Research On Automatic Segmentation Algorithm Of Liver Tumor Based On CT Image

Image segmentation is a key research topic in the field of computer vision.It is also especially important in medical image processing.It can obtain lesions of interest to doctors in complex images,helping doctors to make more accurate stage assessments and more accurate treatment plans.It brings a lot of convenience to doctors.Liver cancer is currently one of the most common cancer diseases in the world.Due to the diversity and spread of tumor shape,automatic segmentation of liver tumor lesions is very challenging.This good isolated nature of radiologists makes artificial intelligence algorithms well used in the field of liver tumor segmentation.This thesis analyzes the advantages and disadvantages of the existing liver tumor segmentation algorithm,and improves the algorithm based on two-dimensional fully convolutional neural network(FCN)and three-dimensional fully convolutional neural network under the existing GPU memory resources.The validity of the proposed method is verified on the public data set.Firstly,the basic theoretical knowledge of medical image segmentation algorithm based on deep learning is introduced,which lays a theoretical foundation for the subsequent introduction of improved algorithm work.Secondly,aiming at the problem of insufficient spatial information in the twodimensional FCN neural network method,a new two-dimensional FCN-based liver tumor segmentation algorithm is proposed.The model uses multi-scale feature extraction module and adds position information in the deconvolution process to fully combine the global information and local information of the image to improve the accuracy of lesion location.By comparing with existing methods,the amount of model parameters is greatly reduced,reducing the need for computer GPU memory.Finally,a lightweight three-dimensional FCN-based liver tumor segmentation algorithm with dual-channel input is proposed,which alleviates the high computational cost and high memory consumption of 3D convolution.We compensate for the imbalance of training samples by randomly sampling and optimizing the objective function,and combine all the improvement points proposed in this thesis to get a hybrid network structure.The experimental results on the public dataset show that,regardless of the network structure based on two-dimensional FCN or three-dimensional FCN,the proposed method reduces the parameter quantity of the model while improving the accuracy of model positioning compared with the original method.