It is an application and research of artificial intelligence technology in medical field to assist medical image information processing.The most important task of voxel based brain tumor image segmentation task is to achieve a high accuracy classification of brain tumor tissue region pixels.However,because of the randomness of the brain tumors in the brain,the traditional segmentation methods have some limitations,and the accuracy can’t meet the requirements of reality.In order to solve the above problems,we make full use of deep learning powerful nonlinear characterization method,so as to understand the image of deeper information;This thesis uses the model of convolutional neural network to handle the brain tumor image classification and annotation tasks;In the training process of the model,we make improvement on the parameter optimization process.The main contents of this thesis are as follows:This thesis introduces the principle and operation method of convolution neural network in the field of image processing.Then based on two-way convolutional neural network architecture,according to local channel structure,12 layers of convolutional network structure is set up by using certain skills;when considering the dependence between pixel tags,by simulating the CRFs model,this thesis designs the input series connection structure to solve the problem;In order to solve the problem of limited input pixel area,this thesis designed the input multi pool model;Finally,the accuracy can reach 83%,which proved the validity of the improved model.The principle of different stochastic gradient descent method is introduced;And then,an improved Adam algorithm is proposed,and it is used to adaptively adjust the learning rate.The improved Adam algorithm is applied to training process of convolution neural network structure about multi pooling and Input series connection structure.Finally,the improved Adam algorithm and other random the gradient descent algorithm is compared,show that the improved Adam has higher efficiency in the process of training,has important significance on the scale of the training data. |