论文部分内容阅读
Purpose: To perform data mining of the multimodal diagnostic MRI images with support vector machine(SVM)model learning based on the combination of image textures and histograms and to optimize the computer aided diagnosing(CAD)system for IDH1 negative or positive evaluation of gliomas.Methods: 120 patients with a subsequent histologically confirmed diagnosis of glioma after surgery were included retrospectively.The texture features were retrieved from dynamic contrast enhancement MRI(DCE-MRI),diffusion weight image MRI(DWI-MRI)and 3-D arterial spin labeling(3D-ASL)images by using our previously optimized parameters.The individual parameters from DCE-MRI,DWI-MRI and 3D-ASL were calculated at the pixel level so as to build the relevant histograms.An SVM learning technique using leave one out cross-validation(LOOCV)was applied to these multimodal MR imaging-based texture features or histograms separately or in combination for the IDH1 prediction.The kernels,dataset threshold,number of features for SVM to obtain a satisfying accuracy were optimized.Results: Our pilot analysis revealed that data scaling,sample augmentation with SMOTE or linear kernel model helps improve the diagnostic accuracy of SVM learning.Thus,we applied the optimized SVM learning with the above mentioned procedures.Compared with conventional parameters(maximal accuracy = 79.6%)derived from DCE-MRI,DWI-MRI or 3D-ASL,the whole tumor multimodal texture feature-(accuracy = 91.5%)or histogram(accuracy = 89%)-SVM demonstrated higher classification accuracies of IDH1,respectively.Combining texture and histogram features further increased the accuracy to 94.5%.A satisfying accuracy of SVM is maintained till the 1-side or 2-side data filter reached 35%or 25%,respectively.Besides,the satisfying classification accuracy was obtained when 250-650 mostly weighted features were selected for SVM learning.Discussion: Combining image textures and histograms of multimodal MRI images for SVM learning significantly improved the accuracy of gliomas IDH1 evaluation.Our optimized data preprocessing and SVM learning procedure is user friendly and might pose potential in the future clinical use.