AI Diabetic Retinopathy Screening: What the Algorithm Actually Does

In 2018, an algorithm became the first AI diagnostic system in any field of medicine to receive FDA authorisation to make a diagnosis without a physician reviewing the image.

The disease it diagnosed was diabetic retinopathy, and the image was an ordinary fundus photograph.

AI-based diabetic retinopathy screening is now deployed at scale in primary care and community settings that have never had access to an ophthalmologist.

The clinical case for it is almost entirely about a mismatch: the number of people with diabetes is enormous, and the number of eye specialists able to screen them all, every year, is not.

Understanding what these systems can and cannot do is now a basic requirement for anyone practising in diabetic eye care.

What Is AI Diabetic Retinopathy Screening?

AI diabetic retinopathy screening uses deep learning algorithms, typically convolutional neural networks, to analyse fundus photographs and automatically detect referable diabetic retinopathy without a human grader in the loop.

Systems in current use fall into two broad categories:

Autonomous systems, authorised to issue a screening result directly to the patient or primary care provider without human over-read
Assistive systems, which flag or pre-grade images for a human reader, aiming to reduce grading workload rather than replace it entirely

The regulatory distinction between these two categories matters enormously – autonomous systems carry a different liability and validation burden than assistive tools sitting alongside a human grader.

The Screening Problem This Solves

The number of people with diabetes worldwide has risen steeply, and annual retinal screening is recommended for essentially all of them
Trained human graders and ophthalmologists cannot scale to meet that volume in many health systems, particularly in low-resource and rural settings
Manual grading is also subject to inter-grader variability, meaning even fully staffed programmes are not perfectly consistent
Non-attendance at screening remains one of the largest single drivers of preventable diabetic vision loss, independent of grading accuracy

The core value proposition is capacity, not just accuracy – these tools exist primarily to screen people who would otherwise not be screened at all.

How the Technology Works

Deep convolutional neural networks are trained on large datasets of fundus images labelled by expert human graders against a reference standard, typically the International Clinical Diabetic Retinopathy severity scale
The network learns to detect the lesions that define disease severity – microaneurysms, dot and blot haemorrhages, hard exudates, cotton wool spots and neovascularisation
Output is typically a binary or graded classification – referable versus non-referable disease, sometimes with severity staging
Image quality assessment is built into most systems, since poor-quality images that cannot be reliably graded need to be flagged rather than misclassified

What the Model Actually Sees

A gradable colour fundus photograph, of the kind captured by tabletop or handheld retinal cameras, is the raw input for these algorithms – the same image type a human grader would review, but processed in seconds rather than minutes.

Clinical Evidence

IDx-DR became the first autonomous AI diagnostic system authorised by the FDA for any field of medicine, based on a prospective trial demonstrating sensitivity and specificity meeting pre-specified thresholds for detecting more-than-mild diabetic retinopathy
Multiple subsequent systems, including tools developed and validated across large screening populations in the UK, India and elsewhere, have shown comparable performance to human graders on external validation datasets
Real-world deployment studies have generally shown good sensitivity for referable disease but performance can vary by population, camera type and image quality compared with the controlled conditions of pivotal trials
Cost-effectiveness analyses generally favour AI-assisted screening for populations with capacity constraints, though the calculation depends heavily on local staffing costs and screening uptake

The gap between pivotal-trial performance and real-world performance is the recurring theme in this literature – external validation on a program’s own population and camera is not optional before deployment.

Integration Into Screening Pathways

Typical Autonomous Workflow

Fundus photographs are captured, often by a technician or trained non-specialist in a primary care or community setting
The algorithm grades image quality and disease status in real time or near real time
Patients graded as non-referable are recalled per the standard screening interval; referable results are sent for prompt ophthalmic assessment
Ungradable images are routed to human review or a repeat photograph

Assistive Workflow

The algorithm pre-grades or triages images before a human grader review, aiming to prioritise likely-referable cases and reduce grader workload on clearly normal images
A human grader remains the final decision-maker in every case

Ungradable-image handling is one of the most practically important design decisions in any of these systems – a program that quietly discards or mishandles ungradable images loses exactly the patients most in need of a careful look.

Limitations and Safety Considerations

Algorithms are trained and validated for diabetic retinopathy specifically, and are not general-purpose fundus interpretation tools – they can miss unrelated pathology such as glaucomatous disc changes or macular degeneration unless separately trained and validated for those conditions
Performance can degrade on populations, ethnicities or camera types under-represented in training data, which is why local validation before deployment is now widely recommended
Image quality remains a major failure point – poor pupil dilation, media opacity and patient movement all reduce gradability
A clear escalation pathway and clinical governance structure is required for referable results, ungradable images and any diagnostic disagreement
Regulatory approval in one jurisdiction or for one algorithm version does not automatically extend to another population, camera, or software update

The single-disease scope of these tools is the point most likely to be misunderstood by non-specialist users – a normal diabetic retinopathy screening result says nothing about the optic nerve or macula in isolation.

Prognosis for the Field

AI screening is moving from novelty to standard infrastructure in diabetic eye care.

Multiple national screening programmes have already incorporated AI grading, either autonomously or as an assistive first pass
Expansion toward multi-disease screening – simultaneous detection of glaucomatous and age-related macular changes from the same image – is an active area of development
Continued external validation, post-deployment monitoring and clear accountability structures remain the main barriers to broader trust and adoption, more than the underlying algorithmic accuracy
The realistic long-term role is not replacing ophthalmologists but extending screening reach to populations who currently receive none at all

The technology has already answered the accuracy question convincingly – what remains is largely a health-systems question of deployment, governance and equitable access.

References

Abramoff MD, Lavin PT, Birch M, et al. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy. npj Digital Medicine. 2018.
Gulshan V, Peng L, Coram M, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA. 2016.
Ting DSW, Cheung CY, Lim G, et al. Development and validation of a deep learning system for diabetic retinopathy and related eye diseases. JAMA. 2017.
Artificial Intelligence in Diabetic Retinopathy Screening. EyeWiki, American Academy of Ophthalmology.
Diabetic Retinopathy Screening. StatPearls, NCBI Bookshelf.