| BackgroundLung cancer is one of the malignant tumor with the highest morbidity and mortality from all cancers in China as well as in all other countries.Because this cancer has no obviously clinical symptoms in the early stage,most patients are diagnosed when the tumor develops into rather late stage and surgery becomes impossible.Over the past decade,the treatment of non-small cell lung cancer(NSCLC)has evolved from previous cytotoxic chemotherapy to a phase of individualized targeted therapy based on molecular changes.Targeted therapy is one of the common treatment methods for the advanced NSCLC.Several clinical trials showed that,the epidermal growth factor receptor tyrosine kinase inhibitor(EGFR-TKI)can be administered to NSCLC patients as the first-line medicine,who are sensitive to the epidermal growth factor receptor(EGFR)mutations.This method can significantly improve the patients’ survival rate and quality of life,however it is also limited by the highh price,complicated procedure,reproducibility,and the need of intrusive inspection.Recently,radiomics is able to extract high-throughput parameters(e.g.,texture parameters)in traditional medical image.Based on the artificial intelligence(e.g.,machine learning),it can be used to build prediction model of EGFR mutations.Meanwhile,the diagnosis efficiency,robustness and generality could also be tested and validated.Overall,radiomics provides a new path for the diagnosis and treatment of NSCLC patients.ObjectiveTo explore the predictive value of NSCLC clinical comprehensive predictive model based on imaging features and clinically related risk factors for EGFR gene mutation.MethodsFrom July 2015 to July 2019,we retrospectively collected and screened lung cancer inpatients who received chest CT scanning for lung cancer in the Radiology Department of the Second Hospital of Anhui Medical University.The complete genetic detection results could be obtained from all enrolled patients based on the electronic medical record system(EMR).The preoperative CT scanning data and clinical risk factors(such as age,gender,smoking status,serum carcinoembryonic antigen,etc.)of 151 enrolled patients with pathologically confirmed NSCLC were finally obtained through screening.Firstly,the AI-based radiomics software was used to extract the three-dimensional volume data of all target lesions from the CT images to obtain the high-dimensional radiomics features for the construction of the radiomics model,which was regarded as the radiomics signature.Secondly,three prediction models(clinical model,radiomics model and comprehensive prediction model)were finally constructed through the logistic regression analysis and other modeling methods by using various risk factors.Finally,the prediction efficiency were compared by using the receiver operating characteristic curve(ROC)of all proposed models based-on independent test data sets,and finally selected the optimal one,taking EGFR mutation status as the gold standard.ResultsThe radiomics signature and clinical risk factors such as age,gender,smoking status and glitch signs were associated with EGFR mutation.The constructed radiomics signature could distinguish the EGFR mutation status,the area under the curve(AUC)of 0.864 and 0.831 in the training and validation group,respectively.After combined with the clinical model(AUC= 0.720 and 0.742 in training and validation group respectively),the AUC of comprehensive predictive model were 0.893,0.846 in training and validation group,respectively.ConclusionThe comprehensive prediction model could adopt the high-dimension radiomics features(the first-order features,gray level co-occurrence matrix features,gray level run length matrix and gray level size zone matrix,etc.)and clinical risk factors,and is able to reflect internal relevance between the predictors and the EGFR mutations status.This proposed advanced biomarkers combined with radiomics and clinical risk factors could be expected to support the clinical decision-making preoperatively. |
PDF Full Text Request