The Prediction Of Gliomas Grading Based On Radiomics Features

Glioma is the most common malignant tumor in the central nervous system.According to malignant degree,it can be divided into high-grade glioma and low-grade glioma.In recent years,researches at the genetic level have proved to be of great value in the prediction and prognosis of glioma.However,the grading study of glioma on the molecular level alone is incomplete.Therefore,accurate grading of glioma at the tissue level is the basis for molecular diagnosis,and plays an important role in clinical diagnosis and treatment of patients with glioma.Currently,tumors are clinically graded with surgical or biopsy methods.Due to the invasiveness and heterogeneity,this method has significant limitations.As an non-invasive tool,magnetic resonance imaging(MR)plays an important part in the diagnosis of glioma.Most of the studies on MR images of glioma are for single modal or focus on the first-order characteristics of MR images,such as the mean value and the statistical features such as the maximum and minimum values,which cannot fully represent phenotype information of tumors.Based on this situation,this article uses the sample set of brain glioma MR images provided by BRATS in 2015,and uses PyRadiomics,an open-source data package in Python,to achieve the high-throughput extraction of large amounts of from multi-mode glioma MR images.This essay also uses the support vector machine-recursive feature elimination to select features to obtain an optimal feature subset,which can represent the presentational feature of brain glioma,and then uses the XgBoost algorithm to train the classifiers.As a comparison,ERT and S VM are also used for training of classification.In this paper,we first extracted 420 radiomics features from the four-modality MR images of glioma.These feature data are huge in quantity,have many unrelated features and noise,so cannot be directly used for training research of classifier.So we need to do feature selection first.In this paper we use the support vector machine-recursive feature elimination(SVM-RFE)in the feature selection process.We use 5-fold cross-validation in the whole process and perform five feature selections on the sample set to obtain five optimal feature subsets,then choose features commonly included in these feature subsets as the result of feature selection.It can be found that most texture features and features in the T1c modality are extracted,indicating that texture features and T1c-modality images play an important role in the classification of high-and low-malignant grade glioma.This point of research finds consistency with current domestic and foreign research conclusions.In this paper,we use XgBoost algorithm to train the classifier.We use 5-fold cross-validation to evaluate the performance of the classifier.At the same time,we use the ERT and SVM algorithms to train the classifier as a comparison.The experimental results show that the XgBoost classifier has accuracy,recall ratio,precision ratio and F1,which are 91.25%,91.63%,91.25%and 91.06%respectively,both higher than the ERT and SVM classifiers.This point also shows that the XgBoost algorithm used in this paper has great advantages.The classification trainer designed in this article has high precision and can be used to assist doctors in diagnosing of malignant degree,which is the basis for further molecular diagnosis.