Structurally Guided Channel Attention Networks: SGCA-Net

In this paper, we propose Structurally Guided Channel Attention Networks (SGCA-Net), a principled way to guide the channel attention of CNNs. Convolution operator constructs features maps by using both channel and spatial information within the receptive fields of its filters. Prior research has investigated the impact of strengthening the representational power of CNNs using channel attention modules. In this work, we guide the channel attention of networks using feature vectors that contain clinically relevant information. We do so by attaching guided attention modules into a state-of-the-art network architecture, and guiding these attention modules with domain knowledge using feature vectors. Experiments on a dataset of 5512 posterior retinal images, taken using a wide angle fundus camera, show that SGCA-Net achieves 0.983 and 0.948 AUC to predict plus and normal categories, respectively. SGCA-Net achieves higher performance than CNNs without attention modules and CNNs with unguided attention modules.