Convolutional Neural Network for Classifying Retinal Diseases from OCT2017 Dataset

Abstract

Optical Coherence Tomography (OCT) is an imaging modality that offers real-time, non-invasive high-resolution imaging in the biomedical field. It is widely utilized in ophthalmology to perform diagnostic imaging of the anterior eye and retina structures. Several methods based on traditional image processing and classical machine learning have been widely applied to detect and classify retinal diseases from OCT images with various weaknesses, particularly complex rules and long processing times. Recently, several studies of deep learning in multiple fields, including medical, have manifested promising results. However, it is still rarely explored to detect retinal disease on OCT images. Thus, in this study, we performed several state-of-the-art deep learning image classification methods to confirm the best for the sizeable OCT2017 database accommodating eighty thousand images. MobileNet-V2 achieved the highest accuracy of 99.6% compared to the others. The model achieved high performance of accuracy and has a fast computational time at 0.0124 seconds/image. Hence, it is promising to implement in real-time while assisting ophthalmologists.