For years, we have been hearing about the AI revolution that will transform the field of retina. Multiple AI algorithms for retinal diseases have received FDA clearance,1,2 including LumineticsCore (formerly IDx-DR, Digital Diagnostics), EyeART (EyeNuk), Altris AI (Altris), and AEYE-DS (AEYE), all of which are designed more for primary eye care settings than the retina clinic.3 Outside of the United States, several AI algorithms for biomarker identification have been cleared for clinical use. These newest developments exist alongside traditional computer vision methods for tasks such as OCT retinal layer segmentation, retinal thickness mapping, and retinal nerve fiber layer quantification.4

Although these innovations represent important steps in the application of AI to the field of retina, such technology has failed to make a significant effect on the day-to-day activities of most retina clinics. What prevents this technology—already shaping the cutting-edge consumer products that surround us—from revolutionizing the practice of retina? Here, we focus on five key hurdles to the development and implementation of AI in retina.

Hurdle No. 1: Data Limitations

The deep-learning algorithms that power today’s AI technologies rely on large, diverse datasets with high-quality, often labelled images.5 However, accessing ophthalmic data with these characteristics can be difficult. While there are several publicly available fundus photograph and OCT datasets for diseases such as diabetes and AMD, the data domain does not entirely overlap with that generated in a typical US retina clinic, which has hindered model accuracy in real-world settings.6-8

Data collected during routine patient care might better represent the model’s end use and could be advantageous for model training. However, real-world data can present challenges in data quality and heterogeneity. Imaging protocols that are expeditious but adequate for a busy retina clinic may lack the quality or completeness necessary for AI model training or inference. Images acquired under uniform parameters—including high-resolution scans and small step sizes—may improve model performance but can be difficult to implement in the real world due to increased acquisition times and disruptions in clinical workflow.

Hurdle No. 2: Clinical Integration

Even with high-quality data and effective algorithms, integration into clinical workflows is a major hurdle to AI implementation. AI tools must seamlessly fit into the existing processes of patient care without disrupting efficiency. Retina clinics already rely on many applications for the acquisition and display of clinical and imaging data. Therefore, AI systems that require separate interfaces and additional clicks will inevitably struggle to gain traction. Retina specialists should be involved early and often in device development to ensure that a high-performing device fits naturally into existing practice ecosystems.

Developers may be stymied in achieving this due to barriers of interoperability. Unlike other medical fields (eg, radiology), ophthalmology lacks a universal imaging standard. Additionally, electronic health record proprietors rarely have ready routes for integration of third-party AI modules. Adoption of Digital Imaging and Communications in Medicine (DICOM) and Fast Healthcare Interoperability Resources (FHIR) may improve device interoperability, and the ophthalmic community must continue to advocate through professional organizations for mandating these standards.

Hurdle No. 3: Practice Economics

Integrating AI programs into the clinical workflow requires time and resources; they may require specific clinical workflows, image types, and third-party applications. Practices must see a path to a return on their investment if they are expected to shoulder the burden of these costs. In a fee-for-service model, favorable reimbursement is the most immediate mechanism to support AI device implementation. In 2021, CPT code 92229 was introduced to reimburse the automated analysis of retinal images in the identification of diabetic retinopathy, typically using the devices mentioned above in a primary care setting.9 Billing codes have been approved for AI-enabled home OCT monitoring as well.10 However, there is no billing code that is used with frequency by retina specialists. Moreover, while billing codes are necessary, they may not be sufficient; insurers must also be willing to cover these costs.

Hurdle No. 4: Regulation

FDA regulation of AI technologies in health care is important to ensure algorithms are safe and accurate; those that are fully automated must meet a higher standard than those that are merely assistive to the doctor. However, the current regulatory framework for AI in medicine is still evolving.7 Additionally, as with drug development, regulation of AI can be costly and time intensive, particularly for first-in-class de novo approvals.

The field of radiology, a frontrunner in AI with hundreds of FDA-approved AI-based algorithms, serves as a good case study.11 Many of the radiology AI programs underwent approval through a 510(k) pathway, which involves a less burdensome review process if devices are substantially similar to previously approved devices.11 This suggests that approvals in ophthalmology will proliferate as initial de novo approvals are obtained.

Another unique consideration is that AI models can be dynamic; they can learn and evolve as they process more data. This adaptability raises regulatory questions about how to ensure patient safety and efficacy over the lifecycle of the AI system.12

Hurdle No. 5: Data Protection

AI systems in retina clinics will have access to sensitive patient information, including images, medical histories, and diagnostic results. This introduces risks related to data breaches and cyberattacks. Ensuring that patient data is encrypted and protected is paramount, particularly as AI systems become more integrated into clinical practice and electronic health records.13 Data may also be used by AI developers for applications unbeknownst to the data sharer, for instance, developing valuable intellectual property. Where strong legal protections seem insufficient, approaches such as federated learning may be necessary, whereby AI models may be applied to or trained on data without the data leaving its local network.6

FORGING AHEAD

While these challenges to the implementation of AI in retina are considerable, they can be overcome. In many ways, the most difficult problems—those of computing and technology—have already been solved. Ophthalmology, and retina specifically, has always been at the technological forefront in medicine, eager to adopt and adapt. We as a field should and will continue this leadership through collaboration with technology partners to overcome the implementation obstacles highlighted here. We are certain that our field will meet the moment and bring the benefits of AI to our patients.

1. Chia MA, Antaki F, Zhou Y, Turner AW, Lee AY, Keane PA. Foundation models in ophthalmology. Br J Ophthalmol. 2024;108(10):1341.

2. Danese C, Kale AU, Aslam T, et al. The impact of artificial intelligence on retinal disease management: Vision Academy retinal expert consensus. Curr Opin Ophthalmol. 2023;34(5):396-402.

3. Rajesh AE, Davidson OQ, Lee CS, Lee AY. Artificial intelligence and diabetic retinopathy: AI framework, prospective studies, head-to-head validation, and cost-effectiveness. Diabetes Care. 2023;46(10):1728-1739.

4. Mares V, Nehemy MB, Bogunovic H, Frank S, Reiter GS, Schmidt-Erfurth U. AI-based support for optical coherence tomography in age-related macular degeneration. In J Retina Vitreous. 2024;10(1):31.

5. Litjens G, Kooi T, Bejnordi BE, et al. A survey on deep learning in medical image analysis. Med Image Anal. 2017;42:60-88.

6. Lim JI, Rachitskaya AV, Hallak JA, Gholami S, Alam MN. Artificial intelligence for retinal diseases. Asia Pac J Ophthalmol. 2024;13(4):100096.

7. Kras A, Celi LA, Miller JB. Accelerating ophthalmic artificial intelligence research: the role of an open access data repository. Curr Opin Ophthalmol. 2020;31(5):337-350.

8. Kermany DS, Goldbaum M, Cai W, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell. 2018;172(5):1122-1131.e9.

9. Wolf RM, Channa R, Lehmann HP, Abramoff MD, Liu TYA. Clinical implementation of autonomous artificial intelligence systems for diabetic eye exams: considerations for success. Clin Diabetes. 2024;42(1):142-149.

10. CPT codes for home OCT established [press release]. Notal Vision. January 8, 2020. Accessed October 16, 2024. notalvision.com/index.php?p=actions/asset-count/count&id=23993

11. Muehlematter UJ, Bluethgen C, Vokinger KN. FDA-cleared artificial intelligence and machine learning-based medical devices and their 510(k) predicate networks. Lancet Digit Health. 2023;5(9):e618-e626.

12. Vokinger KN, Gasser U. Regulating AI in medicine in the United States and Europe. Nat Mach Intell. 2021;3(9):738-739.

13. Lim JS, Hong M, Lam WST, et al. Novel technical and privacy-preserving technology for artificial intelligence in ophthalmology. Curr Opin Ophthalmol. 2022;33(3):174-187.