Systemised AI for Surfer’s Disease

Artificial intelligence is frequently used while surfing the web, but it’s now being applied to a condition associated with surfing—pterygium—according to Carolin Elhardt MD.

While the condition usually occurs in the corner of the eye, in severe cases it can grow to cover the cornea. It is commonly known as ‘surfer’s eye’ due to prolonged exposure to UV radiation as its primary cause.

As pterygium can present a variety of symptoms and clinical features, Dr Elhardt said the decision regarding whether to perform surgery is not standardised. The decision is ultimately based on the surgeon’s judgement, other factors, and their experience.

An AI-driven deep learning model uses multilayered neural networks to simulate the complex decision-making power of the human brain. This system can identify and analyse key patterns in complex, high-dimensional data, thus streamlining the diagnostic process of a disease like pterygium.

Dr Elhardt and her colleagues developed a deep learning model that processed information on 328 pterygium patients who were then split into two categories based on corneal specialist analysis: surgery indicated and surgery not indicated. Clinical information included in the model were age, artificial tear use, presence of symptoms, ocular surface irritation, sphere shape, astigmatism, and visual acuity.

The clinical raw data was entered into a deep learning model and fed through a rectified linear unit multilayer perceptron neural network. To act as a control, images based on the Scheimpflug principle were fed through a ResNet-18 multilayer neural net as part of the same process. Both data sets were then connected in a final layer of analysis to create a binary result.

The independent corneal specialists placed 55% of the patients in the ‘surgery indicated’ category and 45% in the ‘surgery not indicated’ category. Dr Elhardt reported that the deep learning model achieved a top mark of 95% accuracy compared to the specialist findings when it combined the analysis of both the raw clinical data and Scheimpflug principle control images.

Data processed solely via the deep learning model achieved a 90% accuracy rate, and the information processed via the Scheimpflug principle control images achieved an 82% accuracy rate. Dr Elhardt also found that different factors were prioritised by the deep learning model during its decision-making process, with both sphere size and astigmatism weighed more heavily than other considerations.

In her concluding remarks, Dr Elhardt said the deep learning model they created is capable of assisting physicians in the decision-making process. It should particularly help in achieving higher rates of patient satisfaction by supporting well-considered, evidence-based decisions, which can affect pterygium surgery by a wide variety of factors.

Dr Elhardt presented at the 2025 EuCornea conference in Prague.

Carolin Elhardt MD is an ophthalmology resident at Ulm University Hospital, Germany. carolin.elhardt@uniklinik-ulm.de