A New Method for Retinal Image Semantic Segmentation Based on Fully Convolution Network

Retina images contain a lot of useful information for medical judgment, blood vessel extrusion, the ratio of the arteriovenous width and whether there is lesion area are vital to disease judgment, it is difficult to draft a unified standard for artificial judgment due to subjectivity. Traditional approaches to obtain the three indicators mentioned above include image processing and machine learning, these approaches have relatively poor accuracy or too many restrictions. In order to solve these problems, we propose a customized fully convolutional network, RI-FCN, based on image semantic segmentation for retina image detection. In our proposed method, there are five convolution layers, three down-pooling layers and two up-pooling layers. This structure can classify every pixel into predefined categories and show in different colors and small features can also be presented which is vital in the detection of blood vessel extrusion. Using the RI-FCN model, identification accuracy rate of arteriovenous width ratio, extrusion and lesion area can be increased to 92.23%, 90.99% and 98.13% respectively.