Unlocking diagnosis potential: CNN in multi-class classification of corneal ulcer

Abstract

This research explores the multi-class classification of Corneal Ulcers using a deep-learning approach. The SUSTech-SYSU dataset, obtained from Sun Yat-sen University's Zhong-shan Ophthalmic Center, contains 712 images of patients affected with various types, grades, and categories of Corneal Ulcers. The captured dataset after fluorescein staining is used to improve deep learning models. A deep learning convolutional neural network (CNN) architecture is applied to train pre-Trained models for ECU image classification, and a customized model is generated to improve validation and test accuracy. The improvised dataset contains 7,200 training images,3,000 testing images, and 1,800 validation images for evaluation. The customized model uses a hierarchical architecture for feature extraction and classification, employs the loss function with categorical cross-entropy, and uses the Adam optimizer for multi-class classification. Hyper-parameter tuning is performed using the validation set to optimize model performance. The customized model validation accuracy is 90%, and the training accuracy is 99%. This research aims to develop automated Corneal Ulcer classification, potentially improving ophthalmologists' productivity in diagnosing and curing corneal infections.