A Study For Predicting Cervical Lymph Node Metastasis In Patients With Papillary Thyroid Carcinoma Based On CT Radiomics

Objective: To study the value of radiomic features extracted from pre-contrast phase,arterial phase and venous phase for predicting cervical lymph node metastasis(CLNM)of papillary thyroid carcinoma(PTC)and furthermore analyze predictive performance of combined phase to CLNM in contrast-enhanced CT scan by using radiomics analysis,aiding to explore the necessity of multi-phase CT scan.Methods: Clinical data and CT images of 1213 patients underwent total thyroidectomy or lobectomy were collected in our hospital from January 2017 to June2020.197 patients with PTC were screened by pathological result and exclusive or inclusive criteria and 512 CT images with the thickness of 1mm including pre-contrast phase,arterial phase and venous phase in 197 patients with PTC were gathered for radiomic study.The 512 CT images consisted of 193 pre-contrast phases,131 arterial phases and 188 venous phases,which only 124 patients contained all of three phase CT images.Regions of interest(ROIs)was segmented manually on the slice with the largest area of PTC lesion in 512 CT images by ITK-SNAP software and 107 radiomic features were extracted from each ROI by using Pyradiomics software.Grid Search CV method and 5-fold cross validation were performed with all 107 radiomic features and512 CT images,which were divided into training set and test set at a ratio of 4:1,to investigate and select optimal classified parameter by CLNM status.The result showed that optimal classified parameter was following:{‘max＿depth’: 60,‘max＿features’ :‘sqrt’,‘min＿samples＿leaf’: 8,‘n＿estimators’ : 20}.Firstly,124 patients who had all three phase CT images were randomly divided into training set and test set at a ratio of9:1.The random forest(RF)classification models of three phases,which were precontrast phase(n),arterial phase(a)and venous phase(v),respectively,were established based on best features set [top(k)] selected by Select KBest algorithm and the optimal classified parameter optimized by Grid Search CV method and tested by 10 fold cross validation to achieve top(k)and 10-fold receiver operating characteristic(ROC)curve.Secondly,CT images of pre-contrast phase,arterial phase and venous phase in 124 patients were reconstructed four combined phases,which were indicated as the following:n+a,n+v,a+v and n+a+v,and each combination was divided into training set and test set at a ratio of 9:1.Similarly,the RF classification model of four combined phases were established by CLNM status and tested with 10-fold cross validation to achieve top(k)and 10-fold ROC curve.The maximal average area under curve(AUC)of and accuracy were used for assessing the model.R-language software was used for statistical analysis.t-test and Chi-square test was performed for continuous variables and categories variables,respectively.P < 0.05 indicates significant difference statistically.Results: The ROC of three RF models built by best radiomic features extracted from pre-contrast images,post-contrast arterial and venous phase images of CT scan for predicting CLNM correctly showed maximal average AUCs: 0.843、0.775、0.783 and corresponding accuracy: 0.767、0.695 and 0.726,respectively.The ROC of four RF model created by best radiomic features extracted from combined phases of n+a,n+v,a+v and n+a+v for predicting CLNM showed maximal average AUCs: 0.829、0.829、0.777、0.821 and corresponding accuracy: 0.759、0.765、0.724 and 0.733,respectively.Compared to arterial phase and venous phase,the radiomic feature sets extracted from CT images of pre-contrast phase showed higher AUC value to predict CLNM(p<0.05).There were no significance statistically in four AUCs from precontrast phase and three combined phase,which indicated n+a,n+v,and n+a+v,respectively,while the combined radiomic features extracted from CT images of a+v combined phase showed lower performance to predict CLNM(p<0.05).Conclusions: 1.Radiomic features extracted from pre-contrast phase,arterial phase,venous phase and four combined phases including n+a,n+v,a+v,and n+a+v in CT scan could correctly predict CLNM in patients with PTC.2.Compared to those from CT images of arterial phase and venous phase,radiomic features extracted from CT images of pre-contrast phase show better performance to predict CLNM.3.The combined phases of enhanced-contrasted CT cannot improve the ability of the prediction for cervical lymph node metastasis in patients with papillary thyroid carcinoma.