Retinal Image Segmentation Based On Semantics And Multi-scale Information

In recent years,the increasing popularity of electronic products has brought convenience to people.At the same time,due to long-term irregular use of eyes,the prevalence of ocular diseases is also increasing year by year.Some of ocular diseases can lead to blindness if they are not diagnosed and treated in early times.The computer-aided diagnosis system on retinal images can quickly and effectively solve the diagnosis problems of ocular diseases by analyzing the morphological characteristics of the vessels and centerlines in the retinal image.Therefore,it is of great significance to obtain accurate results of retinal vessels and centerline segmentation for auxiliary diagnosis of ophthalmology.In recent years,there have been a lot of studies to solve the problem of high-precision segmentation of retinal vessels and centerlines.Most of these methods are based on encoder-decoder structures,adding well-designed modules to improve segmentation results.However,these methods don’t make full use of the semantic features and multi-scale features of the network.High-level semantic features and multi-scale features can provide positioning and global information for blood vessels and centerlines and solve the problem of the large changes in vessels scale,so they are crucial to improve the segmentation accuracy.This paper proposes two segmentation algorithms in order to make better use of the semantic and multi-scale features of the network and further improve the segmentation accuracy of blood vessels and centerlines:(1)Semantic and multi-scale aggregation network.This network mainly solves the problem of accurate segmentation of retinal vessels.The semantic aggregation module is designed to integrate high-level features into the shallow layer and make full use of semantic information to promote capillary recognition.A multi-scale aggregation module is designed to extract information at different scales to solve the problem of different vessel sizes and large scale changes.(2)Deep semantic and multi-scale cross-task aggregation network.The network segments blood vessels and centerlines by joint-learning way.The front network extracts the common features of blood vessels and centerlines by parameter sharing,and fuses high-level semantic information for preliminarily segmentation of blood vessels and centerlines.The end network use the unique but complementary information of the two tasks to further improve the preliminary segmentation results.In this paper,the two proposed segmentation algorithms are verified on several public fundus datasets respectively.The experimental results show that the two algorithms can achieve high results in the two tasks of blood vessel segmentation and centerline extraction,and surpass several high-precision segmentation algorithms in recent years,which also lays a good foundation for the construction of computer-aided diagnosis system based on retinal image.