The Classification Of Lung Adenocarcinoma And Lung Squamous Carcinoma Using Machine Learning

Lung cancer has the highest morbidity and mortality among malignant tumors in the world,and it is also the most common malignant tumor in China.Different histological subtypes of lung cancer have different biological behaviors,the degree of malignancy,the tendeny of metastasis and the sensitivity of different treatment require the histological subtype diagnosis of lung cancer.There are many clinical sampling methods to determine histological classification,but they are all invasive.Computed Tomography(CT)imaging is an effective method for the disease diagnosis.Based on CT images,this study applied several methods in machine learning to explore the differentiation between lung squamous cell carcinoma(Sq CC)and adenocarcinoma(ADC).The radiomics mrthod was used to extract features form CT image lesions,including intensity features,shape features,and texture features.The synthetic minority oversampling technique(SMOTE)method was used for data expansion to eliminate the imbalance between different subtypes.The analysis of variance(ANOVA)and the least absolute shrinkage and selection operator(LASSO)methods were used to select features,which significant for distinguishing Sq CC from ADC.Using support vector machine(SVM),logistics regression(LR)and multilayer perceptron(MLP)methods to construct the machine learning models,and grid search and 5-fold cross-validation were used for tune model parameters.The receive operating characteristic curve(ROC),sensitivity,and specificity were performed in test set to evaluate these models.The results showed that 71 features have good effects,among the machine learning models,MLP and LR showed good performance.SVM model performance needs to be improved.Deep learning was then used to classify the two subtypes of lung cancer end-to-end.The transfer learning method was used due to insufficient data.Using residual networks(Res Net)with group normalization to construct Res Net GN network,and the source network which used to identify benign and malignant of lung nodules.Divided the data into training and test sets,then rotated and mirrored the training data.After preprocessed of training set,cut out different size of the region of interest(ROI)to train different models which were pre-trained Res Net GN network.After optimizing the network parameters,the 2D model had a more effective performance than 3D model in 193 cases of data,and transfer learning method greatly improved the performance of the model.In this study,three traditional machine learning and deep learning methods with different input sizes were compared to deeply mine image information.By extracting the image features of lung cancer,a discriminative model of lung adenocarcinoma and squamous cell carcinoma based on CT images was constructed.This study had proved that machine learning methods have powerful capability to release image information.After multi-center verification,machine learning methods may be applied to clinical practice to assisting physicians.This paper consists of 30 figures,4 tables,15 formulas,and 51 references.