Artificial Intelligence-Based Anterior Segment Imaging For Dry Eye Disease

Dry eye disease (DED) that affects up to half the population is demanding in terms of healthcare burden. We aim to assess whether artificial intelligence (AI) can be harnessed to identify patients with DED who need specialist care using images that can be acquired in community settings.

Technology such as artificial intelligence (AI) can be harnessed to identify people who need specialist care, and facilitate the care continuum between specialists and primary care practitioners. Managing a high volume of people with minimal manpower resources will increase care sustainability and make healthy living accessible to all.

123 patients managed in the cornea clinic in Singapore National Eye Centre from 2017 to 2021 were involved in this study.

High quality cornea images were acquired from patients with dry eye symptoms in a tertiary hospital with a slit-lamp biomicroscope and cobalt blue filter after fluorescein dye instillation. Two trained ophthalmologists independently classified images into 160 abnormal (referral) and 101 normal\minimal staining (non-referral) images. We developed a deep learning system for automated segregation of corneal staining (split 7:1:2 for training : validation : testing using the EfficientNet_b0 image classification model)

We validated this system using two datasets. For the internal test set, area under curve (AUC) = 0.97, with internal test accuracy of 90.4%. For the threshold of 0.50, the specificity is 100%, sensitivity is 84.4% and the Matthews Correlation Coefficient (MCC) is 0.822. For the external test set, AUC = 0.960 and external test accuracy of 77.4%. For our threshold of 0.50, the specificity is 100%, sensitivity is 63.2% and MCC is 0.632.

A single image of the stained ocular surface may be able to assist decision making in the primary care setting and patients with milder conditions can then be managed in primary care and rescreened at long intervals.