Research And Application Of End-to-end Machine Learning Models In Diagnosis Of Lung Nodules

Lung cancer is one of the most serious malignant tumors of human health,and its early manifestations are solitary pulmonary nodules.Computed tomography(CT)is a commonly used imaging method for the detection of pulmonary nodules.With the rapid development of computer technology and the sharp increase of medical image data,computer aided diagnosis(CAD)system is widely used.The clinical application of CAD system relieves doctors' workload and improves work efficiency.Therefore,it is of practical significance to study the effective algorithm for diagnosis of benign and malignant pulmonary nodules.The traditional CAD system based on lung CT images includes the stages of lung parenchyma segmentation,feature extraction and classifier design.And its problems are as follows: complicated process,many artificial intervention.To solve these problems,the end-to-end machine learning models are applied.At present,the commonly used end-to-end machine learning models include massive training artificial neural network and convolutional neural network.The core processing unit of the massive training artificial neural network is the fully connected network,which has several problems,such as complex structure,excessive parameters and time-consuming training.Therefore,this thesis studied the application of convolution neural network in the classification of benign and malignant pulmonary nodules.First,designed convolutional neural network model.The first step was to select some structural parameters,such as the network depth,the number and thesize of convolution kernel.The second step was to determine the training parameters,such as activation function type,basic learning rate and the learning rate attenuation strategy.Through a large number of comparative experiments,the model structure was determined and several parameters were optimized.Finally,the model was applied to the classification of benign and malignant pulmonary nodules.Secondly,in view of the small data set,the convolutional neural network model was prone to over-fitting,then a data enhancement method was proposed to extend the training samples to prevent over-fitting.The specific operation was to divide the region of interest into a number of local sub-regions,then the sub-regions containing nodule related features were selected,and added them to the training samples.So that,the convolutional neural network model was trained with a large sample set.The experimental results showed that the convolutional neural network model designed in this thesis had a good performance in the classification of benign and malignant pulmonary nodules.The accuracy and AUC values of the evaluation indexes were 89.5% and 0.896 respectively.Secondly,in order to verify the validity of the data enhancement method proposed in this paper,the extended samples were used to train AlexNet,rd-CNN and the convolutional neural network model designed in this paper.The test results showed that the performance of three kinds of network models were improved.The accuracy and AUC value of the self-designed convolution neural network model were increased to 92.5% and 0.933 respectively.Therefore,the research in this paper has certain reference value for improving the performance of computer aided diagnosis system.