GC-Net: Global and Class Attention Blocks for Automated Glaucoma Classification

Glaucoma is an irreversible vision loss, which develops gradually without obvious symptoms. It is hard to detect in early stages and diagnostic procedure is a time-consuming work. Therefore, early screening and treatment are essential to protect vision and maintain quality of life. In previous work of glaucoma classification, convolutional neural network (CNN) has been used in lots of researches and got a good performance. However, the convolution operator only focuses on local information in feature extraction and context information will be lost to a large extent. Attention block pays more attention to global information, which has full coverage of the whole feature extraction. In this paper, a novel CNN model embedded with two attention blocks is proposed. Global attention block (GAB) has advantages on extracting global attention maps and focusing on context information for fundus images. We also put forward class attention block (CAB) to focus on the characteristics of each disease category and reduce the impact of data set imbalance. By combining the above modules and CNN backbone, our GCNet is constructed for glaucoma classification task, which can be trained in an end-to-end manner. We verify our model through two public dataset experiments and both of them show that our global and classes attention network (GC-Net) produces the best performance compared with the baseline CNN models and other existing state-of-the-art deep learning models.