G-EYENET: A CONVOLUTIONAL AUTOENCODING CLASSIFIER FRAMEWORK FOR THE DETECTION OF GLAUCOMA FROM RETINAL FUNDUS IMAGES

摘要

Glaucoma is one of the leading causes of visual impairment in the world. It deteriorates the optic nerve fibers over time, and cannot be cured once it reaches the later stages. Hence early detection is of utmost importance for the aging society. In this paper, we propose a novel deep learning multi-model network termed G-EyeNet for glaucoma detection from retinal fundus images. G-EyeNet consists of a deep convolutional autoencoder and a traditional convolutional neural network (CNN) classifier sharing the encoder framework. The multi-model network is jointly optimized for minimizing both image reconstruction error and the classification error based on a multi-task learning procedure. Extensive training experiments are performed on publicly available datasets HRF, DRISHTI-GS, RIM ONE v.3. and evaluated on the publicly available DRIONS-DB dataset. Unsupervised training of the encoder framework helps in learning a good distribution of the input, which helps in classification. This architecture is especially effective when the training dataset is small, which is usually the case in medical imaging. Experimental results show that an area under the curve (AUC) of the receiver operator characteristic (ROC) curve of 0.923 is achieved, which is better compared to the state-of-the-art deep learning algorithms.