Xianxian Luo1, Wenghao Wen1, Jingru Wang2, Songya Xu3, Yingying Gao4, Jianlong Huang1. 1. Faculty of Mathematics and Computer Science, Quanzhou Normal University, Quanzhou 362000, China; Fujian Provincial Key Laboratory of Data Intensive Computing, Quanzhou 362000, China. 2. Department of Ophthalmology, 2nd Affiliated Hospital, Fujian Medical University, Quanzhou 362000, China. 3. Faculty of Educational Science, Quanzhou Normal University, Quanzhou 362000, China. 4. Department of Ophthalmology, 2nd Affiliated Hospital, Fujian Medical University, Quanzhou 362000, China. Electronic address: gaoyingying1968@163.com.
Abstract
PURPOSE: We aim to present effective diagnostics in the field of ophthalmology and improve eye health. The purpose of this study is to examine the capability of health classification of Meibomian gland dysfunction (MGD) using Keratography 5M and AlexNet method. METHOD: A total of 4,609 meibomian gland images were collected from 2,000 patients using Keratography 5M in the hospital. Then, MGD dataset for eyelid gland health recognition was constructed through image pre-processing, labelling, cropping and augmentation. Furthermore, AlexNet network was used to identify the eyelid gland health. The effects of different optimization methods, different learning rates, Dropout methods and different batch sizes on the recognition accuracy were discussed. RESULTS: The results show that the effect of model recognition is the best when the optimized method is Adam, the number of iterations is 150, the learning rate is 0.001, and the batch size is 80, then, the overall test accuracy of health degree is 94.00%. CONCLUSION: Our research provides a reference to the clinical diagnosis or screening of eyelid gland dysfunction. In future implementations, ophthalmologists can implement more advanced learning algorithms to improve the accuracy of diagnosis.
PURPOSE: We aim to present effective diagnostics in the field of ophthalmology and improve eye health. The purpose of this study is to examine the capability of health classification of Meibomian gland dysfunction (MGD) using Keratography 5M and AlexNet method. METHOD: A total of 4,609 meibomian gland images were collected from 2,000 patients using Keratography 5M in the hospital. Then, MGD dataset for eyelid gland health recognition was constructed through image pre-processing, labelling, cropping and augmentation. Furthermore, AlexNet network was used to identify the eyelid gland health. The effects of different optimization methods, different learning rates, Dropout methods and different batch sizes on the recognition accuracy were discussed. RESULTS: The results show that the effect of model recognition is the best when the optimized method is Adam, the number of iterations is 150, the learning rate is 0.001, and the batch size is 80, then, the overall test accuracy of health degree is 94.00%. CONCLUSION: Our research provides a reference to the clinical diagnosis or screening of eyelid gland dysfunction. In future implementations, ophthalmologists can implement more advanced learning algorithms to improve the accuracy of diagnosis.