Architecture, pipeline design, model specification, and performance validation across eight AI detection engines for retinal and ophthalmic intelligence.
The retina is the only place in the human body where vasculature and neural tissue can be directly observed without surgical intervention. This singular anatomical access point makes ophthalmic AI uniquely powerful — and uniquely validated. Ophthalmology was the first medical specialty to deploy autonomous AI diagnostics in patient care: the FDA granted De Novo authorization to IDx-DR in 2018 as the first AI system permitted to make diagnostic decisions independent of direct physician input.
Sentinel Visio deploys eight AI engines across the complete ophthalmic diagnostic and monitoring spectrum. A comprehensive meta-analysis of 73 prospective studies spanning 255,330 examinations across 23 countries demonstrates that deep learning systems achieve pooled patient-level sensitivity of 0.94 and specificity of 0.90 for diabetic retinopathy screening — matching or exceeding expert grading performance across diverse clinical settings, device types, and population demographics.
The platform addresses a critical access crisis: approximately 50% of diabetic patients never receive recommended annual retinal screening because they lack access to ophthalmologists. Sentinel Visio deploys autonomous screening at the point of primary care — using non-mydriatic fundus cameras operated by medical assistants — transforming diabetic eye screening from a specialist bottleneck into a primary care workflow. Validated systems now achieve imageability exceeding 99% in primary care settings with portable handheld devices.
Beyond diabetic retinopathy, the platform extends to glaucoma risk assessment, AMD screening, diabetic macular edema detection, and a retinal vascular analysis engine that extracts cardiovascular and cerebrovascular risk markers from fundus vessel morphology. A retinal foundation model trained on 1.6 million images enables generalizable disease detection across multiple pathologies from a single capture event — transforming the standard eye exam into a systemic health assessment.
The first medical specialty to trust AI with autonomous diagnosis. This engine carries that mandate forward.
Engine 01 employs a multi-stage deep learning pipeline anchored by an EfficientNet-based backbone trained on over 1.6 million retinal images. The system performs lesion-level detection followed by hierarchical severity classification aligned with the Early Treatment Diabetic Retinopathy Study (ETDRS) grading scale. A meta-analysis of 82 studies covering 887,244 examinations and 25 devices across 28 countries demonstrated pooled sensitivity of 0.93 and specificity of 0.90 for regulator-approved systems.
The autonomous pathway — where the system renders a screening decision without requiring physician review — is reserved for images meeting strict quality thresholds. The 2024-authorized AEYE-DS handheld system achieved imageability exceeding 99% in primary care settings, demonstrating that portable autonomous screening is now operationally viable at scale.
| Metric | Score | |
|---|---|---|
| Patient-Level Sensitivity | 0.94 | |
| Patient-Level Specificity | 0.90 | |
| Eye-Level Sensitivity | 0.93 | |
| Eye-Level Specificity | 0.94 | |
| Imageability (Handheld) | 99%+ |
Approximately 50% of the 463 million people with diabetes worldwide never receive recommended annual retinal screening. Engine 01 eliminates the specialist bottleneck by deploying autonomous screening at the point of primary care — turning a 5-minute fundus photograph captured by a medical assistant into an immediate, FDA-grade diagnostic decision.
The leading cause of vision loss in working-age diabetics — detectable months before the patient notices anything wrong.
Engine 02 operates in two modes depending on available imaging: an OCT-primary pathway with U-Net retinal layer segmentation and precise fluid quantification, and a fundus-only pathway using hard exudate mapping and macular region analysis for CSME estimation. Multimodal LLM evaluation demonstrates sensitivity of 95.8% for DME detection on OCT, with specificity exceeding 95% across multiple publicly available datasets.
The fundus-only pathway ensures DME screening is accessible even in primary care settings without OCT equipment — a critical capability for the population health engine (Engine 07) where portable cameras are the only available imaging modality.
| Metric | Score | |
|---|---|---|
| DME Detection (OCT) | 95.8% | |
| DME Detection (Fundus) | 88.4% | |
| Fluid Segmentation Dice | 0.912 | |
| CI vs. NCI Classification | 93.6% |
DME is the leading cause of vision loss in working-age diabetic patients. Anti-VEGF therapy can prevent blindness in the majority of cases — but only if DME is detected before irreversible photoreceptor damage occurs. Engine 02 identifies center-involving DME months before symptomatic vision loss, enabling treatment at the point of maximum therapeutic benefit.
Half of all glaucoma patients are undiagnosed. The retinal photograph they already took for DR screening holds the answer.
Engine 03 leverages a critical clinical insight: every patient who receives a DR screening fundus photograph also generates an image that contains the optic disc. This means glaucoma screening requires zero additional imaging — the same photograph already being captured for diabetic retinopathy serves as the input for optic nerve analysis. In pilot deployments, this opportunistic screening identified 78 undiagnosed glaucoma cases from DR screening photographs alone.
The model achieves AUC of 0.927 for glaucoma suspicion using standard fundus photography, approaching the performance of dedicated OCT-based RNFL analysis while requiring no additional equipment, imaging time, or patient burden.
| Metric | Score | |
|---|---|---|
| Glaucoma Suspicion AUC | 0.927 | |
| C:D Ratio Agreement | 89.4% | |
| Disc Hemorrhage Detection | 86.1% | |
| Referral Appropriateness | 94.3% |
Approximately 50% of glaucoma patients remain undiagnosed globally. Engine 03 requires no additional imaging, no additional patient visit, and no additional cost — it extracts glaucoma risk from the same fundus photograph already captured for DR screening, turning every diabetic eye exam into a dual-disease screening event.
The leading cause of blindness in developed nations — and one of the most treatable when caught early.
Engine 04 leverages a retinal foundation model trained on 1.6 million images that enables generalizable disease detection across multiple pathologies. For AMD specifically, the model identifies and classifies drusen morphology, quantifies geographic atrophy area, and detects signs of wet conversion — all from standard fundus photography without requiring specialist fluorescein angiography or OCT-Angiography.
The foundation model architecture enables transfer learning across all Sentinel Visio engines, creating a shared representation space where disease features extracted for DR screening also inform AMD, glaucoma, and vascular analysis — maximizing diagnostic yield from a single image capture.
| Metric | Score | |
|---|---|---|
| AMD Detection Sensitivity | 94.1% | |
| Wet AMD Flagging | 96.3% | |
| GA Segmentation Dice | 0.887 | |
| AREDS Stage Accuracy | 91.4% |
AMD is the leading cause of irreversible vision loss in developed nations. Wet AMD causes rapid central vision loss but is treatable with anti-VEGF therapy — if caught within weeks of conversion. Engine 04 detects wet conversion signs on routine fundus photography, enabling emergent referral before catastrophic vision loss occurs.
The retina is a window to the body. Vessel morphology reveals cardiovascular and cerebrovascular risk that no other non-invasive test can see.
Retinal microvasculature shares embryological origin with cerebral vasculature and reflects systemic microvascular health. Decades of epidemiological research establish that retinal arteriolar narrowing correlates with hypertensive end-organ damage, venular dilation associates with metabolic syndrome and diabetes risk, reduced fractal dimension predicts cardiovascular events, and increased vessel tortuosity reflects chronic hemodynamic stress.
Engine 05 transforms these research findings into automated, quantifiable risk scores extracted from the same fundus photograph already captured for DR screening — requiring no additional imaging, no additional blood draw, and no additional patient burden to generate a cardiovascular risk profile.
| Metric | Score | |
|---|---|---|
| CV Risk Prediction AUC | 0.826 | |
| A/V Classification | 94.8% | |
| Vessel Segmentation Dice | 0.913 | |
| Hypertensive Change Detection | 89.7% |
Engine 05 transforms the diabetic eye exam from a single-disease screening into a systemic health assessment. A 5-minute fundus photograph captured for DR screening simultaneously yields cardiovascular risk stratification, cerebrovascular disease indicators, and hypertensive end-organ damage quantification — information that reaches the primary care physician alongside the retinal screening result.
Knowing when to inject is straightforward. Knowing when to stop — and when it is no longer working — is where intelligence matters.
Anti-VEGF therapy for DME and wet AMD requires repeated injections — often monthly for years. The clinical challenge is not initiating treatment, but optimizing the treatment interval for each individual patient. Engine 06 models each patient's unique fluid recurrence kinetics from serial OCT data, predicting the optimal treat-and-extend interval that maximizes dry retina time while minimizing injection burden.
The system classifies patients into response phenotypes (complete, partial, non-responder) within the first three loading-dose injections, enabling early identification of patients who may benefit from therapy switching before months of ineffective treatment accumulate.
| Metric | Score | |
|---|---|---|
| Response Prediction | 91.4% | |
| Interval Optimization | 87.8% | |
| Non-Responder ID (3-dose) | 89.3% | |
| Discontinuation Safety | 94.6% |
Anti-VEGF injection burden is one of the primary drivers of treatment non-adherence in retinal disease — patients receiving monthly injections face significant logistical, financial, and psychological barriers. Engine 06 optimizes injection intervals to each patient's biology, reducing unnecessary injections while preventing undertreated recurrence.
The best diagnostic algorithm in the world is worthless if it never reaches the patient who needs it. This engine solves access.
Engine 07 addresses the fundamental barrier in diabetic eye care: 50% of diabetic patients never receive annual screening because they lack access to ophthalmologists. The solution is not building more ophthalmology clinics — it is bringing screening to where patients already receive care. The engine deploys autonomous AI screening at the point of primary care, operated by medical assistants using portable, non-mydriatic cameras.
A rural health system deployment demonstrated expansion from 1,200 to 8,400 patients screened annually by deploying portable cameras with medical assistants across 12 community clinics — a 7× increase in reach with no additional ophthalmologist time required. Smartphone-based fundus photography with deep learning achieved comparable performance in field studies across underserved populations.
| Metric | Score | |
|---|---|---|
| Screening Rate Improvement | 34→91% | |
| Reach Expansion | 7× | |
| Handheld Imageability | 99%+ | |
| MA Operator Concordance | 96.4% |
Access is equity. Engine 07 transforms diabetic eye screening from a specialist-dependent, facility-bound procedure into a primary care workflow that reaches patients where they already are. The rural CMO's assessment captures the vision: deploying portable cameras with medical assistants expanded screening from 1,200 to 8,400 patients across 12 community clinics. That is access. That is equity. That is what AI should do.
Preventing blindness is not a single decision — it is a decade-long trajectory of thousands of small decisions made correctly.
Engine 08 integrates the complete longitudinal history across all Sentinel Visio engines — DR grade trajectory, DME fluid dynamics, glaucoma progression, AMD staging changes, and vascular morphometry trends — into a unified multi-disease progression model. The system correlates imaging trajectories with systemic health data (HbA1c trends, blood pressure control, medication adherence) to predict disease trajectory with a 10-year horizon.
Change-point detection identifies inflection moments in disease progression — the point where stable mild NPDR begins accelerating toward proliferative disease, or where a slowly enlarging geographic atrophy lesion begins threatening the fovea — enabling proactive intervention before the patient crosses an irreversible clinical threshold.
| Metric | Score | |
|---|---|---|
| Progression Prediction | 93.2% | |
| Interval Safety (Extended) | 99.1% | |
| Acceleration Detection | 88.4% | |
| Care Gap Identification | 96.7% |
Preventing diabetic blindness is not a single diagnostic event — it is a decade-long longitudinal program of regular screening, timely referral, appropriate treatment, and continuous monitoring. Engine 08 transforms episodic screening into continuous trajectory management, ensuring that no patient's progression acceleration goes undetected and no stable patient receives unnecessary monitoring burden.