Automated classification of dry eye type analyzing interference fringe color images of tear film using machine learning techniques

Abstract

The unstable balance in secretions of lipids and aqueous fluid to tear film is a significant cause of dry eye disease (DED). Arita et al. demonstrated a simple but very effective method that classifies dry eye types to the aqueous deficient dry eye (ADDE) and the evaporative dry eye (EDE) by focusing on the dry eye type-unique appearances of interference fringe colors and patterns of tear films. We thought this simple classification is very helpful for diagnoses and treatments. However, diagnostic bias by unskilled observers remains an issue to be solved.

The artificial intelligence (AI)-based support for diagnosis is one of the hottest topics in the field of ophthalmology research. We expected that the AI-based model would reduce bias in DED-type diagnoses. Many studies have been reported targeting retinal diseases like age-related macular degeneration and/or diabetic retinopathy. Most of the works established AI-based predicting models using images taken by fundus cameras and/or optical coherence tomography (OCT) devices to capture disease-related structural disorders. In contrast, the interference fringes dynamically change the colors and patterns spatiotemporally. To the best of our knowledge, there is no AI-based model studied for distinguishing ADDE and EDE using interference fringe images. However, an AI-based study classifying the condition of the tear lipid layer by analyzing the textures of interference fringes compared to the device-unique grades has been reported. This suggested the possibility of using the unstructured characteristics, such as colors and/or complexities of interference fringes, as the numerical image features when building AI-based prediction models. In this study, we first examined several types of image characteristics extracted from the colors and patterns of fringes to obtain effective image features for the DED-type classification. We then evaluated whether the AI-based models would have sufficient abilities for this type of prediction by comparing their diagnoses with those made by an ophthalmologist skilled in this classification (the founder of this type classification).