Transfer Learning Coupled Convolution Neural Networks in Detecting Retinal Diseases Using OCT Images

Abstract

Optical coherence tomography (OCT) in diagnosing retinal images is an extensive technique for detecting the wide-ranging diseases related to retina. In this paper, the authors have considered three diseases, viz. diabetic macular edema (DME), choroidal neovascularization (CNV), and drusen. These diseases are classified using six different convolutional neural network (CNN) architectures. The purpose is to compare among the six different CNNs in terms of accuracy, precision, F- measure, and recall. The architectures used are coupled with or without transfer learning, and a comparison has been drawn as to how the CNN architectures work when they are coupled with or without transfer learning. A dataset has been considered with the mentioned retinal diseases and no pathology. The designed models could identify the specific disease or no pathology when fed with multiple retinal images of various diseases. The training accuracies obtained for the six architectures, viz. four convolutional layer deep CNNs, VGG (VGG-16 and VGG-19) and Google’s Inception [Google’s Inception v3 (with or without transfer learning)], and Google’s Inception v4, are, respectively, 87.15%, 91.40%, 93.32%, 85.31%, and 83.63%, respectively, while the corresponding validation accuracies are 73.68%, 88.39%, 86.95%, 85.30%, and 79.50%. Thus, the results so obtained are promising in nature and establish the superiority of the proposed model.