Diabetes would increase the risk of retinal vascular disease, and may further develop into diabetic retinopathy in severe cases. Among all pathological features of diabetic retinopathy, microaneurysms, bleeding, hard exudates, and soft exudates are usually typical. In recent years, with the development of deep learning, intelligent assisted diagnostic medicine has become a trend. The premise of intelligent aided diagnosis is that the corresponding lesion area can be extracted qualitatively and quantitatively. Therefore, a model of fundus lesion detection is proposed in this paper with cascade architecture parameter optimization, which effectively solves the multi-scale and small target problems of fundus lesions. The comprehensive test accuracy of detecting lesions on DDR dataset can reach 0.380, which is better in detection performance than the current mainstream detection network.