A Pyramid Spatial Attention Network For Fovea Localization

Fovea localization plays an important role in the treatment of Diabetic Retinopathy (DR). Fovea is located in the center of the macula and is the most sensitive part of the fundus structure. Accurate fovea localization is challenging in the cases that the lesions of diseases such as DR and Age-related Macular Degeneration (AMD) appear near the fovea. In this paper, we propose a deep regression network based on a pyramid spatial attention module (PAM) for fovea localization. Fovea localization is obtained from coarse to fine through a three-stage cascade network. The PAM attention module is embedded in the first stage to focus on the spatial relationship between optic disc and macula. A coarse candidate area can be got in the first stage. Then the surrounding structure information such as vessels and colors in the coarse candidate area will be learned in the second and third stages to obtain fine localization. The proposed framework is verified on GAMMA and IDRID public datasets, and experiments show that the robustness of localization is significantly improved by the pyramid scheme and spatial attention.