Optimising deep learning models for ophthalmological disorder classification
S. Vidivelli, P. Padmakumari, Chembian Parthiban, A. Dharunbalaji, R. Manikandan, Amir H. Gandomi
Abstract
Fundus imaging, a technique for recording retinal structural components and anomalies, is essential for observing and identifying ophthalmological diseases. Disorders such as hypertension, glaucoma, and diabetic retinopathy are indicated by structural alterations in the optic disc, blood vessels, fovea, and macula. Patients frequently deal with various ophthalmological conditions in either one or both eyes. In this article, we have used different deep learning models for the categorisation of ophthalmological disorders into multiple classes and multiple labels utilising transfer learning-based convolutional neural network (CNN) methods. The Ocular Disease Intelligent Recognition (ODIR) database is used for experiments, and it contains fundus images of the patient's left and right eyes. We compared the performance of two different optimisers, Stochastic Gradient Descent (SGD) and Adam, separately. The best result was achieved using the MobileNet model with the Adam optimiser, yielding a testing accuracy of 89.64%.