Residual Networks and Deep-Densely Connected Networks for the Classification of retinal OCT Images

Abstract

Diabetic macular edema (DME) and drusen macular degeneration (DMD) are two se-vere vision-threatening diseases affecting the mac-ula of the eye. This study presents deep learning based classification models for retinal optical co-herence tomography (OCT) images that distinguish between healthy eyes, DME and DMD cases. The work involves the use of residual models, and deep and densely connected networks for the analysis of OCT images. This involves pre-initialisation of the model using the ImageNet dataset, followed by fine-tuning with OCT images. Both models are very powerful and suitable for real-time use in clinical practise. ResidualNets use skip connections, where the output of the previous layer is added to the layer before it. DenseNets use dense connections between the convolutional layers of the network, which allows deeper supervision between layers. This makes it easier for the model to learn the complex feature maps of the images of OCT in each layer of the network. The models are trained and evaluated using the Mendeley OCT dataset, a publicly available SD-OCT dataset for the retina. We calculate the F1 score, accuracy, precision and recall to evaluate the models. The models provide excellent performance without requiring any pre-processing steps. The promising performance of the computerised systems prove that they can serve as automatic recognition tools to assist ophthalmologists.