The Value Of CT Radiomics In Differential Diagnosis Of Pelvic Rhabdomyosarcoma And Yolk Sac Tumor In Children

Objective: To differentiate the pelvic rhabdomyosarcoma（RMS）and yolk sac tumor（YST）in children by CT radiomics model,and to evaluate its value for diagnosis of RMS.Materials and methods: We retrospectively analyzed 68 cases of pelvic RMS（n=37）and YST（n=31）in children,which have performed plain and contrast CT cans and were pathologically proved.Collect clinical and imaging data of each patient.We segmented tumor area and extracted radiomics features from CT non-enhancement phase,arterial phase and venous phase images.Respectively,1321 initial features were extracted from the three phases,we used 10-fold cross validation strategy,and used LASSO method to reduce the dimension of training set for every fold,and selected five radiomics features with the highest relations to the identification of RMS and YST to set up three machine learning models of support vector machine（SVM）,logical regression（LR）and random forest（RF）.The receiver operating characteristic（ROC）curves were used to evaluate the predictive efficacy of the models,and to calculate predictive indicators such as area under curve（AUC）of ROC,ccuracy,sensitivity and specificity.The AUC differences of ROC curves of multiphase CT were compared by Delong test.Results:1.General data: of the 68 patients,37 were in RMS group,15 males and 22 females,aged 0.5 to 11.7 years.The specific sources included: 28 cases in pelvic cavity and extraperitoneal area,6 cases in perianal area and sacrococcygeal area,3 cases in vagina;31 cases in YST group,2 males and 29 females,aged 0.4 to12 years,including 16 cases in ovary,1 case in pelvic area,9 cases in perianal area and sacroccygeal area,and 5 cases in vagina.Clinical manifestations of tumor: patients were mainly treated with abdominal pain,abdominal mass,dysuria or defecation and other first symptoms.Conventional CT performance: tumor size,density,boundary,uniformity,CT enhancement,sacrococcygeal bone destruction and ascites had no significant difference between RMS and YST（P > 0.05）.Both RMS and YST showed large solid mass in the pelvis,with uneven and slightly low density,without calcification,fat or bleeding.In arterial phase,tumor vascular shadow was seen,in venous phase and delayed phase,the lesions were further enhanced,and the enhanced range was expanded.However,RMS was more likely to have lymph node or distant metastasis than YST（P = 0.045）.Therefore,it was difficult to identify the two by routine radiological examination.In this study,54%（20 / 37）of RMS were misdiagnosed as YST or could not be distinguished from YST on CT.2.The results of feature selection of radiomics: after dimensionality reduction of the features applied by LASSO combined with 10 fold crossvalidation method.（1）there were 5 texture features（wavelet-HLH_glszm_SmallAreaLowGrayLevelEmphasis 、 wavelet-LHH_glrlm_RunVariance 、 wavelet-LLH_glcm_Correlation 、 wavelet-LLH_glcm_Correlation and logarithm_gldm_DependenceEntropy）selected for nonenhancement phase;（2）2 intensity（original_firstorder₁0Percentile and square_firstorder₁0 Percen tile）,3 texture features（squareroot_glrlm_RunEntropy、original_glcm_MaximumProbability and logarithm_glrlm_RunEntropy）were selected for arterial phase;（3）2 intensity（exponential_firstorder₁0Percentile and wavelet-HLH_firstorder_Minimum）,3 texture features（squareroot_glcm_InverseVariance 、 wavelet-HLH_gldm_LargeDependenceHighGray Level Emphasis and wavelet-HHL_glcm_SumEntropy）were also selected for venous phase.3.Classification results predicted by radiomics model: based on the above optimal feature set,three classifier models of SVM,LR and RF were established.（1）RF model had the best efficiency in distinguishing RMS from YST,and CT arterial phase showed higher classification efficiency.（2）In the test set,the AUC,accuracy,sensitivity and specificity of the arterial phase were 0.83（95% CI: 0.72⁰.91）,0.81,0.84 and 0.77,respectively.（3）Relatively low prediction results were found in venous phase and nonenhancement phase.the AUC,accuracy,sensitivity,and specificity of the the venous phase and non-enhancement phase were 0.78（95% CI:0.66⁰.87）,0.75,0.84,0.65,and 0.71（95% CI:0.58⁰.81）,0.72,0.81,and 0.61,respectively.（4）According to Delong test,there was no significant difference in AUC between arterial phase and venous phase,arterial phase and non-enhancement phase,venous phase and non-enhancement phase（P = 0.684,0.491,0.801）.The prediction performance of the SVM and LR models based on the best CT arterial phase were: the AUC,accuracy,sensitivity,and specificity of the SVM model in the test set were 0.78（95% CI:0.67⁰.88）,0.74,0.74 and 0.65,respectively;the AUC,accuracy,sensitivity,and specificity of the LR model were 0.77（95% CI:0.63⁰.84）,0.74,0.71 and 0.71,respectively.RF models based on CT arterial phase generally outperformed other models.Conclusions:1.After feature dimensionality reduction,in the optimal five radiomics feature sets selected for CT 3 phases images,the selected texture features were significantly better than the first-order features.（1）The radiomics features of GLDM,GLCM,GLRLM and GLSZM with diagnostic efficacy were extracted in the non-enhancement phase,which indicated that there was a difference in texture and gray heterogeneity between RMS and YST tumors;（2）The firstorder,GLCM and GLRLM were extracted in the arterial phase,which indicated that there was a difference in the heterogeneity of the voxel intensity,gray-scale length and distribution between RMS and YST tumors;（3）The firstorder,GLDM and GLCM were extracted in the venous phase,which indicated that there was a difference in voxel intensity and gray-scale distribution between RMS and YST tumors.2.Among SVM,LR and RF,the RF model had the best performance in distinguishing RMS from YST.Compared with the previous radiomics research,it was verified that RF classifier was a commonly used classification model,which had excellent classification performance in small sample,nonlinear and high-dimensional data.3.The ROC curve of RF classifier based on three-stage CT images showed that AUC and accuracy were more than 0.7,which were significantly higher than the accuracy of our radiologists in the diagnosis of RMS in the image report（46%）.And the arterial phase showed the relatively high prediction results,however,by Delong test,there was no significant difference in AUC between arterial phase and venous phase,arterial phase and non-enhancement phase,venous phase and non-enhancement phase（P > 0.05）.In summary,the radiomics model based on CT images was preoperatively useful in differential diagnosis of RMS and YST,which may help to increase the confidence of radiologists in the diagnosis of pelvic solid tumors in children,and increase confidence of radiologist in the diagnosis of pelvic solid tumors in children.