Liver Tumor Segmentation Network Based On Feature Fusion And Edge Guidance

In clinical,it is very important to use CT and other medical images to diagnose liver diseases.However,due to the large changes in the shape and size of liver tumors,the current automatic method for segmenting liver tumors cannot meet clinical needs.Therefore,automatic and accurate segmentation of liver tumors remains a major difficulty.The design of this network is mainly considered from two aspects:enhancing the global context extraction ability of the network and enhancing the network’s ability to perceive image edge information,so as to improve the accuracy of the network for liver tumor segmentation on abdominal CT images.The work of this paper mainly consists of the following parts:(1)This paper proposes a feature fusion network based on global and local Feature Fusion(GLF-Net).The network adopts the classic encoder-decoder structure,and a double-branch encoder is designed.Among them,the global feature extraction branch uses a gated axial multi-head self attention layer for feature extraction,and the local feature extraction branch uses ResNet34 as the basic encoding module.This paper also proposes a global and local feature fusion module.This module first implements a spatial attention mechanism for local features to more accurately focus on spatial information.Then,the processed local features are fused with the global features obtained from the global feature extraction branch through the feature fusion module.This module well integrates the global information provided by gated axial multi-head self attention layer and the local information provided by convolution.(2)This paper proposes an edge-guided fusion network(EGF-Net).This network is improved on the basis of GLF-Net.There are two improvements.One is the addition of an edge feature coding module to fuse the first and deepest information from local feature extraction branches,and adds an edge supervision to the fused features,making the fused features more sensitive to the edge of the target.Finally,the feature is passed to the decoder of each stage to realize the edge guidance of the network.Another improvement is the improvement of the global and local feature fusion module of the previous network,which reduces the amount of network parameters without losing the segmentation accuracy.(3)This paper verifies the superiority of the proposed model and the effectiveness of each module of the model through a series of ablation experiments and comparative experiments.The generalization of the network is demonstrated through other public datasets and clinical datasets provided by the Second Hospital of Shandong University.The design of the network in this article mainly considers two aspects:enhancing the global context extraction ability of the network and enhancing the network’s perception of image edge information.This article proposes a segmentation network based on global and local Feature Fusion(GLF-Net)and an edge guided feature fusion network(EGF-Net).The effectiveness of the proposed networks were verified through a series of ablation and comparative experiments conducted on a public dataset.The generalization of the network is demonstrated through other public datasets and clinical datasets provided by the Second Hospital of Shandong University.