| Objective 1.To explore the differential diagnostic value of radiomics based on T2WI images for primary intracranial lymphoma,glioblastoma,and brain metastases.2.To explore the differential diagnostic value of radiomics based on T1WI-enhanced images for primary intracranial lymphoma,glioblastoma,and brain metastases.Materials and Methods 1.General informationRetrospective analysis of 38 cases of primary intracranial lymphoma,51 cases of glioblastoma,and 66 cases of brain metastases treated in our hospital from January 2010 to January 2020,including 72 males and 83 females,with an average age 58.65±11.34 years.All patients were scanned by 3.0T magnetic resonance equipment(GE Discovery HD750 3.0TMRI and Siemens verio 3.0TMRI)according to standard scanning specifications,and axial T2WI and T1WI-enhanced images were obtained.Inclusion criteria :(1)the lesion was confirmed by pathology after surgery or puncture,and the diagnosis was clear;(2)the number of brain metastatic tumors was no more than 3,no large edema and small tumor on the imaging,and the case was misdiagnosed in clinical work;(3)complete image data;(4)no chemoradiotherapy or hormone therapy was given before MRI examination.Exclusion criteria :(1)the image artifact is large and the resolution cannot attain the requirements;(2)no contrast-enhanced MRI was performed.2.Image segmentationThe axial T2WI and T1WI-enhanced images of the collected patients were imported into Radcloud,Huiying Medical Technology Co,Ltd.in DICOM format.Without understanding the clinical data and pathological diagnosis of the patients,the lesion ROI was manually drawn layer by layer to obtain the VOI of the lesion,including cystic degeneration,necrosis and hemorrhage inside the lesion.3.Extraction and screening of feature valuesThe Radcloud platform is used to extract quantitative image features from two sets of MR images,which can be divided into four groups,namely first-order statistical features,shape features,texture features and higher-order statistical features,with a total of 1409 feature values.Then the extracted feature values were reduced by Variance Threshold,Select KBest and Lasso Method respectively,and the optimal feature values that can be used for differential diagnosis were selected.4.Statistical analysisCorrelation analysis of general data was performed on SPSS 22.0 statistical software.If each group of data conforms to the normal distribution and has homogeneous variances,one-way anova was adopted;if not,the rank sum test of multiple groups of data was used.If the P value is less than 0.05,the difference is statistically significant.The rest of the statistical analysis was performed on Radcloud platform of radiomics.All VOIs were randomly divided into training data sets and test data sets according to 4:1,and six classifiers(K Nearest Neighbor,Support Vector Machine,Extreme Gradient Boosting,Random Forest,Logistic Regression,Decision Tree)were used to establish the classification model of the selected optimal feature values.The ROC curve was used to analyze the predictive performance of each classifier.The larger the area under the curve(AUC)of the test group,the higher the diagnostic performance of the model was.The AUC,Sensitivity,Specificity and F1 value of different classifiers were calculated.The higher the value of F1,the better the stability of the model.Result 1.General informationThere were 38 cases of primary central nervous system lymphoma in the group,including 18 males and 20 females,with an average age of 59.29±13.01 years;51 cases of glioblastoma,including 26 males and 25 females,with an average age of 55.27±11.05 years;66 cases of brain metastases,including 28 males and 38 females,with an average age of 60.88±9.99 years.Among the primary tumors were 47 lung cancers,8 breast cancers,2 renal cancers,2 cervical cancers,2 rectal cancers,1 pleural mesothelioma and 1 bladder cancer.There were no significant differences in age and gender between the three groups(P> 0.05).2.Feature value screening results(1)1409 feature values were extracted based on conventional T2WI images,and 893 feature values were obtained after dimensionality reduction by Variance Threshold,and 439 feature values were obtained after dimensionality reduction by Select KBest.Finally,20 optimal feature values were obtained by the Lasso method,including 1 shape feature,3 first-order features,11 texture features,and 5 higher-order features.(2)1409 feature values were extracted based on T1WI-enhanced images,and 898 feature values were obtained after dimensionality reduction by Variance Threshold,and 506 feature values were obtained after dimensionality reduction by Select KBest.Finally,35 optimal feature values were obtained by the Lasso method,including 2 shape feature,4 first-order features,10 texture features,and 19 higher-order features.3.Results of prediction models based on radiomics(1)The AUC values of the identification model constructed by KNN classifier in T2WI images are the best,which are 0.92,0.77 and 0.89 respectively.(2)The AUC values of the identification model constructed by LR classifier in T1WI-enhanced images are the best,which are 0.93,0.95 and 0.82 respectively.Conclusion 1.Radiomics methods based on conventional T2WI images and T1WI-enhanced images can differentiate and diagnose primary intracranial lymphoma,glioblastoma,and brain metastases.2.Among the six classifiers,the LR classifier based on T1WI-enhanced images has the highest diagnostic performance and the best model stability. |
PDF Full Text Request