Mohammad AskarPublic working copy · medical review pending

Retinal biomarkers and AI · 162

Retinal imaging and AI: biomarkers, evidence and intended use

Read fluid and tissue together with vision and disease context. A useful measurement is not automatically a validated treatment-selection tool.

Public working copy · medical review pending. Source, translation and medical checks remain open. Read the limitations in the text and references. These educational materials do not replace individual medical assessment. Published: 2026-11-18T19:45:52+01:00. Internal project references lead to this notice. They are not publicly accessible primary sources; the source limitations remain stated in the text.

Separate morphology, prognosis, comparative treatment benefit and regulatory intended use before acting on an AI output.

Terminology and draft status

Terminology review is not complete. Selected actual WHO UMD queries and their limits are recorded in terminology.md; they do not approve the whole translation.

01Define disease and the decision

Diabetic macular oedema (DME) and macular oedema after retinal vein occlusion (RVO) can threaten vision and need timely assessment. They share vascular leakage and inflammatory pathways, but diabetes-related microvascular disease and venous obstruction are not identical mechanisms; neovascular age-related macular degeneration (nAMD) is a distinct disease, not simply the same final pathway. 1

The export cites DME in 5.5% of people with diabetes, RVO in 28 million globally and 300000 in Germany. The underlying dates, case definitions and population denominators are untraced; retain these as original burden estimates, not verified current prevalence or comparable counts of one disease. The original ranking of DME as a primary working-age blindness cause and RVO as the second vascular visual-loss cause is likewise context-dependent and not verified as a universal current ranking. 1

Hypoxia can stimulate vascular endothelial growth factor (VEGF) and mediators such as interleukin-6 (IL-6), IL-8, monocyte chemoattractant protein-1 (MCP-1) and tumour necrosis factor-alpha (TNF-α). Endothelial adhesion pathways including ICAM-1, leukocyte recruitment, altered tight junctions and impaired fluid balance can contribute to leakage; this is a mechanistic framework, not proof that every focus is a macrophage or every eye needs a steroid. 1

IL-10 is also named in the export, but cytokines have context-dependent regulatory effects and should not all be called uniformly pro-inflammatory. Glial/microglial activation, macrophage infiltration, neurodegeneration and neovascularisation may coexist with leakage; neither increased flow nor ischaemia is a necessary universal sequence in every retinal disease. 1

02Ask what the biomarker actually predicts

A biomarker is an objectively measured characteristic used for a defined purpose. Diagnostic, monitoring, prognostic and predictive roles differ; the same OCT feature can support more than one role without being validated for every role. Explain the intended decision and limits to patients rather than promising personalized success. 1 2

The FDA–NIH BEST resource distinguishes predicting outcome from predicting a different effect of one treatment versus another. A correlation between DRIL and later acuity in treated eyes alone cannot show that switching to steroid improves outcome compared with continuing anti-VEGF. 2

Ask what the biomarker actually predicts
TypeDefinition and purposeClinical example and limit
DiagnosticDetect or confirm diseaseThickness supports DME/RVO assessment but does not determine cause alone.
MonitoringRepeated status measurementThickness/fluid track response; anatomy and vision can diverge.
PrognosticEstimate future courseDRIL association with acuity is not proof of a preferred drug.
PredictiveDifferential treatment effectHRF alone does not establish steroid benefit over anti-VEGF.

03Read OCT structure together with function

Optical coherence tomography (OCT) depicts thickness and fluid compartments. Intraretinal fluid (IRF) and larger cystoid spaces may accompany tissue damage and poor vision, but thickness and acuity do not change in lockstep. The original outer-nuclear-layer cyst threshold over 250 µm is not a validated universal duration test or proof of steroid superiority. 1 3

Hyperreflective foci (HRF) are small bright OCT signals; the source uses less than 30 µm. Activated microglia are one proposed substrate, not histological proof for every focus. Location, reflectivity, shadowing and definition affect counts; associations with worse vision and reductions during treatment do not establish a drug-specific predictive biomarker. 1

Disorganization of retinal inner layers (DRIL) means loss of distinguishable boundaries between inner layers on OCT. It may reflect neural structural damage after oedema, but does not prove irreversible loss of every pathway. A retrospective cohort of 96 participants and 120 eyes associated greater DRIL with worse acuity and early change with later acuity; this was not a randomized steroid-switch trial. 1 4

A separate retrospective study of 33 dexamethasone-treated eyes did not demonstrate prognostic value of DRIL or HRF for visual outcome during its first 4 months. Small, heterogeneous observational studies cannot establish that persistent DRIL under anti-VEGF mandates an early steroid switch. Consider the actual DME phenotype, prior response, lens status, pressure and treatment burden with guideline-supported options. 3 5

Manual counting can be time-consuming and variable. Automated segmentation may improve consistency, but repeatability, scan quality and agreement with a clinically meaningful reference must be tested; a precise fluid volume in nanolitres is not by itself evidence of improved vision. 1

The source defines HRF as small punctate signals at least as reflective as retinal pigment epithelium (RPE), but device, location and definition affect identification. DRIL concerns indistinct boundaries of the ganglion-cell/inner-plexiform complex, inner nuclear and outer plexiform layers; it is a structural OCT description rather than a direct measurement of every neural connection. 1

04Treat cytokine findings as exploratory evidence

Aqueous or vitreous cytokines may correlate with oedema, but the compartments are not interchangeable. In a prospective interventional series, 11 patients with bilateral DME contributed 22 eyes, one receiving triamcinolone and the other bevacizumab, with samples at baseline and 4 weeks; 6 cataract eyes provided controls. This small paired-eye study examined biological change rather than a validated selection rule. 1 6

Triamcinolone reduced a broader measured cytokine set while bevacizumab mainly reduced VEGF in that series. The short-term biochemical and anatomical findings do not prove better long-term vision or justify invasive sampling for routine drug choice; correlated fellow eyes, nonidentical controls and multiple analytes limit interpretation. 6

The original Japanese ranibizumab study, claiming better and faster response with higher baseline VEGF, is not uniquely identified from the export. Preserve the hypothesis without a numerical effect or clinical selection threshold. A future point-of-care cytokine test analogous to glucose monitoring is a research aspiration, not an available standard of care. 1

The export additionally cites two unnamed meta-analyses of raised IL-6, IL-8, MCP-1 and VEGF in DME/RVO, and a severity study in which IL-6/IL-8 increased but VEGF did not differ significantly across DR stages. Their identities, sample compartments, effect sizes and confounding remain unresolved; nonsignificance does not establish a constant VEGF level. 1

05Distinguish AI tasks and intended use

AI classification assigns disease or severity; segmentation measures regions such as IRF, subretinal fluid (SRF) or pigment epithelial detachment (PED); prediction estimates future activity, atrophy or treatment need. These are different tasks with different reference standards. The source’s EURETINA survey figure of 75% willingness is untraced and does not establish uptake, accuracy or patient benefit. 1

The six-system table retains every original product and separates historical claims from the source checked on 2026-11-08. RetinAI’s manufacturer distinguishes CE-marked models from research-use-only features; RetInSight Fluid and GA monitoring have specific inputs and purposes. The deepeye manufacturer announcement dated 2025-05-21 reports a CE mark, superseding the original forecast “expected mid-2025”; neither an announcement nor a platform name proves all current local indications. 7 8 9

The FDA’s actual DEN230043 indication is home imaging of hypo-reflective spaces in NV-AMD between scheduled assessments, not the original broad DME/RVO claim. It is not intended for diagnosis, autonomous treatment decisions or replacement of in-office assessment. A negative alert or poor scan must not delay review of new symptoms. 10

Autonomous diabetic-retinopathy screening is a separate use from OCT-guided treatment support. FDA K240058 covers AEYE-DS with Topcon NW400 or Optomed Aurora for adults with diabetes without previously diagnosed retinopathy; historical IDx-DR De Novo evidence does not validate every current LumineticsCore version or prove that its EU certification was not renewed. 11 12

The original EURETINA survey reports 75% future willingness and 72% expected diagnostic/monitoring support; the questionnaire, response denominator and selection are untraced. Publication growth since 2014 is a historical trend claim, not evidence that clinical benefit grows at the same rate. 1

Distinguish AI tasks and intended use
SystemDeveloper / originFunction and disease scopeStatus and boundary
RetinAI Discovery®RetinAI; Bern spin-off in exportFluids Report; nAMD, GA, DR/DME, CME in exportHistorical CE claim; separate CE models from research modules and verify IFU.
Vienna Fluid MonitorRetInSight; Vienna spin-off in exportIRF/SRF/PED segmentation in AMD; nanolitresHistorical MDR/CE report; version, scanner and intended use require checking.
deepeye TPS®deepeye MedicalnAMD fluid annotation, activity and 12-month injection predictionOriginal mid-2025 forecast superseded by manufacturer 21 May 2025 CE announcement; certificate/IFU unread.
Scanly® Home OCTNotal Vision, Inc.Patient-operated imaging; original nAMD/DME/RVO claim exceeds checked NV-AMD indicationFDA De Novo DEN230043; no autonomous treatment/diagnosis or replacement visits; current CE absence unverified.
RetinSight GA MonitorRetInSightPhotoreceptor/RPE loss area in GAHistorical MDR/CE directory entry, not proof of all current versions or visual benefit.
LumineticsCore™Digital DiagnosticsFundus DR screeningHistorical IDx-DR FDA evidence; current EU nonrenewal/cost attribution unverified.
Optomed Aurora AEYEAEYE Health / OptomedHandheld fundus DR screening, including access beyond eye clinicsFDA K240058 clearance for specified camera/software and population; no blanket diagnosis or verified current no-CE claim.

06Validate the whole clinical workflow

Potential gains include workload reduction, reproducible quantification, screening access and planning support. The export’s universal claim that clinician plus AI outperforms either alone lacks a specific study; effects depend on reader expertise, case mix, interface and error handling. Segmentation accuracy is not automatically visual benefit or cost effectiveness. 1

Check external validation, patient diversity, leakage between training and test data, camera/OCT compatibility, calibration and performance on ungradable images. The source flags predominantly white training populations as a possible bias, not a description established for every model. A system must identify failure and support an appropriate referral pathway. 1 10 11

Define who reviews output, protects patient data, acts on alerts and remains responsible for follow-up under the applicable jurisdiction. Preserve clinical skills to avoid deskilling, and explain uncertainty empathetically; AI supports a clinician’s contextual judgement rather than providing a guarantee of safe treatment. 1

The original text simultaneously describes substantial Asian training data and uses the phrase “white AI” for underrepresentation. These are not interchangeable population descriptions. Audit each dataset, geography and external validation, including African populations, rather than assigning one ethnic composition to all AI; Heidelberg-only compatibility is likewise a model-specific claim to verify. 1

What are the main learning points of this lecture?

The chapters connect findings, meaning and next steps. Check diagnosis, alternatives and source limits against the individual clinical findings.

Which findings require urgent assessment?

Rapid visual deterioration, severe pain, marked inflammation, pressure rise or suspected infection/neurological disease require urgent assessment. Apply the topic-specific warning signs and clinical context.

References and access status

  1. 1
    Immutable original export; own complete lecture export

    Complete export read; original attribution and any recording/transcript or appointment not independently verified.

  2. 2
    FDA–NIH BEST: prognostic versus predictive biomarkers

    Historical Review 039 access dated 8 November 2026 explicitly reused; Actual indexed selected explanation and treatment-interaction text read; whole BEST resource unread.

  3. 3
    NICE NG242 selected DME recommendations

    Historical Review 039 access dated 8 November 2026 explicitly reused; Selected actual DME recommendations read during Review 035 and explicitly reused; whole evidence reports and current product labels unread.

  4. 4
    Sun et al. 2014, PMID 25058813

    Historical Review 039 access dated 8 November 2026 explicitly reused; Complete actual NCBI XML abstract read; full cohort methods, supplementary analyses and patient-level data unread.

  5. 5
    OCT biomarkers after dexamethasone 2023, PMID 37316930

    Historical Review 039 access dated 8 November 2026 explicitly reused; Complete actual NCBI XML abstract read; full retrospective methods and supplements unread.

  6. 6
    Sohn et al. 2011, PMID 21782151

    Historical Review 039 access dated 8 November 2026 explicitly reused; Complete actual NCBI XML abstract read; allocation, paired-eye analysis and full multiple-testing methods unread.

  7. 7
    RetinAI manufacturer model library

    Historical Review 039 access dated 8 November 2026 explicitly reused; Actual indexed current manufacturer model distinction read; no complete certificates or module instructions inspected.

  8. 8
    Royal College of Ophthalmologists AI directory: RetInSight entries

    Historical Review 039 access dated 8 November 2026 explicitly reused; Actual indexed selected Fluid/GA entries read; directory reporting is not direct inspection of all certificates or local deployment.

  9. 9
    deepeye manufacturer CE announcement 2025-05-21

    Historical Review 039 access dated 8 November 2026 explicitly reused; Actual dated manufacturer announcement read as indexed text; certificate/IFU not inspected. An undated older portal still says assessment pending; dated announcement has clearer provenance.

  10. 10
    FDA DEN230043 decision summary, pages1–3 selected

    Historical Review 039 access dated 8 November 2026 explicitly reused; Actual indication, limitations and device-description text on pages1–3 read; remaining performance tables, complete IFU and later modifications not appraised.

  11. 11
    FDA AEYE-DS K240058 indication and selected study design

    Historical Review 039 access dated 8 November 2026 explicitly reused; Actual decision letter, indication page and selected study-design text read; all performance tables and IFU unread.

  12. 12
    FDA historical IDx-DR DEN180001

    Historical Review 039 access dated 8 November 2026 explicitly reused; Actual indexed selected pivotal-design and classification text read; not a current LumineticsCore IFU or an EU certificate renewal record.