Image Segmentation Method Of Liver Tumor Based On Improved V-Net Model

At present,liver malignant tumor is still one of the most common cancers in the world.It is also the main cause of human death and poses a huge threat to human health.In order to effectively reduce cancer mortality,patients still need to have a comprehensive physical examination in advance and receive treatment as soon as possible.However,due to the low contrast between the liver tumor and its surrounding soft tissues,the boundary is not obvious,and for different patients,the shape,volume,number and location of the liver tumor are significantly different.At the same time,a variety of scanning schemes lead to all kinds of problems such as noise interference in CT images.Due to the difficulty of liver tumor segmentation,it is hard to obtain accurate segmentation results only by relying on the naked eye of doctors.This paper proposes a method to achieve accurate segmentation of liver tumors based on the V-Net model and using 3D CT images as the input of the model.In order to improve the segmentation accuracy,some improvements to the original model are proposed.First of all,because in the CT image of liver tumors,in addition to the tumor area of interest,there are some background information such as soft tissues that are not needed,and the original V-Net network cannot identify the tumor information well.Compared with the small lesions,those large lesions are easier to locate and find.It is proposed that the attention mechanism module is integrated into the original network so that the model(AGV-Net)can focus more on the regions of interest and smaller tumors.Then,due to the existence of extremely small liver tumors in the CT image,it occupies a relatively small proportion in the entire CT image,the original V-Net model may have problems such as missed detection.While the Dice loss function in the model focuses on the global information of the image and has a negative impact on the back propagation.To solve this problem,the combined loss function is proposed based on the attentional mechanism model(AGV-Net),and the experimental results show that the proposed improvement has certain effect on the segmentation of liver tumor.Finally,in view of the problem that the increase of network layers will lead to the slow convergence speed of the model,it is proposed that add batch normalization layer(BN)and group normalization layer(GN)to the AGV-Net network to normalize the input data and improve the convergence speed and segmentation accuracy of the model.Then,because the input data is three-dimensional,there will be problems such as poor computing ability of the network model caused by excessive network parameters.Therefore,the deep separable convolution is proposed to replace the ordinary convolution to improve the computing speed of the network model.Experiments have proved that the convergence speed of the model is faster after adding the normalization layer to the network,and the convergence speed of the network when the group normalization layer is added is faster than the convergence speed of the batch normalization layer network.In addition,after using the deep separable convolution,the parameter amount of the network has been reduced to a certain extent,and the calculation speed of the model has also been improved a lot.According to the segmentation index table of the network model,it can be seen that the proposed DN-AGV-Net network has a greater improvement in various indicators than previous comparison models,and can segment liver tumors well.