CT-based Radiomics Features For The Prediction Of Mediastinal Lymph Node Metastasis

Objective: The purpose of this research was to explore CT radiomics features associated with Mediastinal lymph node metastasis(MLNM)in NSCLC,then build and assess the predictive power of CT-based Radiomics models in the identification of MLNM in NSCLC.Furthermore,we find the radiomics features to discriminate normal-sized MLNM between normal-sized benign mediastinal lymph nodes(MLN)and demonstrated their diagnosis efficiency.Finally,we aim to explore different radiomics features between MLNM from adenocarcinoma(AD)and MLNM from squamous cell carcinoma(SCC).Method:1.CT-based Radiomics models for the prediction of MLNM in NSCLCA total of 169 MLNs with pathological diagnosis were retrospectively collected from January 2019 to December 2021 in 134 patients diagnosed with NSCLC at Shaoxing People’s Hospital,including 80 metastasis MLNs and 89 benign(chronic lymphadenitis)MLNs.The MLNs were randomly split into two groups: training and testing set,with a ratio of 7∶3.ITK-SNAP was utilized to delineate ROIs in plain and venous phase CT images of MLNs,and Analysis-Kinetics was employed to extract 396 radiomic features.Standardization,Max-Relevance and Min-Redundancy(m RMR)and least absolute shrinkage and selection operator(Lasso)algorithm,logistic regression were employed to select feature in the training set,while logistic regression was employed to construct a prediction model.The receiver operating characteristic(ROC)curve was employed to estimate the diagnostic efficiency of the models,calibration curve was utilized to assess the calibration degree,clinical application was then evaluated though the decision curve analysis(DCA).2.CT-based Radiomic features to differentiate Normal-sized MLNM in NSCLCA total of 107 normal-sized MLNs in 87 NSCLC patients with complete pathological and clinical data from January 2019 to December 2021 in Shaoxing People’s Hospital were collected retrospectively,including 31 metastasis MLNs and 76 benign MLNs with histopathology diagnosis as gold standard.CT radiomic features were extracted from the region of interest(ROI)delineated on plain-phase and venous-phase CT imaging of MLNs.Mann-Whitney U test/T-test,Spearman correlation and then LASSO algorithm was used to select features,Receiver operating characteristic(ROC)curves for each feature were constructed,the area under the curve(AUC),sensitivity and specificity were calculated to assess the diagnostic efficacy of radiomics features.3.Different CT Radiomic features between MLNM of AD and MLNM of SCCA total of 76 MLNMs in 59 NSCLC patients from January 2019 to December 2021 in Shaoxing People’s Hospital were collected retrospectively,including 48 adenocarcinoma MLNMs and 28 squamous cell carcinoma MLNMs with histopathology diagnosis as gold standard.CT radiomic features were extracted from the region of interest(ROI)delineated on plain-phase and venous-phase CT imaging of MLNMs.Mann-Whitney U test/T-test,Spearman correlation and then LASSO algorithm was used to select features,and then analysis the differences of features between MLNM of AD and MLNM of SCC.Result:1.After the dimension reduction by m RMR and Lasso,the logistic regression model with 8 features in plain-phase showed satisfactory diagnostic performance for predicting MLNM of NSCLC.The AUC,sensitivity,specificity were 0.9/0.84,0.97/0.92 and 0.72/0.66 in the training/test sets,respectively.In venous phase,6 features were selected to build the radiomics model,The AUC,sensitivity,specificity were 0.88/0.75,0.81/0.74 and 0.77/0.71 in the training/test sets,respectively.The classifications and observations were in good accord,as evidenced by calibration curves.DCA confirmed the clinical benefits of radiomics models.2.5 features in plain-phase CT and 4 features in venous-phase CT were showed significant association with MLNM of NSCLC with the areas under the ROC curves(AUCs)of 0.623-0.732 and 0.634-0.696.The association of radiomics features demonstrated AUC value of MLNM predicting performance was 0.852 and 0.739 respectively in plain and venous phase with sensitively of 0.806,0.645 and specificity of 0.803,0.803.3.After a serious of selection,we found Inverse Difference Moment_All Direction_offset4_SD,Voxel Value Sum,Short Run Low Grey Level Emphasis_angle90_offset1,Low Intensity Emphasis,Grey Level Nonuniformity_All Direction_offset7_SD were significant different between MLNM of AD and MLNM of SCC in plain-phase.Voxel Value Sum,Long Run High Grey Level Emphasis_angle0_offset1,Short Run High Grey Level Emphasis_angle0_offset4 were significant different between two groups in venous-phase.Conclusion:(1)We found CT radiomics features associated with MLNM in NSCLC,and models based on CT radiomics features are useful in the prediction of MLNM in NSCLC,which can assist in clinical staging.(2)CT radiomics features has the potential to accurately differentiate normal-sized MLNM and normal-sized benign MLN from NSCLC,they can help to diagnosis MLNM for early NSCLC.(3)we found different CT radiomics features between MLNM from AD and MLNM from SCC,which may help to reflect internal structural of MLMN of different pathological types.