Radiomics Based On MRI For Differentiating Benign And Malignant Of Soft Tissue Tumors And Predicting Histopathological Grade Of Soft Tissue Sarcomas

Purpose To explore the value of radiomics model for preoperatively predicting the pathological differentiation of benign and malignant soft tissue tumors and the grade of soft tissue sarcoma using MRI.Materials and Methods In the first part,the retrospective study enrolled 68 patients who had undergone 3.0T magnetic resonance imaging before surgery.FS-T₂ WI and FS-T₁WI+C were collected to extract the radiomics features.The least absolute shrinkage and selection operator?LASSO?algorithm was used to reduce the features and select valuable ones for preoperative pathological diagnosis.K-nearest neighbor?k NN?and support vector machine?SVM?was trained using 5-fold cross-validation to develop three classification models?FS-T₂ WI,FS-T₁WI+C as well as combined FS-T₂ WI and FST₁WI+C?from the selected features.These three models were evaluated by the receiver area under the operating characteristics curve?AUC?.In the second part,thirty-five patients who were pathologically diagnosed with soft tissue sarcomas and their histological grades were recruited.All patients had undergone MRI before surgery on a 3.0T MRI scanner.Radiomics features were extracted from FS-T₂ WI.We used the LASSO regression method to select features.Then three machine learning classification methods,including random forests,k NN,and SVM algorithm were trained using the 5-fold cross validation strategy to separate the soft tissue sarcomas with low-and high-histopathological grades.In the third part,51 selected patients who were pathologically diagnosed of STSs were recruited in this study.The acquired ADC VOI automatically extracts the radiomics features.Then the LASSO algorithm was used to reduce the features and select valuable ones for preoperative pathological diagnosis.The parameters with statistical differences in the univariate analysis were incorporated into the multivariate logistic regression analysis to determine the independent predictors related to the grading,and the radiomics nomogram was constructed by the independent predictors.ROC,calibration curve and decision curve analysis are used to evaluate the diagnostic efficacy and clinical value of the model.Results In the first part,the model that used SVM method achieved the better performance than k NN,FS-T₂ WI,FS-T₁WI+C as well as combined FS-T₂ WI and FS-T₁WI+C MRI model showed a diagnostic accuracy of 0.86?AUC,0.88±0.07?,0.83?AUC,0.86±0.08?,0.87?AUC,0.89±0.08?,respectively.In the second part,the model that used SVM method achieved the best performance among the three methods,with AUC values of 0.92±0.07,accuracy of 0.88.In the third part,peritumoral high signal intensity on T₂-weighted images and rad-score was found to be the independent factors for predicting the STSs grade.The AUC?95% CI?and accuracy of the rad-score was 0.87?0.77-0.97?and 0.82,respectively.The AUC?95% CI?and accuracy of the radiomics nomogram integrating peritumoral high signal intensity on T₂-weighted images and rad-score was 0.90?0.81-0.99?and 0.88,respectively.Conclusions Radiomics features could be used as candidate biomarkers for distinguishing benign and malignant soft tissue tumors and predicting the grade of soft tissue sarcoma noninvasively before surgery.Part One Preoperative Pathological Differentiation of Benign and Malignant Soft Tissue Tumors based on Radiomics of Conventional MRIPurpose To explore the value of radiomics for preoperatively predicting the pathological differentiation of both non-malignant and malignant soft tissue tumors using MRI.Materials and Methods 1.Clinical data From October 2012 to October 2018,A total of 68 patients?mean age,51 years;age range,17–90 years;35 men and 33 women?with STTs who had undergone Siemens 3.0 T MRI including FS-T₂ WI and FS-T₁WI+C in our institution were recruited.34 were classified as non-malignant and 34 were classified as malignant.2.MRI equipment and method All examinations were performed by a Magnetom verio 3.0 T MRI?Siemens,Erlangen,Germany?with suitable coils to use depends on the location of the lesion.3.Tumor segmentation Import the acquired FS-T₂ WI and FS-T₁WI+C original DICOM images into the postprocessing platform Radcloud?Huiying Medical Technology Co.,Ltd?.The regions of interest?ROI?were delineated by hand along the outline of the lesion on FS-T₂ WI and FS-T₁WI+C,and the computer automatically generated the three-dimensional volume of interest?VOI?.Doctor 1 performed tumor segmentation in all 68 patients,and Doctor 2 performed tumor segmentation in 30 patients who were randomly selected among the 68 patients to analyze the interobserver agreement of radiomics feature extraction.4.Radiomics Feature Extraction and Selection The acquired VOI automatically extracts the radiomics features by using the platform mentioned above.For each radiomic feature,the intraclass correlation coefficient?ICC?was calculated to quantify reproducibility between the test-retest scans.Then the leastabsolute shrinkage and selection operator?LASSO?algorithm was used to reduce the features and select valuable ones for preoperative pathological diagnosis.5.Statistical Analyses k NN and SVM analysis were trained using 5-fold cross-validation to develop three classification models?FS-T₂ WI,FS-T₁WI+C as well as combined FS-T₂ WI and FST₁WI+C?from the selected features.These two models were evaluated by the receiver area under the operating characteristics curve?AUC?and statistical differences between every two models in every machine learning method were compared by De Long test.Results 1.Radiomics Feature Selection 1029 quantitative imaging features were extracted from each FS-T₂ WI and FS-T₁WI+C,there were 983 and 977 features with an excellent reproducibility?ICC higher than 0.75?were included in the further selection process.And Lasso feature selection algorithm was used to select 4 features from each sequence to build the radiomics model.2.Results of differentiation of Benign and Malignant Soft Tissue Tumors based Radiomics of MRI The model that used support vector machine classification method achieved better performance than k NN machine learning methods,and the FS-T₂ WI,FS-T₁WI+C as well as combined FS-T₂ WI and FS-T₁WI+C MRI model using SVM showed a diagnostic accuracy of 0.86?AUC,0.88±0.07?,0.83?AUC,0.86±0.08?,0.87?AUC,0.89±0.08?respectively.Conclusions 1.The radiomics based on FS-T₂ WI,FS-T₁WI+C as well as combined FS-T₂ WI and FS-T₁WI+C MRI model using k NN method has the ability for differentiation of benign and malignant soft tissue tumors,and these three radiomics model shows almost the same role in the differentiation of benign and malignant soft tissue tumors.2.The radiomics based on FS-T₂ WI,FS-T₁WI+C as well as combined FS-T₂ WI and FS-T₁WI+C MRI model using SVM method has the ability for differentiation of benignand malignant soft tissue tumors,and these three radiomics model shows almost the same role in the differentiation of benign and malignant soft tissue tumors.3.The model using the SVM achieved the better diagnostic performance than k NN.Part Two Soft tissue Sarcomas: Preoperative Predictive Histopathological grading based on Radiomics of FS-T₂WIPurpose The purpose of this study is to develop a radiomics model for predicting the histopathological grades of soft tissue sarcomas preoperatively through MRI.Materials and Methods 1.Clinical data Thirty-five selected patients who were pathologically diagnosed of STSSat our hospital between August 2013 and May 2018 were recruited in this study including 22 males and 13 females from 31 to 85 years old and their mean age is 59±12 years.Lesions were graded according to the FNCLCC system.2.MRI equipment and method All examinations were performed by a Magnetom verio 3.0 T MRI?Siemens,Erlangen,Germany?with suitable coils to use depends on the location of the lesion.3.Tumor segmentation,Radiomics Feature Extraction and Selection Import the acquired FS-T₂ WI original DICOM images into the post-processing platform Radcloud?Huiying Medical Technology Co.,Ltd?.The regions of interest?ROI?were delineated by hand along the outline of the lesion on FS-T₂ WI,and the computer automatically generated the three-dimensional volume of interest?VOI?.The acquired VOI automatically extracts the radiomics features by using the platform mentioned above.We used the LASSO regression method to select features.4.Statistical AnalysesThen three machine learning classification methods,including random forests,knearest neighbor and support vector machine algorithm were trained using the 5-fold cross validation strategy to separate the soft tissue sarcomas with low and high histopathological grades.Results 1.Radiomics Feature Selection The radiomics features were significantly associated with the histopathological grades.Quantitative imaging features?n=1029?were extracted from fat-suppressed T₂-weighted imaging,and five features were selected to construct the radiomics model.2.Results of predictive histopathological grading soft tissue sarcomas based Radiomics of MRI The model that used support vector machine classification method achieved the best performance among the three methods,with AUC values of 0.92±0.07,accuracy of 0.88.Conclusions The radiomics classifiers based on fat-suppressed T₂-weighted imaging using SVM,k NN and RF show satisfying performance for the classification of the histopathological grades.The radiomics model using the SVM achieved the best diagnostic performance among the three methods.Part Three Preoperative Prediction of Histopathological grading of Soft Tissue Sarcomas based on Radiomics of DWI-ADCPurpose To explore the value of ADC map-based radiomics features as well as nomograms constructed with conventional MRI morphological features and radiomics features in predicting the pathological grade of soft tissue sarcoma.Materials and Methods 1.Clinical data 51 selected patients who were pathologically diagnosed of STSs at our hospital between August 2013 and September 2019 were recruited in this study including 26 males and 25 females from 12 to 85 years old and their mean age is 55±16 years.Lesions were graded according to the FNCLCC system.2.MRI acquisition Twenty-two patients used Magnetom Verio 3.0T system?Siemens,Erlangen,Germany?and used the corresponding coils to acquire images according to the lesion.Twentynine patients used the Discovery MR750 w 3.0T system?General Electric Healthcare,GE,Milwaukee,USA?and acquired images using the corresponding coils according to the lesion.Siemens and GE workstations automatically generate ADC plots based on DWI.Both two Doctors performed tumor segmentation in all 51 patients to analyze the interobserver agreement of radiomics feature extraction.3.Tumor segmentation and Radiomics Feature Extraction Import the original DICOM images of the DWI and ADC images into the the postprocessing platform Radcloud?Huiying Medical Technology Co.,Ltd?.Because DWI has a higher resolution than the ADC map,the ROI is first delineated on a DWI with a b value of 800 s / mm2,and then copied to the corresponding ADC map.A threedimensional VOIs is automatically generated by the computer.The acquired VOI automatically extracts the radiomics features by using the platform mentioned above.4.Radiomics Feature Selection and Rad-score For each radiomic feature,the intraclass correlation coefficient?ICC?was calculated to quantify reproducibility between the test-retest scans.Then LASSO algorithm was used to reduce the features and select valuable ones for preoperative pathological diagnosis.5.MRI characteristics Tumor size,border,signal intensity heterogeneity on T₂-weighted images and peritumor edema were evaluated on conventional MRI.6.Development of radiomics nomogram and statistical analysisUnivariate analysis of tumor size,margin,signal heterogeneity,Peritumoral high signal intensity on T₂-weighted images and Rad-score?an independent t-test or Mann–Whitney U-test were used for continuous variables as appropriate.Fisher's exact test or the ?2 test was used for comparing categorical variables between the two groups,as appropriate?.Then,the parameters with statistical differences in the univariate analysis were incorporated into the multivariate logistic regression analysis to determine the independent predictors related to the grading,and the radiomics nomogram was constructed by combining the independent predictors.ROC,Bootstrap,calibration curve and decision curve analysis are used to evaluate the diagnostic efficacy and clinical value of the model.Results 1.Radiomics selection and calculated rad-score 1409 quantitative imaging features were extracted from the ADC map.In a further selection process,1371 features with high repeatability were selected.The Lasso was used to select the three features with the largest coefficients to build a rad-score.Univariate analysis found that tumor margins,Peritumoral high signal intensity on T₂-weighted images,and signal intensity heterogeneity on T₂-weighted images were statistically different between the low and high-grade groups?p <0.05?.The above three factors and rad-score were put into multivariate logistic regression.Peritumoral and rad-score was found to be the independent factors for predicting the STSs grade.The AUC?95%CI?and accuracy of the rad-score were 0.87?0.77-0.97?and 0.82 respectively.2.Development of radiomics nomogram The AUC?95%CI?and accuracy of the radiomics nomogram was of 0.90?0.81-0.99?and 0.88,respectively.The calibration curve and DCA demonstrate the clinical utility of the nomogram.The predictive power of the nomogram was higher than the rad-score alone.Conclusions 1.Rad-score and peritumoral high signal intensity on T₂-weighted images are independent predictors of STSs grade and has good predictability.2.The nomogram model based on Rad-score and Peritumoral high signal intensity on T₂-weighted images can provide clinicians with personalized graded prediction probabilities.The prediction performance of nomogram model is better than the radscore.