Abstract: To address the issues of misgrading and insufficient focus on lesion edge information in diabetic retinopathy images,a retinopathy grading algorithm combining dynamic spatial Transformer and multi-level fusion is proposed.Firstly,the retinal images are processed through the PVT v2 backbone network for initial extraction of lesion information.Secondly,a contour enhancement module is introduced in the first three layers of the network to highlight lesion edge features,thereby improving the algorithm’s localization perception of lesion pixels.Thirdly,a dynamic spatial attention module is designed at the network’s lower layers to effectively connect global and local spatial information,enhancing the algorithm’s ability to extract deep semantic information.Finally,a multi-level gated fusion module is constructed to filter out non-diagnostic information while performing multi-level fusion of diagnostic information,further improving the accuracy of retinopathy grading.Experiments on IDRID and APTOS 2019 datasets show that the QWK are 91.71% and 89.89% respectively,the Acc on IDRID dataset and the AUC on APTOS 2019 dataset are 79.61% and 93.06% respectively.The experimental results demonstrate that the proposed algorithm has significant application value in the field of retinopathy grading.

Key words: retinopathy grading, dynamic spatial attention, contour enhancement module, multi-scale gated fusion module