Research And Implementation Of Classification Model Of High And Low Level Brain Glioma Based On Radiomics

Brain glioma is the second leading cause of cancer deaths for adolescent worldwide.Accurate grading of gliomas is critical to optimizing treatment and prognosis.Histopathological analysis of tumor tissue by needle biopsy is the main method for grading of glioma Currently.However,brain puncture is associated with a higher risk of surgery.It is prone to sampling errors and has occasional complications.Currently,magnetic resonance imaging based brain scan provides an effective non-invasive means for diagnosis and grading of glioma.In recent years,"Radiomics" has carried out in-depth mining of MR image data.High-throughput features are extracted from massive medical image data,and quantitative analysis was used to improve the diagnostic efficiency and accuracy of glioma.In this paper,the radiomics method is used to study the classification of gliomas based on MR images.High-throughput features were extracted from MR images by interactive segmentation method,and 10 important image features were screened by feature dimension reduction method.Finally,high-and low-level classification models were constructed by machine learning method,provide decision support for personalized treatment of glioma.The main research content of this paper consists of the following three parts:(1)Study the segmentation of the region of interest(ie,the lesion portion)of the glioma MR image.Aiming at the problem that the boundary of the tumor region is blurred due to the uniform distribution of gray scale in MR image,an interactive segmentation algorithm based on local structure tensor and classical GrowCut is proposed.By analyzing the local structure tensor and neighborhood gray similarity of the image,the evolutionary function of the gray-scale similarity in the neighborhood is considered in the classical GrowCut algorithm,and the segmentation precision of the tumor image boundary is improved.The effectiveness of the method is verified by comparative experiments.(2)Study the high-throughput feature matrix construction of the lesion area.A variety of high-throughput features such as first-order,morphology and texture were extracted from the three aspects of statistics,biomorphology and image texture.Aiming at the problem of excessive number of extracted feature data,a variety of dimensionality reduction methods are used to screen out highly distinctive feature combinations,reducing redundant or irrelevant features.Statistical analysis is performed on multiple high-throughput features after dimensionality reduction to verify the statistical significance of the data.(3)Study the construction of high and low level classification models based on machine learning.Through the research on SVM and random forest algorithm,the characteristics of each are analyzed.Combining the advantages of SVM and integrated learning,an SVM integrated classification algorithm based on AdaBoost is proposed.This method realizes the hyperparameter optimization of the classification model through grid search,and constructs a classification model of glioma image level,and has passed the experimental verification.The classification accuracy of the method.