Improved Detection of Fundus Lesions Using YOLOR-CSP Architecture and Slicing Aided Hyper Inference

Diabetes affects millions of people worldwide, and diabetic retinopathy is a complication of diabetes. Brazil is the sixth in the world in the incidence of diabetes and has the highest numbers in Latin America with 15.7 million adults affected by this condition. Typically diabetic retinopathy is identified by lesions such as hard and soft exudates, microaneurysms, and vitreous hemorrhages. Early diagnosis of these lesions is essential to prevent the progression of the disease to severe stages and the consequent loss of vision. As the disease diagnosis is based on image analysis, it is possible to use deep learning models to detect artifacts in the retina. This article proposes a new method that uses a YOLOR-CSP architecture combined with the Slicing Aided Hyper Inference framework to detect fundus lesions. The proposed method was trained, adjusted, and evaluated using the DDR dataset, obtaining a mAP equal to 38.08%. The proposed method achieved values of AP equal to 40.90%, 46.60%, 26.10%, and 38.70% for hard exudates, soft exudates, microaneurysms, and vitreous hemorrhages, respectively, surpassing similar studies found in the literature.