A Deep Learning-Based Classification Study for Diagnosing Corneal Ulcers on Ocular Staining Images

Abstract

Corneal ulcer is a common disease worldwide and is one of the leading causes of corneal blindness. Diagnosis of the disease requires expertise, and the number of experienced ophthalmologists is not sufficient, especially in underdeveloped countries. For this reason, it is necessary to develop technology-based decision support systems in the diagnosis of the disease. However, the number of studies on this subject is not sufficient. In this study, CNN-based classifications were performed using corneal ulcer images obtained by an ocular staining technique, consisting of 712 samples and three classes. In addition to the AlexNet and VGG16 state-of-the-art architectures, which are widely used in the literature, a CNN model proposed for this study was used for classification. In the classifications performed by applying data augmentation, 95.34% accuracy with AlexNet, 98.14% with VGG16, and 100% accuracy with the proposed model was obtained. The findings were compared with similar studies in the literature. It was concluded that the accuracy rates obtained with all of the models used in the study were generally higher than similar studies in the literature, and the accuracy obtained with the proposed CNN model was higher than all of the peers. In addition, the success of the proposed model compared to other models with more complex structures revealed that it is not always necessary to use complex architectures for high accuracy.