Grading And Classification Of Glioma Based On Wavelet Scattering Network

Glioma is a common tumor disease in neural system.Different personalized treatment plans based on different types of glioma will help prolong the prognosis of patients.At present,glioma grading and the subtype detection is mainly derived from biopsy,it is dangerous due to the location specificity of glioma.How to predict the glioma grade and subtype nondestructively is very important for making the personalized treatment plans.Radiomics,as an emerging discipline,by extracting highthrough image features and constructing the relationships between image features and glioma types,it can predict the glioma type from the radiomic features.However,due to the lack of robustness of radiomic features,its predictive ability is limited.Therefore,how to extract more robust and effective features is of great significance for the preoperative diagnosis of glioma.To deal with this issue,this work proposes to use wavelet scattering network to predict the subtype and grade of glioma,detailed as follows:(1)Propose to use wavelet scattering features to predict glioma grade.Quantitatively extract traditional radiomic features and wavelet scattering-based radiomic features from the tumor region in multimodal MRI images.Then establish a prediction model for the histological grade of glioma after feature screening.Under different data,the proposed method was compared with the traditional radiomics for histological prediction of glioma,and the ability of the features extracted from the surrounding area of the tumor on the prediction of glioma was also discussed.In addition,the robustness of the features extracted by the method under the interference of rotation,exchange,and noise addition is analyzed.The results demonstrated that the proposed method could promote the glioma grade prediction accuracy.(2)In view of the problem that most of the existing research rely on delineating the ROI of the tumor area,a deep learning combined with wavelet scattering method is proposed to predict the molecular subtypes of gliomas.This method uses deep learning features,wavelet scattering features and combined with multi-scale ideas.The prediction of molecular subtypes is achieved in the case of the largest tumor area bounding box.At the same time,visual methods and quantitative methods are used to further analyze the superiority of the proposed method in extracting features.