The Application Study Of Dual-mode Ultrasound Radiomics In The Prediction Of Cervical Lymph Node Tuberculosis

Objective:Based on B-mode ultrasound(B-US)and contrast-enhanced ultrasound(CEUS)images,the radiomics features of patients with swollen lymph nodes in the neck are extracted.The ultrasound imaging features and radiomics features are combined to construct a clinical and radiomics combined model to predict cervical lymph node tuberculosis(CLNT).Nomogram is built for visualization and clinical decision curves are used to quantify net benefits within different threshold ranges in order to evaluate the clinical benefits of different models.Methods:After strict inclusion and exclusion criteria screening,this study included 214 patients with swollen lymph nodes in the neck confirmed by gold standard from January2018 to September 2022 at the Hangzhou Red Cross Hospital.According to the gold standard,the included cases were divided into tuberculosis group and non-tuberculosis group.All enrolled patients underwent B-mode ultrasound(B-US)and contrastenhanced ultrasound(CEUS)in our hospital and the clinical data was complete and could be exported in DICOM format.All the patients included in the study were randomly divided into a training set(n=150)and a validation set(n=64)at a ratio of7:3.1.A retrospective analysis of the relationship between ultrasound imaging features and CLNT was conducted.Independent predictive factors related to CLNT were identified through univariate and multivariate analyses to construct a clinical model(Model 1)and calculate its diagnostic efficiency.2.Two sonographers manually segmented the regions of interest and extracted the radiomics features from both B-US and CEUS modalities.After normalization and using univariate analysis(ANOVA),least absolute shrinkage and selection operator(LASSO)and five-fold cross-validation to select the best features,B-US radiomics model(Model 2),CEUS radiomics model(Model 3),B-US and CEUS combined model(Model 4)were established.Receiver operating characteristic curves(ROC)were plotted for different models and the area under the curve(AUC),sensitivity,specificity and accuracy were compared.The AUC comparison between different models was performed by using Delong’s test.3.After selecting the optimal model among the three radiomics models,ultrasound-independent predictive factors were added to construct a clinical and radiomics combined model(Model 5).Delong’s test was used to compare the AUC and diagnostic efficacy of the clinical prediction model,radiomics model and clinical and radiomics combined model.Nomogram was used to visualize Model 5.HosmerLemeshow test and the calibration curve were used to evaluate the consistency between the predicted results and actual results of Model 5.Decision curve analysis was used to evaluate the net benefits of the three models in clinical application.Results:1.This study included 214 patients,101 males and 113 females,with an average age of 45.1±18.4 years.Among them,there were 117 cases of neck lymph node tuberculosis and 97 cases of non-tuberculosis(27 cases of lymphoma,31 cases of reactive hyperplasia and 39 cases of metastatic lymph nodes).Randomly allocated at a ratio of 7:3,there were 150 cases in the training set and 64 cases in the validation set.There was no significant statistical difference in age and gender between the training set and the validation set(P>0.05).2.Univariate analysis of ultrasound imaging features showed that the boundaries,lymphatic gates,ring hypoecho,peripheral soft tissue echo enhancement,enhancement patterns,perfusion defects,ring enhancement and enhancement modes of lymph nodes in the training set were statistically significant(P<0.05).After multivariate analysis,ring hypoecho and perfusion defects(P<0.001)were taken as independent risk factors for predicting CLNT.A clinical prediction model(Model1)was constructed.The AUC of the model in the training set and validation set were 0.821 and 0.754 respectively.The sensitivity,specificity and accuracy of the model in the validation set were 59.4%,81.3% and 57.9% respectively.3.After univariate analysis of variance,LASSO dimensionality reduction and 5-fold cross-validation,Model 2 ultimately obtained 2 radiomics features.The AUC in the training and validation sets were 0.769 and 0.817.In the validation set,the sensitivity,specificity and accuracy were 90.6%,62.5% and 88.5%.Model 3 selected three radiomics features with the AUC of 0.850 and 0.810 in the training and validation sets.In the validation set,the sensitivity,specificity and accuracy were 93.8%,68.8%and 81.3% respectively.Model 4 ultimately obtained five features with the AUC of0.880 and 0.853 in the training and validation sets.In the validation set,the sensitivity,specificity and accuracy were 78.1%,81.3% and 79.7% respectively.Delong test showed no significant statistical differences in the AUC among the three models in the validation set(P > 0.05).However,the AUC and specificity of Model 4 were the highest among the three models,indicating a balanced sensitivity and specificity.4.The optimal radiomics model,the B-US and CEUS combined model(Model 4),was combined with ultrasound features to construct a clinical and radiomics combined model(Model 5),with AUCs of 0.928 and 0.893 for the training and validation sets.In the validation set,the sensitivity,specificity and accuracy were 87.5%,81.3% and 84.4%respectively.In the validation set of the Model 1,Model 4 and Model 5,although the specificity of the three models was equal,Model 5 had the highest AUC,sensitivity and accuracy.Delong test showed that in the validation set,the AUC of Model 5 was higher than that of Model 1(P<0.05),the AUC of Model 4 was higher than that of Model 1and the AUC of Model 5 was higher than that of Model 4,but the difference was not statistically significant(P>0.05).5.Nomogram could more intuitively present the diagnostic efficacy of Model 5 in predicting CLNT.The calibration curve had a high degree of overlap with the standard curve and Hosmer-Lemeshow test showed P≥0.05,indicating a good fitting effect.Clinical decision curve analysis showed that within the threshold of 0-0.9,model 5 can provide CLNT patients with the maximum net benefit and its efficacy was significantly superior to that of other models.Conclusion:1.The clinical model(Model 1)based on ultrasound features has a certain predictive ability for CLNT,but its stability is poor,with low sensitivity and accuracy in the validation set.2.B-US and CEUS combined model(Model 4)has higher predictive ability for CLNT than the clinical model(Model 1).Although the AUC comparison between the two models is not statistically significant,the former significantly improves sensitivity and accuracy in the validation set.3.The clinical and radiomics combined model(Model 5)exhibits the best predictive ability for CLNT,with the highest AUC,sensitivity and accuracy in the validation set,and Nomogram shows the results more intuitively in model 5.Clinical decision curve analysis shows that Model 5 has greater clinical application value in predicting CLNT and helps to achieve early precision treatment.