Deep Learning in Opthalmology: Iris Melanocytic Tumor Intelligent Diagnosis

Recently, Convolutional neural networks (CNN) have shown a growth due to their ability of learning different level image representations that helps in image classification in different fields. These networks have been trained on millions of images, so they gained a powerful ability of extracting the rightful features from input images, which results in accurate classification. In this research, we investigate the effects of transfer learning based convolutional neural networks for the iris tumor malignancy identification as it is notoriously hard to distinguish an iris nevus from an iris tumor. Features are transferred from a CNN trained on a source task, i.e. ImageNet, to a target task, i.e. iris tumor datasets. We transfer features learned from AlexNet and VGG-16 that are trained on ImageNet, to classify three different iris images types which are: iris nevus unaffected, iris cysts, and iris melanocytic tumors. The employed pre-trained models are modified by replacing their feedforward neural network classifier, Softmax, by a support vector machine (SVM) that is expected to slightly boost their performance (AlexNet-SVM and VGG16-SVM). All employed models are trained (fine-tuned) on a 60% of the available large dataset of iris images in order to investigate their power of generalization when trained using large amount of data. The networks are also tested on 40% of the data. The best performance was achieved by the VGG16-SVM which scored a high accuracy of 96.27% and strong features extraction capability as compared to the other models. Experimentally, it was seen that adding SVM contributed in improving the network performance compared to original models which use a feedforward neural network classifier.