Glaucoma screening is the attempt to find optic nerve damage before the patient notices any loss of vision. That is harder than it sounds, because glaucoma takes the peripheral field first and the central field last, so most people have no symptoms until the disease is advanced. A 2014 meta-analysis in Ophthalmology estimated that 76 million people aged 40 to 80 would have glaucoma in 2020, rising to 111.8 million by 2040, with a global prevalence of 3.54 percent in that age group. In India, clinicians speaking at World Glaucoma Week events put the number of people living with the disease at close to 12 million and estimate that around 90 percent of cases are undiagnosed.
The damage cannot be reversed, but it can be slowed with treatment once it is found. That makes early detection of glaucoma the point where the most sight is saved. This article explains what an algorithm looks for on a photograph of the optic nerve, how AI glaucoma detection turns those features into a referral decision, what the published evidence shows, and where the approach has limits.
The optic nerve head, or optic disc, is where roughly one million retinal ganglion cell axons leave the eye. On a colour fundus photograph, it appears as a pale, roughly circular area with a central depression called the cup, surrounded by a ring of nerve tissue called the neuroretinal rim. Glaucoma kills ganglion cells, the rim thins, and the cup enlarges. Optic nerve head analysis is the structured assessment of this change.
Clinicians look for a consistent set of signs:
No single sign is diagnostic. Large physiological cups occur in healthy eyes, especially large discs, and high myopia distorts the disc. Human assessment is therefore subjective, and agreement between graders is imperfect. That variability is one of the reasons automated analysis has attracted interest.
Early automated systems segmented the disc and cup explicitly, measured the CDR and applied a threshold. Current systems use convolutional neural networks trained end to end on tens of thousands of fundus photographs that specialists have labelled as glaucomatous or not. The network learns the pixel patterns that predict the label, which in practice include cup size, rim shape and RNFL appearance, without being told to measure any of them.
In deployment, the sequence has four steps.
A 2019 study from Google, published in Ophthalmology, tested which disc features drove referral decisions. For glaucoma specialists and for the algorithm alike, the features that mattered most were a vertical CDR of 0.7 or more, neuroretinal rim notching, RNFL defects and bared circumlinear vessels. The algorithm, in other words, had learned to weigh the same signs a specialist weighs.
Li et al., Ophthalmology, 2018. A deep learning system was trained and validated on 48,116 fundus photographs. Referable glaucomatous optic neuropathy was defined as a vertical CDR of 0.7 or more, or other typical changes such as rim notching, an RNFL defect, or a disc haemorrhage. On the validation set, the system reached an area under the curve (AUC) of 0.986, with 95.6 percent sensitivity and 92.0 percent specificity. Most false negatives occurred in eyes with high myopia, and most false positives in eyes with physiological cupping or pathological myopia.
Ting et al., JAMA, 2017. A system trained on 125,189 images from Singapore was validated on 71,896 images from 14,880 patients in the Singapore national diabetic retinopathy screening programme, a community-based Chinese, Malay and Indian population. For possible glaucoma it achieved an AUC of 0.942, with 96.4 percent sensitivity and 87.2 percent specificity.
Phene et al., Ophthalmology, 2019. A model trained on 86,618 images reached an AUC of 0.945 on a 1,205-image validation set graded by glaucoma specialists, and 0.855 and 0.881 on two further external sets. Its sensitivity exceeded that of seven of ten graders, including two of three glaucoma specialists, with comparable specificity.
Christopher et al., Scientific Reports, 2018. On 14,822 optic nerve head photographs from a racially diverse population, a transfer-learning model reached an AUC of 0.91 overall and 0.97 for moderate to severe disease. The gap between those two numbers is the honest summary of the field. Advanced glaucoma is easy to detect on a photograph. Early glaucoma is harder, for algorithms and for humans.
AI analysis of a fundus photograph is a triage tool. It identifies discs that warrant a full examination. It does not diagnose glaucoma, because the diagnosis also depends on intraocular pressure, the drainage angle, the visual field and, increasingly, OCT measurement of RNFL thickness. The limits are worth stating.
Population screening for glaucoma has historically been judged not cost-effective because the tests were slow, specialist-dependent and prone to false positives. Opportunistic screening of people who are already being photographed for another reason, such as diabetic retinopathy screening, changes that calculation, because the image already exists and the marginal cost of a second algorithm is small.
Early detection of glaucoma depends on noticing structural change at the optic nerve head before the patient notices anything at all. Cup-to-disc ratio, the ISNT rule, rim notching and RNFL defects are the signs clinicians have always used, and deep learning systems have now been shown to weigh the same signs and reach referable glaucomatous optic neuropathy with sensitivity and specificity in the range of 87 to 96 percent across large validation sets. The practical opportunity is to add glaucoma analysis to retinal photographs that are already being captured for diabetic retinopathy screening, so that the undiagnosed majority are found where they already receive care.
Forus Health AI works on this problem through FH-POISE (Precision Ocular Intelligence for Systemic and Eye Health), its retinal image analysis platform for the detection and screening of ocular and systemic diseases. The 3nethra range of portable, non-mydriatic imaging devices that Forus Health manufactures is designed to screen for diabetic retinopathy, glaucoma and other conditions in primary care and outreach settings, and FH-POISE is built to run on the images those devices capture. The aim is the one set out in this article, which is to bring structured optic nerve head analysis to the point of care, so that the patients who need a specialist are identified early and specialist time is spent where it makes a difference.
Sources
Diabetic retinopathy screening is the routine examination of the retina in people with diabetes to detect vascular damage before it threatens sight. The need in India…
