AN EVALUATION OF EYE DISEASES CLASSIFICATION USING RESNET ON FUNDUS IMAGE DATASET COLLECTED FROM THAI BINH HOSPITAL

Abstract

The difference in the input retinal image size, the influence of the disease labelling process, the number of disease labels, and the disease recognition signs of the left and right retinal images are evaluated based on the ResNet model which has a depth of 101, 2048 channels. The seven datasets from Set-1 to Set-7 are created from the original dataset of 5000 fundus images with different labelling and preprocessing of the input images. The results show that the larger the input image is, the more accurate the results are and the signs of eye disease identification in both eyes are the same. In other words, the training dataset of the left eye can be used to classify diseases of the right eye. The results also show that the data set (Set-6), which contains 12 types of labels, is the most balanced and gives the most accurate classification results at 98.08%.