An Automatic Liver Segmentation Algorithm for CT Images U-Net with Separated Paths of Feature Extraction

In this paper, a fully convolutional neural network based on U-net is proposed to segment the liver in CT images. Two modifications are made to the original U-net structure. Firstly, an extra path is added to the original net structure to extract the global features and detail features separately. Secondly, the number of convolutional channels of the original contraction path, the original expansion path and the new path is reduced. These two modifications make the training more rapid and improve the efficiency of the convolution kernel extraction feature. Then, the segmentation results before and after modification is compared in terms of performance, including recall rate and precision rate, to ensure that the modified network can reach even higher than the original network precision. After that, the paper analyzes the reasons why our network can maintain good segmentation effect and summarizes the application prospect of the modified network.