A Combined Model Of CT-based Radiomics Features And Clinical Data To Predict The Risk Of Brain Metastases In Non-small Cell Lung Cancer

Background: Lung cancer is the most common tumor with the highest mortality in the world.Brain metastasis is one of the most common sites of distant metastasis of lung cancer.Patients with non-small cell lung cancer(NSCLC)have a high incidence of brain metastasis.Enhanced MRI screening is the best diagnostic method for brain metastasis.Patients need to follow up with enhanced MRI screening multiple times,which increases the risk of contrast agent deposition and economic burden.Some lung cancer patients with brain metastasis do not have related neurological symptoms.Whether head MRI screening is necessary needs further exploration.Patients with brain metastasis of NSCLC choose the treatment plan according to the number of metastases.The fewer metastases,the less damage the treatment does to normal area of the brain.Therefore,it is necessary to stratified the risk of brain metastasis in NSCLC patients.For patients with high risk of brain metastasis,it is recommended to actively and closely review and develop a comprehensive treatment program which is beneficial and personalized,so as to diagnose and treat earlier,improve the quality of life and prolong the overall survival of patients.For patients with low risk of brain metastasis,the interval of enhanced head MRI screening can be properly prolonged to reduce the financial burden of patients and the risk of MRI contrast agent deposition.CT radiomics can obtain high-throughput and high-dimensional image information non-invasively.It has been reported that CT radiomics features may be associated with brain metastasis in NSCLC.However,it has not been determined which thickness,phase and reconstruction algorithm CT images should be selected for CT radiomics research.Objective: This study aims to analyze CT images with different thickness,phase and reconstruction algorithm to find stable CT radiomics features with predictive value and establish model to predict brain metastasis of NSCLC.Combined with high risk clinical indicators of brain metastasis,a combined model of CT radiomics,clinical indicators and CT imaging features will be established to predict the risk of brain metastasis more comprehensively and efficiently.Methods: Chest CT images and clinical data of 226 patients with pathologically confirmed NSCLC were retrospectively analyzed,of which 61 had brain metastases and 185 did not.The training group(195 cases)used SIEMENS CT scanner and the test group(31 cases)used GE CT scanner.Univariate analysis and multivariate analysis were used to screen clinical indicators and CT imaging features related to brain metastasis.Different observers used semi-automatic delineation method to delineate areas of interest twice in 1mm nonenhanced lung algorithm images(1mm lung algorithm group),2mm nonenhanced standard algorithm images(2mm nonenhanced group)and 2mm enhanced standard algorithm images(2mm enhanced group)respectively.Radiomics features including features of the original image and filters were extracted.Features with good consistency between observers could be retained and used to evaluate the feature consistency between different reconstruction algorithms and phase groups.Three normalization methods,Pearson correlation coefficient,four feature screening methods and four classifier methods were selected for combination.The radiomics model was constructed on the data set of each imaging training groups,and five-fold cross-validation was carried out to compare the effect of the model combination of various feature screening and classifier methods on the crossvalidation group,and the model prediction performance was tested on the test group.The receiver operating characteristic(ROC)curve was used to analyze and evaluate the efficacy of each group and calculate the area under ROC curve(AUC).The Delong test was used to evaluate the differences of AUC in each group.The clinical indicators,CT imaging features and Rad-score related to brain metastasis were included to construct the combined model,and the effectiveness of the combined models was evaluated by ROC curve analysis.The optimal combined model was selected to establish nomogram of the risk of brain metastasis in NSCLC.Results: Univariate analysis showed that there were statistically significant differences in CEA,NSE,CYFRA21-1,pathological type,N stage,surgical resection,extracranial metastasis,tumor location,tumor margin,pleural effusion,cancerous lymphangitis and intrapulmonary metastasis between brain metastasis positive and negative groups.Multivariate analysis showed that CYFRA21-1,pathological type(adenocarcinoma),extracranial metastasis,pleural effusion and cancerous lymphangitis on CT images were independent risk factors for brain metastasis of NSCLC while surgical resection was independent protective factor.Inter-group consistency analysis of CT image omics features showed that 3.8% of features were consistent between different reconstruction algorithm groups,and 33.0% of features were consistent between different phase groups.Minmax-PCC-RFE-SVM method was used to establish the model in the 1mm lung algorithm group and 2mm nonenhanced group,and Minmax-PCC-Relief-LDA method was used to establish the model in the 2mm enhanced group.The AUCs of the training group were 0.651,0.683 and 0.677.The AUC values of the test groups were0.619,0.696 and 0.643,respectively,and Delong test showed no statistical difference in AUC values of each group.The 2mm nonenhanced group with the highest AUC in the training group and the test group was used to calculate the Rad-score,and the combined model was constructed.The AUC of the combined model of radiomics +clinical indicators in the training group and the test group were 0.838 and 0.738.The AUC of radiomics +CT imaging features combined model training group and test group were 0.703 and 0.714.Delong test showed that the AUC values of radiomics + clinical indicators combined model training group were statistically different from those of radiomics +CT imaging features combined model and 2mm nonenhanced group.Based on this,the nomogram of the risk of brain metastasis in NSCLC was established.Conclusion: The reconstruction algorithm and enhanced scanning would affect the reproducibility of CT radiomics features.It was feasible for CT radiomics model to predict brain metastasis brain metastasis of NSCLC.The CT radiomics model of the2 mm nonenhanced group had the best predictive efficacy.The thickness,phase and reconstruction algorithm of CT images had no effect on the prediction efficiency.The combined model of clinical indicators and 2mm nonenhanced group CT radiomics features could improve the predictive efficiency.