Automatic Classification Of Healthy And Sick Lung CT Images With 3D Convolution Neural Network

With the development of science and technology,artificial intelligence has permeated many aspects of human's life.At the same time,doctors increasingly rely on medical image data for the diagnosis of disease.The intelligence of medical image is already an inevitable development trend.This paper does the research on how to screening healthy CT image based on 3D convolution neural network in order to achieve automatic classification for CT lung image which is helpful for reducing the burden on doctors and then lowering the rate of misdiagnosis and missed diagnosis.In the beginning of this paper,the current research of deep learning in medical image is described and the existing problems are presented.After comparing the methods of computer aided diagnosis between traditional machine learning and deep learning,the latter is superior to the former.Then aiming at the limitations of 2D convolution neural network in CT sequence image classification task,3D convolution neural network is proposed to be applied to healthy screening for CT lung images Finally,the validity of 3D convolution neural network in sequence image classification task is proved by experiments.The main contents of this paper are as follows:First,in order to deal with the CT lung data set,the DICOM standard which provides the rule for transmission and storage of medical image is analyzed.Then the relative information contained in the CT images which being to converted into JPG format image by linear window display technology is extracted.After that,the concept of window center and window width and their effects on image brightness and contrast are introduced.Second,the basic structure of convolution neural network which includes the principle of convolution layer and pool layer,the propagation process of the whole network,including forward propagation and error back propagation,and gradient descent algorithm was presented.Some methods for enhancing training effect and convergence speed are also introduced.The experiments in this paper rely on a deep learning framework named Lasagne.Finally,the classification of 2D network for healthy and illness in CT images is carried out and shows the limitations in this task.Third,based on the 2D convolution neural network,3D convolution neural network is improved in order to make better use of the information between adjacent layers of CT sequence image.The differences between the two methods shows advantages of the 3D convolution kernel.Then,the theory of migration learning is applied to this experiment.In addition,the pseudo-color images are synthesized by three window images including the lung window image,the high attenuation window image and the low attenuation window image.Parameters of the network in medical image classification is initialized with the pretrained parameters in natural image classification.Finally experiments show that applying 3D convolution network in the healthy screening task could produce better results.The research of this paper shows that 3D convolution neural network can extract more correlative information between adjacent CT sequences than 2D convolution neural network and perform better in classifying series of images.In addition,the situation of lack of medical data make migration learning to improve the accuracy of classification in this task possible.