A deep learning approach for automated classification of Corneal Ulcers

Abstract

Eye Corneal Ulcer(ECU) has been demonstrated to be the second most common cause of treatable blindness worldwide, after cataracts. It is an extremely prevalent ophthalmic ailment and can cause severe visual impairment or perhaps total blindness. This thesis renders a comprehensive study on the automated classification of corneal ulcers using a deep learning approach. In this research, the SUSTech-SYSU dataset has been utilized which is obtained from Sun Yat-sen University’s Zhongshan Ophthalmic Center, consisting of 712 images of patients with various types, grades and categories of corneal ulcers. These ocular surface images are captured after fluorescein staining, aiding as a valuable resource for the enhancement of deep learning models. The images in the dataset having dimensions of 2592 pixels in width and 1728 pixels in height, depicts close-up views of corneal abrasions under cobalt blue light during eye examinations, which is the particular type of image captured in this dataset. This thesis occupies the deep learning Convolutional Neural Networks (CNN) architecture, which includes InceptionV3, ResNet50 and VGG16 in order to create a pre-trained model for Eye-Corneal-Ulcer (ECU) image classification. In addition, a customized model is built to foster validation and test accuracy. The deep learning models are run on training and testing sets, enabling them to recognize unique criteria and patterns linked with different types of corneal ulcers. For data training, the dataset has been allocated by dividing it into separate folders based on the type of ECU images. Data augmentation is exerted by using the ImageData- Generator to escalate the diversity of the dataset and improve model generalization. The dataset comprises 10,000 training images and 2,000 testing images for evaluating the model. In the customized model, a sequential architecture is implemented, including layers such as Conv2D, max pooling, batch normalization, flatten, and dense layers for feature extraction and classification. For multi-class classification, categorical cross-entropy is employed as the loss function, and the Adam optimizer is used. Hyperparameter tuning has been enacted using the validation set, encompassing various learning rates, batch sizes, and regularization techniques to optimize the performance of the model. The consequence of this research avails the development of automated corneal ulcer classification, potentially facilitating ophthalmologists in diagnosing and curing corneal infections more productively.