Identification of Corneal Ulcers with Pre-Trained AlexNet Based on Transfer Learning

Artificial intelligence methods are often used in the medical field because they give objective and consistent results. In this study, the SUSTech-SYSU data set was used for the automatic classification of corneal ulcers. Classification procedures were carried out using the pre-trained AlexNet model using fluorescein staining images of corneal ulcers, which were divided into 3 different classes and labeled (3 Categories, 5 Types, 5 Grades) in the dataset. Prior to the training of the pre-trained AlexNet model, data augmentation operations were performed on corneal ulcer images. The images labeled in different classes in the data set were evaluated separately for each class and classification operations were performed. As a result of the experiments, it was determined that the images with the Type label among the classes were the most effective class in detecting corneal ulcers, and as a result of the classification, an accuracy of 80.42% was achieved. After the Type-labeled image class, the Category and Grade classes, respectively, were included in the effectiveness ranking.