Hybrid Prior-Guided Diffusion Network for Enhanced Diabetic Retinopathy Diagnosis

Diabetic retinopathy (DR) is one of the leading causes of blindness in adults, so early diagnosis is crucial. Previous studies on the diagnosis of diabetic retinopathy have had difficulty in capturing the dependencies between lesions. In order to model the long-distance dependencies in retinal images, accurately locate the relevant key regions for feature extraction, and cope with unexpected noise in the images, we propose a diagnostic model for diabetic retinopathy diagnosis called hybrid a priori guided diffusion network (HPGNet). The network is based on the standard diffusion model and uses a hybrid prior generator (HPG) to generate global and structurally relevant priors to guide the diffusion process. In the HPG, Bi-Level Routing Attention (BRA) and Feature refinement module (FRM) work in tandem to enhance feature selection and efficiently capture interdependencies between lesion regions. The proposed model has achieved outstanding results on both the APTOS 2019 dataset and the Messidor-2 dataset, outperforming other representative methods.