Research On The Image Classification Of Brain Glioma Based On Improved Support Vector Machine

Radiomics is the key research direction of computer medical science,and it is valued by governments,hospitals and universities all over the world.The main research areas of Radiomics are cancer and tumor.Accurately diagnosis of glioma is more difficult.There is currently no effective method for accurately grading of glioma.Radiomics achieves grading prediction of tumors through image acquisition,tumor segmentation,feature extraction,and data analysis.This thesis uses the Radiomics method to predict the grade of gliomas.Two sets of glioma imaging data were obtained,285 samples of BRATS2017 glioma dataset and 161 samples of glioma dataset from radiology department of Henan Provincial People’s Hospital,respectively,to verify the effects of this method on scientific research and clinical application.All images were manually segmented by radiologists.Feature extraction algorithms were used to extract features from tumor images.357 features including intensity,shape,texture,and wavelet were extracted.For the extracted features,this thesis uses the minimal-redundancy-maximal-relevance criterion(mRMR)for feature selection,and uses LIBSVM and HLSVM algorithms for training and prediction according to different numbers of feature selection.The results show that the subset of 100 features selected by mRMR is optimal feature subset.Further experiments were performed using the optimal subset of features,using a variety of machine learning evaluation indicators to evaluate these algorithms,and plot the ROC curve.The results show that HLSVM has better training accuracy and faster prediction speed than LIBSVM on the glioma dataset.Experiments were carried out on the optimal feature subset using the 10-fold cross-validation method.The results show that the speed of HLSVM is an order of magnitude faster than the speed of LIBSVM while the training accuracy and prediction accuracy are little different.According to the experimental comparison of the two data sets,BRATS2017 data set is selected by many medical institutions and experts with clear images,high resolution,accurate tumor segmentation.and better experimental results than the dataset of Henan Provincial People’s Hospital,but not universal.The experimental results of Henan Provincial People’s Hospital showed that the average prediction accuracy of the 10-fold crossvalidation using the two classification algorithms were 83.8235% and 80.1102%,which was clinically meaningful.After analyzing the experimental results,it is found that there is data imbalance in the data set.At the same time,the support vector machine algorithm used needs to be improved.Therefore,adaptive synthetic sampling(ADASYN)algorithm is used to oversample the data set,and the dynamic cost support vector machine(DCSVM)is proposed.The adaptive dynamic cost support vector machine predicts the glioma of the glioma,and obtains good experimental results.At the same time,LIBSVM,HLSVM,DEC and FSVM are used for comparison experiments.The results show that the algorithm can effectively improve the prediction accuracy of the minority class in the unbalanced data set,and reduce the rate of misdiagnosis.The G-mean index shows that the algorithm can effectively deal with the imbalance of glioma data.The classification prediction algorithm proposed in this thesis has important practical significance for radiomics and clinical medicine.