The Application of Deep Learning in Vessel Extraction with Fundus Image

Abstract

Fundus image is a very prevalent Biomedical data, so it is very important to choose a method to extract blood vessel from fundus images. Extracting vessel from the fundus image can be used for personal identity verification and assist doctor to diagnose diseases [1] such as diabetes. However, traditional extracted methods such as manually annotation area waste of time and energy. There is some another method [2] based on the image recognition, extracting the feature with the filter and segment with mathematical morphology method. But in the fact of the application, this method has low stability and accuracy. Traditional machine learning methods such as SVM [3] are also used for vessel extraction, but this method can only apply to dealing with one-dimensional data while the images showed up as matrix form. So, we have to stretch a matrix line by line or column by column into one-dimensional vectors [4-6]. However, by doing so, we will be face that some important information can’t be reflected because the position relation of rows and rows in the original matrix will be lost. Deep learning can process image data naturally and extract features automatically with convolution, and it is also popular for its good learning ability and high accuracy. U-net [7] is a convolutional network based on deep learning and widely used in biological image extraction, which constructs a U-shaped network from input image and output image. In this paper, building a U-Net convolutional network, which takes fundus image as input and vessel image as output, to extract vessels in fundus image. We established a training set with 20 fundus images and a test set with 20 fundus images. Finally, the accuracy rate of our U-net model on the training set can reach 99 percent, while the accuracy rate on the test set was stable above 99.5 percent.