Attention-Aware Convolutional Neural Network for Age-Related Macular Degeneration Classification

Though age-related macular degeneration (AMD) poses an important personal and public health burden, studies on AMD is hampered by different approaches to classify AMD. In this paper, we propose convolutional neural networks (CNN) based models for fundus retinal images that classify four types of AMD automatically. We use deep residual network (ResNet50) to extract high-dimensional features and be trained end-to-end to classify AMD. Furthermore, we apply attention mechanism to deep residual network (Atten-ResNet) which enables to further select features adaptively. Experimental results show that comparing to HOG-SVM and Visual Geometry Group (VGG), the ResNet50 based method could achieve 17.2% and 12.1% overall classification accuracy improvement. The Atten-ResNet based method has more 0.4% accuracy improvement than ResNet50 based method.