| Liver cancer is one of the common diseases,and quantitative biomarkers extracted from computed tomography(CT)can create a powerful clinical decision tool for the diagnosis of hepatocellular carcinoma.Currently,manual segmentation of tumors is mainly performed by clinicians,which usually takes a lot of time and resources,so it is necessary to extract the liver and tumor area by computer-aided system.In response to the above problems,this paper carries out a study on liver tumor segmentation method based on U-Net network,and the specific research work is as follows:(1)For liver CT image segmentation,this paper proposes a liver segmentation algorithm based on multi-scale attention network.Taking the 2D U-Net network as the basic network,in order to learn more information of the image and prevent overfitting,the traditional convolution block in the U-Net encoder is replaced with a residual block.Secondly,squeeze and attention modules are introduced in the decoder to combine both spatial and channel aspects to learn image features adaptively and improve liver segmentation accuracy.Finally,in order to solve the problem of feature information loss due to convolution and down-sampling,the multi-scale null convolution module is used to replace the transition and output layers to obtain multi-scale image information through different receiving domains,thus further improving the segmentation accuracy.(2)For tumor CT image segmentation,this paper proposes a liver tumor segmentation algorithm based on an improved 3D U-Net network.3D network solves the problem that 2D network does not learn the spatial information,and this network replaces the convolutional structure in U-Net with a densely connected structure,while introducing an attention mechanism to suppress irrelevant regions,which improves the stability and accuracy of the model.To address the problem of large computation brought by 3D convolution,this paper replaces the traditional convolution with depth-separable convolution(DS-Conv)in the coding layer.The depth-separable convolution reduces the memory requirement and computation cost of GPU to a certain extent and achieves high performance.(3)To validate the performance of the liver tumor segmentation model proposed in this paper,a liver tumor dataset was constructed with a collaborating hospital for testing the tumor segmentation effect to verify the generalization ability and robustness of the proposed model,and further,analyze the effectiveness of the method for liver tumor segmentation and the value of clinical research in this paper.The experimental results show that the two-stage segmentation model proposed in this paper,in which the first network extracts the location of the liver and the second network uses the results of the first network to further segment the tumor,the improved model can complete the tumor segmentation more efficiently and accurately. |
