Identification Of EGFR Mutations In Lung Adenocarcinoma With CT Images

The epidermal growth factor receptor(EGFR)is a transmembrane receptor tyrosine kinase,which is important for the growth of tumor cells.The mutation rate of EGFR in Asian patients with lung cancer is about 50%.Accurate indentification of EGFR in lung adenocarcinoma is a prerequisite to make clinical decision.The existing clinical methods(biopsy test)are high-risk,invasive,and difficult to implement repeatedly.Therefore,to achieve dynamic tracking of EGFR mutation in lung adenocarcinoma during the treatment,non-invasive and convenient models using radiomics and convolutional neural networks based on CT images were proposed.This study firstly extracted quantitative features by radiomics method from chest CT images of lung adenocarcinoma patients.Then,a large number of discriminative radiomic features were selected to construct a classification model to indentify the EGFR mutations in lung adenocarcinoma.The numbers of radiomic features which have significant(P-value<0.05)and very significant(P-value<0.001)abilities to distinguish EGFR mutation statues were 186 and 82,respectively.The AUC score achieved by radiomics based model in validation dataset was 0.766,which was significantly better than clinical features baseline(AUC=0.686).To solve the instability during the radiomic feature extraction process,two dimensional convolutional neural network(2D CNN),which was an end-to-end model,was proposed to incorporate the feature extraction and selection steps into the classification step.A high performance(AUC=0.838)was achieved by the fine-tuned 2D CNN model which was built using the mutli-scale input sizes,multi-view slicing method,and multiple tests.This model was significantly better than radiomics based model.To utilize the spatial information of tumor CT images,a 3D CNN model was further proposed.First,this strudy built a pre-trained model by a large3 D medical image dataset of lung nodule classification.Based on this pretrained model,the optimal 3D CNN model was fine-tuned by the EGFR genes data and the corresponding AUC in validation dataset was 0.863.A CT image based fusion model was obtained by fusing radiomics and2D/3D CNN models.Then,the identification ability of the CT image based fusion model was further improved by incorporating the clinical features(gender and smoking history).The fusing model achieved the highest AUC score of 0.881,which means that there are complementarity between the information extracted from CT images and clinical features.Comparied with other state-of-the-art medical images based models,the proposed model was more accurate and closer to the clinical requirements.The results of this thesis indicate that the CT image can identify the mutation status of EGFR gene in lung adenocarcinoma and assist clinical decisionmaking,which provides a non-invasive,convenient,dynamic tracking and efficient detection method for clinical.