| Background and purpose: Mucinous ovarian cancer(MOC)which is a relatively rare subtype of epithelial ovarian cancer,accounting for about 3%-5%.A part of MOC originates from ovary and the other part can be metastasized from the digestive system,pancreas,uterus,breast and so on.Because of the low incidence and similar clinical features,the accurate diagnosis of primary MOC and metastatic MOC is difficult.Radiomics can extract high-dimensional information from the images which is analyzed through machine learning algorithms and not affected by the naked eye,thus playing an important role in the early diagnosis,stage classification and prognosis analysis.In this study,radiomics models and clinical-radiomics models are constructed based on machine learning algorithm for the diagnosis of primary and metastatic MOC.Materials and Methods 67 patients with mucinous ovarian cancer confirmed by pathology were retrospectively analyzed,including 45 patients with primary MOC and 23 patients with metastatic MOC.They were divided into the training set(47 patients)and the test set(20 patients)according to a ratio of 7:3.Next,the software 3D Slicer was used to manually delineate the tumor region layer by layer on the venous phase(VP)image of enhanced CT to generating the region of interest(ROI)and the Radiomics of 3D Slicer was used to extract radiomics features.The t-test,the Wilcoxon signed rank as well as least absolute shrinkage and selection operator(LASSO)were used to screen the radiomics features.And then,the selected features were applied to the construction of the radiomics models with three machine learning algorithm: random forest,support vector machine and logistic regression.The three different radiomics models were evaluated by ROC curve,AUC value,sensitivity,specificity and accuracy.Furthermore,the clinical features of different histological subtypes of MOC were compared including HE4,CA125,CA199,AFP,CEA,PLT,leukocyte,diameter of tumor,unilateral or bilateral and ascites.Similarly,the t-test,the Wilcoxon signed rank and LASSO were used to screen the clinical-radiomics features.The clinical-radiomics models were constructed based on the clinical-radiomics features through the random forest,support vector machine and logistic regression,respectively.Finally,the radiomics models and clinical-radiomics models were evaluated by ROC curve,AUC value,sensitivity,specificity and accuracy.Results(1)In the radiomics model,851 radiomics features were extracted from VP images of enhanced CT,and 6 radiomics features were selected by t-test and LOSSO to establish the radiomics models through 3 kind of machine learning algorithm.(a)In the training set,the AUC value of the radiomics model based on random forest was 0.975(95%CI 0.969-0.981)with the accuracy of 97.1%(95%CI 96.4 – 97.8%),the sensitivity of 96.0% and the specificity of 100.0%.In the test set,the AUC value of the radiomics model based on random forest was 0.725(95%CI 0.700-0.750),with the accuracy of 67.1%(95%CI 64.6 –69.7%),the sensitivity of 71.0% and the specificity of 75.0%.(b)In the training set,the AUC value of the radiomics model based on support vector machine was 0.822(95%CI0.811-0.832)with the accuracy of 84.9%(95%CI 84.1-85.7%),the sensitivity of 96.0%and the specificity of 74.0%.In the test set,the AUC value of the radiomics model based on support vector machine was 0.663(95%CI 0.633-0.692),with the accuracy of 70.4%(95%CI 67.6-73.2%),the sensitivity of 65.0% and the specificity of 73.0%.(c)In the training set,the AUC value of the radiomics model based on logistic regression was 0.824(95%CI 0.807-0.840)with the accuracy of 84.5%(95%CI 62.7-69.2%),the sensitivity of 89.0% and the specificity of 74.0%.In the test set,the AUC value of the radiomics model based on support vector machine was 0.660(95%CI 0.627-0.692),with the accuracy of69.7%(95%CI 67.2-72.1%),the sensitivity of 71.0% and the specificity of 75.0%.The constructed radiomics model was run for 50 times,and the radiomics model based on random forest algorithm was better than the other two models in terms of AUC value and accuracy.(2)In the clinical radiomics model,there are significant differences in age,CA125,CEA and tumor laterality between primary and metastatic MOC(P<0.05).Four clinical features and 851 radiomics features were selected to established clinical-radiomics model after the selection of t-test and LASSO.(a)In the training set,the AUC value of the clinicalradiomics model based on random forest was 0.987(95%CI 0.982-0.991)with the accuracy of 98.7%(95%CI 98.2%-99.1%),the sensitivity of 100.0% and the specificity of 100.0%.In the test set,the AUC value of the clinical-radiomics model based on random forest was 0.799(95%CI 0.775-0.823)with the accuracy of 76.6%(95%CI 74.7%-78.6%),the sensitivity of 100.0% and the specificity of 88.0%.(b)In the training set,the AUC value of the clinical-radiomics model based on support vector machine was 0.965(95%CI 0.959-0.970)with the accuracy of 96.1%(95%CI 95.5-96.8),the sensitivity of89.0% and the specificity of 95.0%.In the test set,the AUC value of the clinical-radiomics model based on support vector machine was 0.724(95%CI 0.702-0.746),with the accuracy of 70.0%(95%CI 67.7%-72.4%),the sensitivity of 88.0% and the specificity of75.0%.(c)In the training set,the AUC value of the clinical-radiomics model based on logistic regression was 0.963(95%CI 0.955-0.972)with the accuracy of 96.4%(95%CI95.5-97.2),the sensitivity of 96.0% and the specificity of 100.0%.In the test set,the AUC value of the clinical-radiomics model based on logistic regression was 0.719(95%CI 0.687-0.750),with the accuracy of 73.8%(95%CI 72.0-75.6%),the sensitivity of 82.0% and the specificity of 50.0%.The constructed clinical-radiomics model was run for 50 times,and the clinical-radiomics model based on random forest algorithm was better than the other two models in terms of AUC value and prediction accuracy.(3)Compared with the radiomics model,the clinical-radiomics model has better resolving ability under the same machine learning algorithm(P < 0.05).Conclusion(1)The radiomics model based on VP images of enhanced CT showed good predictive performance in primary and metastatic MOC.(2)Among the three machine learning algorithms,random forest algorithm showed the best performance in the identification of primary and metastatic MOC in both the radiomics model and the clinical-radiomics model.(3)The discriminant performance of clinical–radiomics model in random forest,support vector machine and logistic regression model is better than that of simple radiomics model.(4)Among the six machine learning models,the model based on random forest algorithm and clinical-radiomics features has the best predictive performance. |
PDF Full Text Request