A Framework for Identifying Diabetic Retinopathy Based on patch attention and lesion location

In order to solve the problem that the existing methods in the field of diabetic retinopathy (DR) intelligent diagnosis have not fully exploited the effective DR lesion information in the fundus map, as well as the problem that the traditional attention mechanism have not fully explored the influencing factors of different lesion category in DR grading. This paper proposed a diagnostic method that fuses multi-level patch attention and lesion location. The method contains a multi-level patch lesion attention generator (MPAG) and lesion location module (LLM). The MPAG generates a attention map containing the lesion level imformation of different fundus patches, which is weighted with the fundus map and classified by a global network. The LLM is able to indicate lesion and generate a localization-based global attention and increasing the weights of lesion details in the classification network. This paper demonstrated the effectiveness of the proposed method through extensive experiments on the DDR dataset, obtained an accuracy of 0.8064.