Analysis Of Magnetic Resonance Spectroscopy Signal Based On Bayesian Model

Brain tumor is a disease raised from abnormal cells formed in the brain.The accurate grading of brain tumor can not only help to provide patients with effective treatment timely,but also avoid applying radical treatment to patients leading to additional physiological damages.The main content of this paper is application of magnetic resonance spectroscopy in brain tumor.Magnetic resonance spectroscopy(MRS)is a popular imaging technique in recent years.As it could be used for the detection of metabolites in living tissues,so it's known as "virtual biopsy" technique.The information about metabolites in the tumor can be collected by chemical shift imaging,analysis of which is an important way to the study of brain tumor.First of all,the de-noising methods based on wavelet analysis were studied in this paper.The original MRS data was extracted from SPAR and SDAT files,which are specific file formats defined by Philips.Wavelet decomposition and wavelet packet decomposition based on entropy were applied to analyze the original MRS signal.Combined with four threshold estimation methods,soft or hard shrinking rules and noise estimation methods,the original MRS signal was de-noised.The methods based on Stein's unbiased risk estimation,hard shrinking rule and level-by-level noise estimation was the best for the frequency domain signal,while the method based on threshold entropy,global threshold estimation and hard shrinking rule performed best for time domain signal.After wavelet de-noising,three machine learning methods,Bayesian neural network,support machine vector and random forest,were explored to build the model for cerebral glioma grading in this paper.Before the data was put into the model,feature selection methods based on LI regulation,mutual information,ANOVA F-test,random forest and SVM-RFE were used to for dimension reduction of the quantified MRS data.Although there were some differences among these methods,GSH,Ala and NAA were the important features in common.Then leave-one-out and grid search were applied to tune the model.At last,the accuracies of Bayesian neural network,support vector machine and random forest were 92.6%,96.3%and 96.3%respectively.It can be concluded that MRS has great value and clinical significance in cerebral glioma grading and the grading models based on Bayesian neural network,support vector machine and random forest are helpful for diagnosis for brain tumor.