Optic Disc Determination in Retinal Images with Deep Features

Abstract

Deep learning has attracted so many researchers in the image processing and machine learning communities. So many new applications have been presented with DL day by day. Thus, in this paper, we propose a hybrid approach to detect the optic disc in retinal images by using deep Convolutional Neural Networks (CNN) and K-Nearest Neighbor (KNN) classifier. More specifically, we extract effective deep features from the fc6 layer of a pre-trained CNN model and classify them into optic disc and non-optic disc classes with KNN classifier. The AlexNet is considered for the pre-trained CNN model which extracts 4096 dimensional feature vector for each patch image. To this end, three retinal image datasets are considered for construction of the training and test sets. 500 optic disc patches and 1565 non-optic disc patches of size 280x280 are collected and then resized to 227x227 for feature extraction and construction of the KNN classifier setting. In addition, 165 retinal images are used for testing the presented hybrid approach. A series of experimental works have been carried out for showing the efficiency of the proposed approach. Accuracy, sensitivity and specificity values are calculated. According to the obtained results, 95.74% accuracy, 84.46% sensitivity and 99.08% specificity values are recorded. These results are quite encouraging for related future works.