centers, primary care offices, and at home.
