A Research Of Chest Radiograph Classification And Lesion Detection Based On Machine Learning

Lung cancer is one of the world's cancers with high morbidity and mortality rates.In 2015,lung cancer accounted for 17.1%(733,000 cases)of the 4.292 million new cancer cases in China,and deaths accounted for 21.1%(610,000 cases)of total cancer deaths(2.814 million cases).The cause of lung cancer is complex,involving genetic and genetic changes.In recent years,with the overall improvement of public health knowl-edge,reduction of environmental pollution,improvement of air pollution,improvement of medical level and use of clinically effective drugs,these have greatly improved the therapeutic effect of lung cancer However,the survival rate and mortality rate of pa-tients with advanced lung cancer are still at a high level.Therefore,early diagnosis and treatment of lung cancer has become the key to the prevention and treatment of lung cancer.An important method for the primary screening of clinical lung cancer is the use of imaging diagnosis.CT means computed tomography,which is a computer-controlled X-ray showing different density images through human tissue for comparative diag-nosis.The misdiagnosis rate of CT is low.However,some scholars believe that CT scans have the risk of patients receiving higher levels of radiation,because X-rays have the potential to make genetic mutations.Compared with CT examination,X-ray have some advantages,eg:small radiation,low cost,and popularization of equipment,es-pecially in the less developed countries of the third world economy.Therefore,X-ray radiographs are used in actual clinical lung cancer screening and related lung diseases.There is still a certain meaning in the work.Image recognition is the use of computer algorithms to process and analyze im-ages for identifying different targets and objects.Image recognition has always been an important direction of computer vision research.It has become a popular applica-tion to collect and control information on intelligent scenes through digital images.The image recognition and positioning technology can effectively process the detection and recognition of specific targets,and can classify and locate the images.At present,image recognition technology has foreseeable broad application prospects in high-tech indus-tries such as image search,target behavior analysis,drones,and autonomous driving,as well as biomedicine and geology.In the early stage of image recognition,some feature extraction methods such as histogram of oriented gradients(HOG)and scale invariant feature transform(SIFT)were used,and then the extracted features were classified by classifier.These characteristics are ultimately based on the subj ective manual features of the research,different applications,different problems,and different choices of de-sign features will have a direct impact on the overall system performance,the early image recognition tasks were all targeted to specific recognition obj ects,and the sam-ple size was smaller than current one.The early image recognition algorithm model had poor generalization ability,and it was difficult to achieve higher recognition re-quirements in practical applications.In recent years,with the development of machine learning technology,more and more successful cases of applying machine learning tech-nology to image recognition tasks have been made.Deep learning belongs to the research field of machine learning,bringing a new wave of research to machine learning.Deep learning in the field of image recognition,speech recognition has made a maj or breakthrough,deep learning solves the gradient diffusion problem.The new method transforms the artificially designed feature extrac-tion into an automatic acquisition feature in image recognition technology.The study of machine learning-assisted chest X-ray diagnosis can be divided into two parts:one is:the classification method of positive chest radiography with lesion and the negative chest radiograph without lesion;the other part is the research on the de-tection and location of lesions in positive chest radiograph.This paper comprehensively analyzes the models that can be used for the above two tasks,using redesigned classifi-cation network,optimization suggestion box generation strategy,adding new functional layer,using effective activation function,feature extraction network replacement,gradi-ent information dynamic correction,second order optimization algorithm and influence function algorithm,three new chest X-ray classification models and three new chest X-ray lesion detection models were proposed.The research results show that the new model proposed in this paper can be better applied to the task of chest disc sorting and chest lesion detection,and expands the application range and depth of computer-aided chest radiography diagnosis.The main research contents and results of the thesis include the following aspects:(1)For the chest radiography classification task,the traditional machine learning algorithm and deep learning algorithm are studied.The application of principal com-ponent analysis in the chest radiography classification task is studied for the problem of excessive image data.Secondly,this paper analyzes the problem that traditional ma-chine learning algorithms are difficult to understand two-dimensional image informa-tion,and proposes a chest radiography classification task based on convolution-based deep learning network.Considering the dramatic change of the number of neurons in the fully connected network,redesign the fully connected layer is proposed,and the FC-migreatd-ResNet model is proposed.The strong classification performance of the model on the chest catalogue database is verified by experiments.The classification accuracy is better than the traditional machine learning model and the original classifi-cation model.(2)For the chest lesion detection task,the Fast-RCNN model was studied.The multi-task network in the model used the Dropout strategy to reduce the possibility of the model fitting the training set,in order to improve the detection performance of the model.Inspired by the Inception network,this paper proposes the VGG-19-BN model,which removes the Dropout strategy and adds the Batch Normalization layer to standardize the output of each layer in the multi-task network,reducing the difference between different samples.The new model was validated in the chest lesion database and improved lesion detection performance compared to the original Fast-RCNN model.(3)For the chest lesion detection task,the Faster-RCNN model is studied.This model uses the convolutional RPN network to generate bounding box for training the network.The 9-anchors bounding box generation strategy in the original algorithm will generate too much low quality bounding box.The low quality bounding box reduces the performance of the model lesion detection.This paper firstly counts the labeling box in the chest lesion database,proposes the 4-anchors bounding box generation strategy,and secondly,compares the ResNet and VGG feature extraction networks.The ResNet-4anchors chest lesion detection model was finally proposed.Through the detection of lesions in the chest lesion database,the experiment result shows that the performance of the new model is significantly improved compared with the original model.(4)For the high threshold lesion detection task,three models of YOLO were stud-ied.The superior model was compared using the chest lesion database and the results showed that the YOLOv3 model had higher detection performance in high-threshold le-sion detection tasks.In order to complete the high threshold lesion detection task better and reduce the missed detection rate of the lesion,through the analysis of the YOLOv3 network structure,it is found that number of some convolutional layer feature maps is reduced too fast,and there is the possibility of losing information,so some new con-volution layers are added which filter kernel size is 1 × 1,the YOLOv3-domain lesion detection model was proposed.The new model was validated on the chest radiography lesion database.The performance of the new model with high threshold lesion detection was improved compared with the original network.(5)For the chest radiography classification task,the gradient descent method for neural network parameter training and the second-order optimization algorithm in nu-merical optimization are studied.In order to train the neural network more effectively and improve the model prediction performance,this paper proposes a second-order op-timization algorithm HFGCSO which uses the gradient information correction strategy.The new algorithm integrates the first-order and second-order information of network parameters,and considers the spatial relationship between the two kinds of informa-tion,which makes the parameter training of the neural network more effective.The HFGCSO algorithm is validated on two data sets.The new algorithm has higher accu-racy in chest radiography classification task than the gradient descent algorithm and the original second-order optimization algorithm.(6)For the chest radiography classification task,the influence function is studied.In order to improve the interpretability of the deep classification network,this paper uses the influence function to quantitatively analyze the chest radiography classifica-tion dataset,and obtains the single training sample influence value which effect the deep classification model.The predictive behavior of the classification model is explained through the influence values of training sample.In order to improve the classification performance of the traditional machine learning model,a tailoring training strategy is proposed.The accuracy of the model classification improved by the tailoring train-ing strategy is improved.The new strategy is effective for the three machine learning classification models on the chest radiography classification dataset.The neural network mimics the human brain neurons and shows the ability to make decisions that are close to or even beyond humans in many tasks.The use of neural network for medical image-assisted diagnosis is the original intention of this paper.The improvement of various classification machine learning algorithms and target detection algorithms is the research content of this paper.In this paper,different methods are used to improve the performance of computer in the diagnosis of chest radiography.The purpose of this paper is to improve the application value of computer-aided diagnosis in medical imaging diagnosis,and to provide technical preparation for accelerating the popularity of computer-aided diagnosis.