Preliminary Study Of MRI-based Computer-aided Diagnosis Of Parasellar Cavernous Hemangioma And Meningioma

Purpose1.To explore the diagnostic efficacy of the same kind of classifier based on computer-assisted diagnosis of different MR sequence images in the diagnosis of parasellar cavernous hemangioma and meningiomas.The aim of this study is to select the best MR sequence for computer-aided diagnosis.2.To explore the diagnostic efficacy of different kinds of classifiers based on computer-assisted diagnosis of the same MR sequence image in the diagnosis of parasellar cavernous hemangioma and meningioma.The aim of this study is to select the best classifier for computer-aided diagnosis.3.To compare the diagnostic value of computer-aided diagnosis based on MRI and DWI in differentiating parasellar cavernous hemangioma from meningioma,and to provide reference for radiologists.Materials and Methods1.Image data acquisition: retrospective analysis the MR images of the parasellar cavernous hemangioma of parasellar meningioma from January 2010 to November2018,among The Second Affiliated Hospital of Dalian Medical University,the First Affiliated Hospital of Wenzhou Medical University,and the First People’s Hospital of Jining City.The criteria for enrollment were as follows:(1)All patients were confirmed by surgery or clinical trial treatment;(2)complete imaging data;(3)the main body of the lesion was located in the area around the cavernous sinus or cavernous sinus,and the cranial fossa and saddle were invaded before and after Districts and bungee areas are not restricted,and the scope of its invasion is not limited.Exclusion criteria were as follows: Images with large image artifacts and resolutions that do not meet the analytical application.Finally,32 cases of cavernous hemangioma in the saddle and 62 cases of meningioma were included.All patients underwent T1 WI,T2WI and enhanced T1WI;26 patients with parasellar cavernous hemangioma and 38 patients with meningioma underwent DWI examination.Due to the susceptibility artifact interference,24 cases of parasellar cavernous hemangioma and 31 cases of parasellar meningioma DWI image were achieved finally.ADC map of 21 cases of saddle spongy Hemangioma and 21 cases of parasellar meningioma were achieved finally due to the low version of the MR scanner software version,which cannot reconstruct the ADC map.2.Image import and volume extraction of interest: The MRI of DICOM format is classified and uploaded to the radiology platform 2.1V [Huiying Medical Technology(Beijing)Co.,Ltd.].Customizing grayscale to achieve image grayscale normalization,then manually delineate the lesion contour as the volume of interest(VOI)in T1 WI,T2WI,DWI,ADC map and enhanced T1 WI.3.Feature calculation and dimensionality reduction: Calculating the image feature value generation which based on feature class and filter-based feature parameters.Three-dimensional dimensionality reduction analysis by variance selection method,single variable selection method and lasso method were used to calculatethe MRI main feature value between two groups of cases in different sequences.4.Classification identification: 80% of VOI in each grouph trained by six classifiers separately,that is K-Nearest Neighbor(KNN),Support Vector Machines(SVM),Extreme Gradient Boosting(XGBoost),Random Forest(RF),Logistic Regression(LR)and Decision Tree(DT),and the remaining 20% of VOIs were tested by the six above classifiers.5.Statistical analysis:1)Comparing and analyze the area under the ROC curve with 95% confidence interval,sensitivity,specificity and accuracy of each classifier based on different MR sequence test results,using F-Score value to evaluate the stability of the model.The larger the F-Score value is,the better the stability the model has.2)In this study,the relationship between the lesion diameter and the detection rate of 64 lesions was statistically analyzed on the DWI image using SPSS 24.0 software.The ROC was used to evaluate the area under the ROC(AUC).The better the resolution,the AUC value is usually > 0.7 has clinical value;according to its sensitivity and specificity,the Yoden index is obtained,the long-distance cut-off value is found,and the accuracy is calculated.The same method was used to perform statistical analysis on the ADC values of the two.Results1.KNN,SVM,XGBoost,RF,LR classifiers based on T2 WI and KNN,LR,and RF classifiers based on ADC images identify the accuracy of the diagnosis of parasellar cavernous hemangioma and meningioma higher than 0.90.The F-Score value is greater than 0.90.2.The accuracy of DT classifier based on enhanced T1 WI,SVM and RF classifier based on T1 WI,XGBoost,DT and SVM classifier based on the ADC images for identifying parasellar cavernous hemangioma and meningioma from 0.80 to 0.90,and the F-Score was ranged from 0.80-0.90.Theaccuracy of the other classifiers for identifying the parasellar cavernous hemangioma and meningioma based on different MR sequences was less than or equal to 0.80,and the F-Score values were less than 0.80.3.In the DWI sequence,the detection rate of cavernous hemangioma was 24/26,and the detection rate of parasellar cavernous hemangioma was 31/38.The average value of the undetected was about 1.95 ± 0.48.The detection rate of conventional MRI lesions was 100%.The ADC value of parasellar cavernous hemangioma was(1.458±0.214)×10-3mm2/s;the ADC value of parasellar meningioma was(0.915±0.151)×10-3mm2/sConclusion1.Computer-aided diagnosis based on T2 WI and ADC images is more valuable when KNN,SVM,XGBoost,RF,and LR classifiers are all used to identify parasellar cavernous hemangioma and meningioma.2.KNN,SVM,XGBoost,RF,LR classifiers based on T2 WI and KNN,LR,and RF classifiers based on ADC images have similar diagnostic efficiency in the identification of cavernous hemangioma and meningioma.3.Computer-aided diagnosis based on conventional MRI is of great value when the size of the parasellar cavernous hemangioma or parasellar meningiomas is too small.