Dual-Branch Attention Network and Atrous Spatial Pyramid Pooling for Diabetic Retinopathy Classification Using Ultra-Widefield Images

Diabetic Retinopathy (DR) is a very common retinal disease in the world, which can affect vision and even cause blindness. Early diagnosis can effectively prevent the disease, or at least delay the progression of DR. However, most methods are based on regular single-view images, which would lack complete information of lesions. In this paper, a novel method is proposed to achieve DR classification using ultra-widefield images (UWF). The proposed network includes a dual-branch network, an efficient channel attention (ECA) module, a spatial attention (SA) module, and an atrous spatial pyramid pooling (ASPP) module. Specifically, the dual-branch network uses ResNet-34 model as the backbone. The ASPP module enlarges the receptive field to extract rich feature information by setting different dilated rates. To emphasize the useful information and suppress the useless information, the ECA and SA modules are utilized to extract important channel information and spatial information respectively. To reduce the parameters of the network, we use a global average pooling (GAP) layer to compress the features. The experimental results on the UWF images collected by a local hospital show that our model performs very well.