Scale-up screening of retina is important for prevention, early diagnosis and timely intervention and treatment of related diseases. In this paper, the construction of intelligent diagnosis system for retinal diseases is based on deep learning technology tools. The fundus image is pre-processed by performing region segmentation and extraction, image denoising and region feature enhancement. Based on the structure of retinal layers, the encoding of retinal boundaries is carried out through a dual boundary representation with the existence of complementary constraint relations, and multi-task consistency constraints are proposed correspondingly. Under the retinal layering of the dual-task boundary concern method, a fully convolutional network is introduced to extract a lesion feature and feed back the structural and spatial location information of this feature. The average performance of the system solution for boundary location segmentation of lesion images (9.80±7.26) is extremely similar to that of manual labeling data (8.08±5.66), and the average accuracy of segmentation of the four major lesion metrics is as high as 0.6773, which is capable of assisting in the localization and detection of lesion features with accurate and efficient retinal layer segmentation performance
