The Study Of Dual-energy CT Based Radiomics Model In Predicting Thyroid Cartilage Invasion Of Laryngeal Carcinoma

Objective:To explore and compare the value of radiomics model based on dualenergy CT venous phase iodine maps and 70 keV monochromatic images in predicting thyroid cartilage invasion of laryngeal carcinoma preoperatively.Methods:This retrospective study collected 111 patients with pathology confirmed laryngeal carcinoma who received preoperative dual-energy CT scanning.The patient’s gender,age,primary tumor site,tumor T stage,N stage were collected,and the patients were divided into thyroid cartilage invasion group and non-thyroid cartilage invasion group according to the postoperative pathology.All cases were randomly divided into training set and test set according to the ratio of 7:3,and the region of interest(ROI)of laryngeal carcinoma lesions was manually delineated layer by layer using ITK-SNAP software.FeAture Explorer(FAE)software was used to extract 851 radiomics features from the volume of interest(VOI)of laryngeal carcinoma lesions in the preoperative dual-energy CT venous phase iodine maps and 70 keV monoenergetic images of each patient.For the radiomics features extracted based on the iodine maps in the venous phase,Pearson Correlation Coefficient(PCC)and Principal Component Analysis(PCA)were used for feature dimension reduction,and further analysis of variance(ANOVA),Kruskal-Wallis test,Recursive Feature Elimination(RFE)and Relief were used for feature selection.Different classification models were constructed based on the finally screened radiomics features and 10 classifiers.For the radiomics features extracted from 70 keV monoenergetic images in the venous phase,PCC was used for feature dimension reduction,ANOVA,RFE,and Relief algorithms were used for feature selection,and different classification models were constructed by combining 10 classifiers.Ten-fold cross-validation method,receiver operating characteristic curve(ROC)and area under curve(AUC)were used to evaluate the performance of different classification models in predicting thyroid cartilage invasion of laryngeal carcinoma.The delong test was used to compare the area under curve of iodine maps radiomics model and 70 keV monoenergetic images radiomics model.Results:1.Among the 111 laryngeal carcinomas,108 were male and 3 were female,ranging in age from 36 to 83 years old,with an average age of 59.5±8.7 years old.Among them,42 cases had thyroid cartilage invasion and 69 cases had no thyroid cartilage invasion.There was no significant difference in age,gender,primary site of tumor and N stage between the thyroid cartilage invasion group and the non-thyroid cartilage invasion group(P>0.05),but there was significant difference in T stage(P<0.05).2.The AUC of all classification models was compared using FAE software in the validation set.The results were show below:(1)Using iodine maps of venous phase,The Random Forest(RF)classification model(RF model)constructed by 16 radiomics features retained after dimension reduction based on PCC and feature selection based on RFE had the best prediction performance,The AUC in the training set,and test set were 1.000,and 0.883,respectively,and the accuracy,sensitivity,and specificity in the test set were 0.853,0.923,and 0.809,respectively.When using the "one-standard error" rule,the model produced a simpler predictive model based on 10 radiomics features.The best algorithm RF through PCC,PCA two feature dimension reduction methods and ANOVA,Kruskal-Wallis test,RFE,Relief four feature selection methods in the training set and test set AUC were greater than 0.750.(2)Using 70 keV monoenergetic images of venous phase,PCC was used for feature dimension reduction,and based on three feature selection algorithms,ANOVA,RFE,and Relief,three models with the best diagnostic performance were finally selected.Among them,the machine learning model(GP model)based on seven radiomics features(wavelet-HHL_glszm_Large Area High Gray Level Emphasis,wavelet-HLH firstorder_Skewness,wavelet-LHH_glszm_ZoneEntropy,waveletLHL firstorder_Median,wavelet-LHL_glszm_ZoneEntropy,wavelet-LLH_firstorder_Skewness,wavelet-LLL_glszm_ZoneEntropy),RFE for feature selection and using Gaussian Process(GP)algorithm as classifier showed the best predictive performance in predicting thyroid cartilage invasion of laryngeal carcinoma,The AUC of the training set and the test set were 0.999 and 0.817.The seven radiomics features finally screened were all wavelet features,and the top three features with weight coefficients were Large Area High Gray Level Emphasis(LAHGLE),Skewness,and ZoneEntropy(ZE).3.The results of Delong test showed that:the AUC of the RF model was greater than that of the GP model in the training set,test set.But there was no statistical difference(training set:Z=0.707,P=0.479,test set:Z=0.639,P=0.522).The AUC of the two models in the training set,test set were all greater than 0.800.Conclusion:The classification model constructed based on the radiomics features of dual-energy CT venous phase iodine maps and 70 keV monoenergetic images can effectively predict thyroid cartilage invasion of laryngeal carcinoma,providing a new non-invasive method for preoperative prediction of thyroid cartilage invasion in laryngeal carcinoma.