Research On Segmentation Algorithm Of Liver Tumor CT Image Based On U-Net Network

In recent years,the incidence of liver cancer has been increasing,causing serious harm to people’s health.At present,hepatectomy is still the main treatment for liver cancer.Therefore,it is of great significance for doctors to segment liver tumors from CT images by means of increasingly advanced computer diagnosis and treatment.The existing segmentation methods of liver tumors have the problems of missing small targets and fuzzy segmentation of tumor edges.To solve these problems,this paper studies the segmentation algorithm of liver tumor based on U-Net network structure,which is most commonly used in medical image segmentation.Aiming at the problem of missing detection in segmentation due to the small proportion of liver tumor in background,a segmentation algorithm of liver tumor based on residual attention mechanism U-Net was studied.Adding the attention mechanism to the jump connection of U-Net network,the attention mechanism can make the key area have the advantage of higher weight in the network training process,learn to focus and distinguish the position related to the object of interest,and the attention perception characteristics will change adaptively.In order to improve the feature extraction ability of small target liver tumors,the convolution layer in the network is changed into bottleneck residual structure,which not only improves the network depth,but also reduces the training parameters,and avoids the problem of gradient disappearance.In addition,liver tumors vary in size,especially for smaller tumors,and the resolution of the feature map will decrease after passing through multiple single types of convolution nuclei.Study problem of the scale of the tumor,the attention under the mechanism of residual pyramid U-Net liver tumor segmentation algorithm,using the pyramid convolution convolution kernels of different depth and size from parallel processing input and the advantages of multi-scale feature fusion,will U-Net of all ordinary convolution convolution,replace with pyramid bottleneck residual thought into pyramid convolution structure,In addition to expanding the receptive field,better capture of tumor edge information.The experimental results show that the algorithm not only solves the problem of missing tumors in small target,but also segmentation the details of the edge of tumor region more accurately.Compared with other algorithms,the effectiveness of the proposed algorithm is proved.