Research On Lung Tumor Image Recognition Method Based On Transfer Learning And Convolutional Neural Network

With the improvement of social economy,lung cancer has slowly become one of the main reasons endangering human safety.Quickly and accurately determine whether the lung has a tumor and whether the tumor is benign or malignant.It can reduce the previous cumbersome examination,prevent unnecessary surgery,reduce the psychological and physical pain of the patient,and can also improve the survival rate of the patient.Therefore,the early diagnosis and treatment of lung cancer have far-reaching significance for the safety of human life.Electronic computed tomography(CT)is used as a screening tool for the diagnosis of lung cancer,but doctors have difficulties in reading CT information,which takes a long time,is prone to fatigue,and causes misdiagnosis.A computer-aided diagnosis system based on CT images of the lungs can assist doctors reduce the amount of reading tasks and help to enhance the accuracy and efficiency of diagnosis.This article focuses on the key technologies of computer-aided diagnosis based on CT images of the lungs.Because it is difficult to obtain massive lung annotation samples and traditional shallow convolutional neural networks are difficult to obtain deep image features and are prone to overfitting,which leads to low classification efficiency and accuracy,we propose an optimization scheme.The main work includes the following:(1)Lung CT data preprocessing.The original image is fuzzy and contains redundant information,which is not conducive to the targeted training of the network,so it is necessary to perform preprocessing operations on the image data.First,the original lung CT image itself is enhanced to expand the gray range of the image and enhance the contrast,make the unclear CT image clear,improve the quality and recognizability of the image,and then use the maximum between-class variance method to segment the CT image,Remove the redundant information in the image,initially realize the segmentation of the CT image,then use morphological methods such as dilation,erosion,open operation and close operation to realize the expansion of the lung parenchyma,and use the regional growth algorithm to realize the sense based on the seed pointExtraction of regions of interest,and finally,the synthetic minority upsampling of the data set to reduce the gap in the sample size of the different types of data and increase the overall sample size,to achieve data imbalance and enhancement processing,and finally obtain data that can improve network performance set.(2)Lung tumor image recognition classification based on DLB-Alex Net.Aiming at the problems of low classification accuracy and overfitting of shallow convolutional neural networks,a deep learning recognition model for lung tumor images is designed.Improve and optimize the Alex Net model,improve the original activation function and the normalization layer,improve the Rectified Linear Unit function to the Leaky-Rectified Linear Unit function,the Local response Normalization is improved to Batch Normalization,and the final output is classified by the Softmax Algorithm for lung image classification.The experimental results show that the improved method improves network convergence speed and accuracy,and improves network robustness.(3)Lung tumor image recognition classification based on transfer learning and DLB-Alex Net.For the problem of how to make the deep neural network effectively use the very limited medical data,the convolutional neural network model pre-trained from the large-scale natural image classification is used as the initial feature extractor to extract the features of the lung tumor image.By organizing the optimal migration matrix to minimize the difference between the multi-core maximum mean of the source and target domain features,the transferred source domain features and corresponding knowledge are used for classification operations,which are then used to classify the target domain features and repeat operations to obtain Annotate information in the target domain,and finally realize the classification and recognition of lung tumor images.Through experimental verification,the proposed method avoids overfitting of network microdata and improves classification efficiency and accuracy..