Dry Eye Disease Classification Using Vision Transformer Deep Learning Technique

Abstract

Eye disease is any condition or disorder that affects the eyes or the visual system, and dry eye disease (DED) is one of the most common conditions affecting millions worldwide. DED is a prevalent ocular condition characterized by inadequate tear production or excessive tear evaporation, leading to discomfort and potential vision impairment. Accurate diagnosis and classification of subtypes of DED are crucial for appropriate treatment selection and management. There are problems on classification methods for DED, and its often rely on subjective assessments, symptom questionnaires, and clinical examinations, which can introduce variability and subjectivity. It also takes significant time and effort to identify the condition and may result in an incorrect selection of medications to treat the disease. Therefore, there is a need to explore more objective and efficient techniques for the diagnosis and classification of DED. This study mainly focuses on classifying DED using the deep learning technique of vision transformers and employs an experimental research design incorporating various image processing techniques. Eight hundred forty-six (846) sclera DED images were used as Ambo University Referral Hospital datasets. Of 846 image datasets, 80% (676/846), 10% (84/846), and 10% (84/846) were used for training, validation, and testing, respectively. Transfer learning-based pre-trained Vit-base models (ViT_B_32) were used for our experimentation. The model evaluation indicated that the Vit-base performance achieved 88% accuracy, an AUC value of 0.8, and a loss value of 0.19. A pivotal focus was on transformers, recognized as potent DL models in Computer Vision (CV) and Natural Language Processing (NLP) tasks. Leveraging this knowledge, we innovated the Vision Transformer (ViT), a versatile model specifically designed for the classification of Dry Eye Disease (DED). However, our model still needs a lot of work. We plan to build on this work by incorporating additional datasets, including other Eye Diseases, customized treatment plans, and others, to make it more robust, accurate, and diverse. This will also allow for greater portability and creative collaboration, enhance the accuracy and efficiency of DED classification, and contribute to advancing research in ocular health.