Deep Learning Based Segmentation Model Of Liver Tumors

Segmentation of liver tumors is one of the important steps in the treatment of liver cancer.However,due to the low contrast and deformed shape of liver tumors in CT images,automatic segmentation is difficult.In addition,the number of para meters in medical liver tumor segmentation models is usually large,which requires the high performance of t he equipment.Therefore,the performance of the liver tumor segmentation model needs to be improved.To solve these problems,two methods of liver tumor segmentation are proposed in this paper:(1)A dense UNet-based liver tumor segmentation method R-Dense UNet was proposedR-Dense UNet is a Denseunet-based method for liver tumor segmentation.It optimizes the Dense UNet network,increases the connection between shallow feature map and deep feature map,and solves the problem of network redundancy.In addition,R-Dense UNet also introduced multi-scale dense junctions and convolution structures of feature maps of different scales to prevent th e loss of liver tumor details during subsampling.The experimental results show that the R-Dense UNet method can further improve the accuracy of liver tumor segmentation results.The R-Dense UNet algorithm was introduced into the Dense UNet subsampling,and its F1 was increased by 0.7397,so as to improve its tumor recognition and classification performance.In summary,R-Dense UNet is a superior method for liver tumor segmentation and can achieve better results in practical applications.(2)A lightweight liver tumor segmentation model AG-VNet based on attention G-VNet was proposedThree key techniques of AG-VNet segmentation method based on attention mechanism: use Ghost module to replace the convolution operation in VNet encoder,add attention mechanism after each layer of convolution operation and apply appropriate loss function.The Ghost module can reduce the number of model parameters,and improve the training speed of the model;The attention mechanism can enhance the feature attention of important information in semantic segmentation.Using the Focal Loss function could solve the problem of uneven positive and negative samples in the liver tumor data set,to fully learn image features and improve the semantic segmentation accuracy of the model.The comprehen sive application of these techniques can make the AG-VNet model achieve better results in liver tumor segmentation.Compared with the classical VNet model and G-VNet model without adding attention mechanism and loss function,the experimental results show that compared with VNet,the total number of parameters in AG-VNet decreases by 3.37 M,and the number of training parameters decreases by3.41 M.In addition,AG-VNet also showed better performance in accuracy,DSC,and Loss.These results indicate that AG-VNet is a better method for liver tumor segmentation and can achieve better results in practical applications.(3)Aiming at the above two methods,AG-VNet was selected as the basic model,and the liver tumor segmentation system was designed and implement ed through system requirement analysis,environment configuration,and interface rendering.