Deep Learning in Image Classification using Dense Networks and Residual Networks for Pathologic Myopia Detection

Abstract

Half of the population in the world predicted will have myopia and one-tenth of the population will have high myopia. Pathologic myopia is the most dangerous form of myopia that can lead to vision loss permanently. The definition of it was updated as the unusual cases were reported. The latest definition of pathological myopia is eyes with posterior staphyloma or myopic maculopathy equal to or higher than category 2 in META-PM. Detection of pathologic myopia requires a high cost because of insufficient specialists worldwide. To produce an efficient cost, artificial intelligence for health care is rapidly adopted. Several ophthalmology studies have been conducted using retinal fundus photographs such as diabetic retinopathy, cataract, age-related macular disease, and pathologic myopia. Nevertheless, pathologic myopia detection has still been a scarce resource due to the unstandardized definition yet. In this study, a public dataset is used. There are 612 images available distinguished into two classes: normal eye and pathologic myopia eye. The augmentation technique was used to create a robust model. ResNet and DenseNet architecture are performed on two different preprocessing and splitting data. Each model also used three variations of the optimizers such as SGD, RMSprop, and Adam to work out which optimizer performs better and fine-tune the learning rate each time the model stops improving. The results showed that the best model on this proposed method provides accuracy, sensitivity, and specificity of 97%, 93%, and 100%. It performed on DenseNet architecture with normalization and standardization preprocessing, 70:20:10 type of data split, and adam optimizer.