Introduction and Context
Age-related macular degeneration (AMD) is the leading cause of irreversible central vision loss in older adults in high-income countries and a rapidly growing public‑health challenge globally. Early identification of AMD—especially the sight-threatening neovascular (wet) form and geographic atrophy (GA)—enables timely referral and management that can preserve vision. Artificial intelligence (AI) screening models applied to retinal images have shown promise for detecting undiagnosed eye disease at scale in non-specialist settings (primary care, community screening, teleophthalmology). However, reliable algorithm development and comparison require a consistent, reproducible reference standard for labeling images used in training and validation.
In April 2026, a multi‑disciplinary panel of retina specialists, general ophthalmologists, imaging experts, and AI specialists published a modified Delphi consensus that defines a reference standard for labeling AMD in images intended for validation of screening algorithms (Domalpally et al., Ophthalmology 2026). The guidance aims to harmonize how images are annotated and interpreted for AI validation, not to replace clinical practice guidelines for patient management.
Why this consensus matters now
– Rapid proliferation of AI tools for retinal disease requires consistent, transparent validation pipelines to ensure safety, generalizability, and regulatory readiness.
– Prior scales and classification systems (AREDS severity scales, Beckman) were developed for clinical trials or epidemiology and varied in modality reliance (color fundus photography vs multimodal imaging).
– Advances in clinical retinal imaging—most importantly spectral-domain OCT—improve detection of key AMD features and alter how ground truth should be established for image-based AI. The consensus addresses that gap.
New Guideline Highlights
Major takeaways from the 2026 Delphi consensus (Domalpally et al.)
– Level 1 reference standard: The Beckman clinical classification of AMD was recommended as the primary (Level 1) reference scale for image‑based labeling used to validate screening algorithms (median score 8; agreement).
– OCT is essential: Use of OCT to identify core AMD features (drusen, geographic atrophy, choroidal neovascularization) reached strong consensus (median scores 8.5–9). OCT should be incorporated when available to confirm structural abnormalities.
– Pigmentary changes: Detection of pigment change on color imaging as a labeling feature did not reach full consensus (median 7.5; uncertain), reflecting variable reliability on different modalities and grader variability.
– Screening age threshold: Panelists did not reach consensus on a single, universal age cutoff for screening (median 8; uncertain), reflecting variation in population risk and program goals.
– Referral thresholds: There was strong agreement on practical referral thresholds—urgent referral for findings consistent with neovascular AMD and nonurgent referral for GA and intermediate AMD (median 9; agreement).
Key implications
– For AI developers and evaluators, Beckman-classified datasets with multimodal corroboration (OCT ± fundus autofluorescence and color fundus photography) should be the preferred validation benchmark.
– Screening studies should document imaging modalities used, grader training and adjudication process, and the reference level (e.g., Beckman Level 1) employed.
Updated Recommendations and Key Changes from Prior Approaches
What changed compared with earlier classification schemes and common practice?
– Ascendancy of Beckman as Level 1 reference: While many prior studies used the AREDS-based severity scales for epidemiology and trials, the panel endorsed the Beckman clinical classification as the pragmatic primary standard for image-based screening validation. The Beckman system provides clinically meaningful categories (no AMD, early, intermediate, late nonexudative (GA), late exudative (neovascular)). (Ferris et al., Ophthalmology 2013)
– Formal inclusion of OCT: Earlier reference standards often relied mainly on color fundus photography. The 2026 consensus formally elevates OCT to a key component of the reference standard for screening validation because of OCT’s superior sensitivity for subretinal/intraretinal fluid, drusen morphology, and atrophic changes.
– Explicit separation of validation labeling from clinical practice guidelines: The consensus clarifies that the reference standard is for labeling images for AI evaluation and not a directive for how care should be delivered—an important distinction for regulators and clinical adopters.
Table: High-level comparison (prior typical practice vs 2026 consensus)
– Modality for reference labeling: Color fundus photography only → Beckman + OCT as preferred multimodal reference
– Primary classification system used: AREDS/varied → Beckman clinical classification (Level 1)
– Role of pigmentary change: Often included → Uncertain reliability; no consensus to require as Level 1 feature
– Referral guidance embedded in labels: Variable → Consensus: urgent for neovascular AMD; nonurgent for GA/intermediate AMD
Topic‑by‑Topic Recommendations
Below are the main domain recommendations distilled into practical points for developers, clinicians, and researchers.
1) Reference grading framework
– Level 1 reference: Use the Beckman clinical classification when labeling images for screening algorithm validation. Beckman categories: no AMD, early AMD, intermediate AMD, late AMD—subdivided into geographic atrophy (GA) and neovascular AMD.
– If additional granularity is needed (e.g., for prognostic models), document how features map to Beckman categories and ensure interrater adjudication.
2) Imaging modalities and evidence hierarchy
– Core modalities for reference labeling: OCT plus color fundus photography (CFP) when available. Autofluorescence (FAF) can support detection of atrophy and should be used when available.
– OCT is required to confirm suspected neovascular AMD (intraretinal/subretinal fluid, pigment epithelial detachments with SRF), to better characterize drusen and reticular pseudodrusen, and to identify atrophy (using CAM definitions on OCT where applicable).
– In resource‑limited settings where OCT is not available, CFP-based Beckman labeling is acceptable but should be annotated as lower-level reference and developers must acknowledge the increased uncertainty.
3) Key imaging features and labeling practice
– Drusen: Document size (small, medium, large) and confluence; correlate CFP appearance with OCT to identify subretinal drusenoid deposits (pseudodrusen).
– Pseudodrusen (reticular pseudodrusen): Where possible, confirm on OCT or near-infrared imaging; these lesions carry prognostic significance and should be noted separately if reliably identified.
– Pigmentary changes: Because of grader variability and limitations in imaging (CFP vs OCT/Af), pigmentary change was not mandated as a core Level 1 label. If used, clearly define criteria and report intergrader agreement.
– Geographic atrophy (GA): Use structural OCT criteria and/or FAF patterns; where available, apply CAM group consensus definitions for OCT-based atrophy. (Sadda et al., Ophthalmology 2018)
– Neovascular AMD (nAMD): OCT-confirmed subretinal/intraretinal fluid or OCT signs of active neovascularization should be used to label nAMD. CFP alone is insufficient for high-confidence labeling of nAMD.
4) Referral thresholds and recommended actions
– Urgent referral: Findings consistent with neovascular AMD (fluid on OCT, acute vision change with new hemorrhage on CFP) → urgent retina referral for evaluation and likely treatment.
– Nonurgent referral: Intermediate AMD (large drusen or substantial intermediate findings) and GA without signs of active exudation → routine retinal referral/monitoring and counseling.
– Negative for AMD or early AMD: Advise routine monitoring, risk‑factor management (smoking cessation, healthy diet), and consideration of AREDS/AREDS2 supplementation per clinical guidelines when appropriate.
5) Age and target population for screening
– No single age cutoff achieved consensus. Programs should define target screening age based on local epidemiology, resource capacity, and program goals. Many programs prioritize adults aged ≥50–55 due to increased AMD prevalence, but flexibility is endorsed.
6) Grader qualifications and adjudication
– Images used for reference labeling should be graded by fellowship-trained retina specialists where possible or by trained ophthalmic graders with adjudication by retina specialists for uncertain cases.
– Document grader training, consensus processes, intergrader agreement metrics, and procedures for arbitration.
7) Dataset reporting and benchmarking
– Validation studies should transparently report: imaging modalities used; classification system (Beckman Level 1 or alternative); grader credentials; intergrader reliability; prevalence of AMD categories; and referral thresholds tied to labels.
– Publicly accessible, well‑annotated reference datasets using these consensus standards are encouraged to improve comparability across algorithms.
Expert Commentary and Insights
Panel composition and perspectives
– The Delphi panel included a broad cross-section of expertise—retina subspecialists, ophthalmologists, AI researchers, and imaging experts. This multidisciplinary makeup strengthened the recommendation set but also reflected real world tensions between ideal reference standards (multimodal imaging) and pragmatic constraints (limited OCT access).
Consensus, controversy, and gray areas
– Strong consensus: Beckman as the principal reference framework and the centrality of OCT for confirming late AMD subtypes.
– Areas of ongoing debate: The role of pigmentary changes as a reliable label; a universal screening age; and how to handle datasets from low-resource settings lacking OCT. The panel recognized equity concerns and recommended transparency when lower-level references must be used.
Regulatory and clinical practice implications
– The consensus is intentionally separate from clinical practice guidelines but is highly relevant to regulators and payers evaluating AI screening solutions. A consistent reference standard can reduce ambiguity in performance claims and enable apples‑to‑apples comparisons across algorithms.
– Developers should not conflate a positive screen from an AI tool with definitive diagnosis—labels are for screening and triage. Clinical workflows must ensure confirmatory evaluation by eye care professionals.
Future needs highlighted by experts
– External, prospective validation of AI tools using the consensus reference; inclusion of diverse populations to ensure generalizability; public reference datasets that use multimodal labeling; and operational research on integration into primary care pathways and teleophthalmology.
Practical Implications
For AI developers
– Use Beckman-defined labels for primary validation sets whenever possible and include OCT-confirmed ground truth for late AMD subtypes. Clearly report modality mix, grader credentials, and interrater reliability.
For clinicians and health systems
– Expect future AI screening outputs to be benchmarked against the Beckman + OCT reference. When implementing AI screening, ensure pathways for urgent referral (suspected nAMD) and for counseling/monitoring of intermediate AMD and GA.
For regulators and payers
– This consensus provides a practical, expert‑backed framework for assessing the validity of AMD screening algorithms. Health systems should require transparent reporting of the reference standard used and performance stratified by imaging modality and patient subgroups.
Patient vignette (illustrative)
Mr. James Parker, 67, attends a community health fair where a retinal image screening program uses an AI tool validated against the Beckman + OCT consensus standard. The AI flags his right eye as “high probability of neovascular AMD.” On expedited referral, OCT confirms subretinal fluid and the retina specialist initiates anti‑VEGF therapy the same week. Result: stabilization of vision. This scenario underscores how a validated AI screen tied to clear referral thresholds (urgent for nAMD) can shorten time to treatment.
Conclusions and Next Steps
The 2026 Delphi consensus offers an important, practical step toward harmonizing how images are labeled for validation of AI-based AMD screening algorithms. By endorsing the Beckman classification as the principal reference and elevating OCT as a core modality for confirming late AMD subtypes, the guidance increases clarity and comparability across studies and products. However, the consensus also acknowledges important uncertainties—particularly around pigmentary changes and universal age-based screening thresholds—and emphasizes transparency and graded evidence reporting.
Moving forward, priorities include building and sharing multimodal, diverse reference datasets annotated to this standard; prospectively validating AI screening tools in real-world settings; addressing equity and access issues where OCT is not available; and closing evidence gaps about how AI-enabled screening affects long-term vision outcomes.
References
– Domalpally A, Chew EY, Eydelman MB, Keenan TDL, Keane PA, Lee AY, Lee CS, Lad EM, Lim JI, Lowenstein A, Schmidt-Erfurth U, Abramoff MD, Collaborative Community for Ophthalmic Imaging Executive Committee and the Working Group for Artificial Intelligence in Age-related Macular Degeneration. Reference Standard for Validation of Age-Related Macular Degeneration Screening Algorithms. Ophthalmology. 2026 Apr 17;133(7):865-873. PMID: 41999903. https://pubmed.ncbi.nlm.nih.gov/41999903/
– Ferris FL 3rd, Wilkinson CP, Bird A, Chakravarthy U, Chew E, Csaky K, Sadda SR; Beckman Initiative for Macular Research Classification Committee. Clinical classification of age-related macular degeneration. Ophthalmology. 2013 Apr;120(4):844-851. doi:10.1016/j.ophtha.2012.10.036.
– Sadda SR, Guymer R, Holz FG, et al.; CAM Consensus Working Group. Consensus definition for atrophy associated with age-related macular degeneration on OCT: Classification of Atrophy Meeting (CAM) report. Ophthalmology. 2018 Nov;125(4):537-548. doi:10.1016/j.ophtha.2017.10.018.
(Additional background on AREDS, OCT biomarkers, and AI validation literature may be consulted in standard ophthalmic and regulatory sources.)
