Vessel Segmentation Method of Fundus Image Based on Multi-scale Features of Fully Convolutional Neural Network

Abstract

A blood vessel segmentation method for fundus images based on multi-scale features of fully convolutional neural networks, comprising the following steps: 1) Preprocessing the retinal images of the fundus; 2) Segmenting the preprocessed images into image blocks for data expansion; 3) Construct a convolutional neural network model, and use the expanded data for network training; 4) Test the trained model to obtain segmentation results. In the present invention, by connecting one encoding and two different decoding structures, and adopting multiple skip connections, it can overcome the shortcomings of the small number of blood vessel image data sets and the low segmentation accuracy caused by low image quality, and more fully integrate images of different depths. Features, and effectively alleviate the gradient disappearance problem caused by the increase of network depth, compared with traditional segmentation methods, it has higher accuracy and higher robustness.

Description

technical field [0001] The invention relates to the fields of medical image processing and computer vision, in particular to a method for segmenting blood vessels in fundus images based on convolutional neural networks. Background technique [0002] The fundus retinal microvascular network is the same as the face, fingerprints, palm prints and other human characteristics, and also has individual uniqueness. The distribution, direction, thickness, curvature and other characteristics of the retinal blood vessels are different for each person, so the fundus information can be used for biometric identification. Furthermore, of all the deeper microvascular circulatory networks in the human body, the only one that can be directly visualized noninvasively is that of the retinal microvasculature. Normally, the vascular network remains unchanged for a long time, but cardiovascular and cerebrovascular diseases such as diabetes and hypertension can lead to changes in the structure and ...

AI Technical Summary

Problems solved by technology

Unsupervised segmentation methods do not require prior label information, but have high requirements for image quality, usually require more additional conditions, and the segmentation effect is slightly worse

The supervised method is mainly based on the extracted feature classifier to achieve the purpose of identifying blood vessels and background. The neural network model needs to extract image features layer by layer, and it is easy to lose a lot of useful information, resulting in the neural network not being able to segment blood vessel images well.

Images