Take-home points

  • Modern management of neovascular AMD focuses on individualized treatment decisions, requiring detailed monitoring of imaging biomarkers on OCT at every patient visit.
  • AI-based image analysis can reliably assess these biomarkers and, importantly, exceeds human capability by quantifying lesion extent in three dimensions.
  • Although specific clinical use cases for AI-powered image analysis already exist, broader clinical adoption will depend on further supporting evidence.
  • Future applications of AI go beyond detection and quantification; they include predictive modeling for visual outcomes and treatment needs, as well as data-driven discovery of novel subclinical imaging biomarkers.

Bios

Dr. Munk is a uveitis and medical retina specialist. She is chief scientific officer and head of research at Augenarzt-Praxisgemeinschaft Gutblick AG, a professor at Inselspital, University Hospital Bern, adjunct professor at Northwestern University in Chicago, director of Eyegnos Consulting and chief medical officer at Isarna Therapeutics.

Dr. Waldstein is a vitreoretinal surgeon and chair of ophthalmology at Mistelbach-Gänserndorf Hospital, Austria; medical director of Gänserndorf Medical Center and associate professor of ophthalmology at Karl Landsteiner University of Health Sciences in Krems, Austria.

Disclosures: Pertinent to this article, Dr. Munk is a consultant for Eyegnos consulting and RetinAI. Dr. Waldstein consults for Alcon, Abbvie, Bayer, Boehringer Ingelheim and Roche.

 

In the treatment of neovascular age-related macular degeneration with anti-VEGF therapy, daily clinical decision-making largely revolves around tailoring injections to the specific needs of individual patients. This includes, for instance, adjusting treatment intervals—whether maintaining, shortening or extending them—based on the retinal microstructural dynamics observed on optical coherence tomography. 

However, modern treatment approaches are becoming increasingly complex and may require precise differentiation between various fluid subtypes or even measurement of pathological features. To this end, artificial intelligence holds potential to assist in assessing and quantifying biomarkers, and moreover, in determining individual patient prognoses.

In this article, we’ll discuss AI tools used for quantifying AMD-related biomarkers and predictive AI models that derive treatment needs or visual outcomes from those biomarkers.

 

Biomarkers that impact outcome and treatment needs

A growing body of literature describes how OCT-based prognostic biomarkers correlate with treatment demands and visual outcomes. For example, intraretinal fluid is typically associated with poorer visual acuity outcomes and is aggressively treated by most retina specialists. However, not all intraretinal fluid results from active exudation; degenerative intraretinal fluid—often seen over areas of fibrosis or RPE atrophy—shouldn’t necessarily preclude extending treatment intervals if the fluid pockets remain stable.

Subretinal fluid, while not strongly linked to worse visual outcomes, has been found to correlate with increased retreatment needs in standardized clinical protocols. SRF-tolerant treatment regimens often require quantification of SRF height or volume to determine if injection intervals can be safely extended.

Similarly, sub-RPE fluid has minimal impact on vision but is also associated with higher treatment frequency. Thus, accurate differentiation between fluid subtypes is essential, and fluid measurement is necessary to assess both stability and clinically meaningful cut-offs when advanced protocols are to be employed.

Figure 1. AI-based OCT layer and fluid segmentation in neovascular AMD. Representative OCT B-scans (top) with automated detection of intraretinal fluid (IRF), subretinal fluid (SRF) and pigment epithelial detachment (PED) using an AI-based segmentation algorithm (RetinAI). Quantitative analysis of fluid compartments (bottom) shows dynamic changes in IRF, SRF and PED volumes over time, illustrating treatment response and disease activity monitoring.

How AI can help in neovascular AMD

In the past decade, advances in computing infrastructure and graphics processing, combined with the availability of large ophthalmic image datasets, have driven a surge in AI technologies with measurable impact on retinal care. Today, several AI models with clinical application in wet AMD are commercially available, with even more developed in research settings. Following are ways AI can be put to use for managing wet AMD.

• AI for detection and quantification of biomarkers. Several AI systems are now commercially available to detect, classify and quantify fluid biomarkers (intraretinal, subretinal and sub-RPE fluid) in OCT scans of patients with neovascular AMD. These include the Notal OCT Analyzer (USA), RetinAI Discovery (Switzerland) and RetInSight Fluid Monitor (Austria). Most are based on standard deep learning (DL) architectures and have been validated against humans, showing comparable performance. Typically cloud-based, these systems require clinicians to upload OCT scans via a web interface.

In our view, there are three main clinical scenarios where such AI systems offer utility.

First and foremost, complex treatment regimens (e.g., SRF-tolerant protocols) require analyzing fluid stability over time, including precise measurement of its volume and/or height—metrics that, in practice, can only be reliably obtained with AI-assisted analytics. However, clinical evidence comparing AI-based approaches to current standards of care remains limited, which may explain the slow adoption of AI in routine practice. A large volume of intraretinal or subretinal fluid is typically considered a sign of disease activity, while complete absence of fluid usually supports a decision of inactivity. However, intermediate volumes—often in the range of just a few nanoliters—don’t always translate directly into activity assessments by the treating specialist. This illustrates why precise quantification through AI may help standardize interpretation and reduce subjectivity in clinical decision-making.

Second, virtual nurse- or technician-led clinics could benefit significantly. OCT scans acquired during these appointments could be pre-screened by AI to identify normal versus pathologic scans, as well as referral-requiring vs. non-referral-requiring scans reducing the workload for retina specialists.

Third, for future home-based OCT systems, AI-based analysis will be a critical component for efficient remote patient monitoring and management. In this setting, the use of pre-specified thresholds—as well as dynamically adjusted, patient-specific thresholds—will be essential to ensure both safety and individualized management.

• AI for prediction of outcomes and treatment requirements. A more ambitious application of AI lies in predicting treatment outcomes and required treatment frequency—an area currently under active research. Typically, explainable machine learning regression models are trained using biomarker quantifications from AI pipelines, in combination with visual acuity and other clinical parameters. This data, collected at baseline and during the initial loading phase (typically three injections), allows the model to estimate clinical outcomes and determine key contributing variables.

At present, prediction models achieve good accuracy mainly by stratifying patients into broad categories such as high- versus low-treatment demanders. To enable more granular forecasting—and potentially provide drug-specific recommendations—substantially larger and more diverse datasets will be required. Initial proof-of-concept studies show that AI can outperform human experts in predicting both visual outcomes and treatment needs. However, the current level of accuracy is still insufficient for clinical deployment. A major limiting factor is the lack of large and diverse training datasets beyond those sourced from clinical trials. Nevertheless, once these models reach a clinically applicable level of accuracy, they could serve as tools for improved patient counseling, enhance patient adherence and persistence, and allow efficient clinic management and planning. Initiatives such as the Global RETFound Consortium and AI-READI are expected to provide the large, high-quality datasets needed to build more reliable and clinically useful prediction models.

 
Using AI to explore novel biomarkers
 

Traditionally, our understanding of OCT biomarkers has been based on clinical and pathological observations made by human experts. However, AI systems are also capable of identifying novel patterns in image data in an unbiased fashion, without pre-defined human input. Advanced AI algorithms—particularly those using hybrid or unsupervised learning methods such as autoencoders—have already identified new, subclinical OCT biomarkers beyond the classical fluid or atrophy/fibrosis-related findings. While still in experimental stages, these discoveries may eventually lead to expanded diagnostic capabilities, pending validation in prospective clinical trials.

 

Bottom line

AI-based detection and quantification of biomarkers in wet AMD is no longer theoretical—it’s available. While several potential clinical applications exist, the lack of robust scientific evidence demonstrating improvement in clinical outcomes or workflow efficiency continues to hinder widespread adoption. Regulatory, data storage and processing constraints further limit widespread implementation. Advanced predictive models for disease progression and treatment needs are showing promise but remain in early developmental stages. Future research may not only refine these predictive tools but also expand our clinical vocabulary through unbiased AI-based exploration of imaging biomarkers. RS