Research On FMRI Data Feature Analysis And Auxiliary Diagnosis Based On Elastic Net

Functional Magnetic Resonance Imaging（fMRI）data contains a large amount of important information that can be used for medical auxiliary diagnosis.Therefore,fMRI data analysis has always been the focus of clinical research and an important basis for medical auxiliary diagnosis.In order to obtain important features in fMRI data,feature selection is an effective method.It can remove a large number of irrelevant features and screen out features that are strongly related to the classification task.Therefore,it is of great value to carry out research on feature selection algorithms under fMRI data.The feature selection algorithm used in this paper is Elastic net（EN）,which uses L₁ regular term to select features,and L₂ regular term to compress coefficients and realize group selection.In the EN algorithm,the same weight is applied to all coefficients,which may make unimportant variables be retained,and the EN algorithm does not have Oracle properties,that is,it cannot guarantee unbiased estimation of parameters with non-zero coefficients.There are three main tasks in the thesis for the above problems:First,an improved EN algorithm is proposed based on correlation analysis.In order to improve the problems of EN,two weight functions based on the correlation between features and labels are introduced.When feature selection is performed,the correlation between features and labels is also considered,so that the weights of important features become larger,and the weights of unimportant features are compressed to achieve the effect of selectively retaining important variables.Second,a new iterative formula of the improved EN algorithm is given,and the convergence of the iterative formula is proved through theoretical deduction.At the same time,it is derived that the improved EN algorithm can achieve group selectivity,and proves that the improved EN algorithm has Oracle properties,that is,the estimated value of the improved EN algorithm can correctly estimate non-zero parameters with a probability of 1.Third,based on the requirements of the background project,this paper conducts numerical experiments on the use of fMRI data of chronic spontaneous urticaria and four sets of open source high-dimensional small sample data.Experiments show that compared with the EN algorithm,the application of the improved algorithm for feature selection can significantly improve the accuracy of the SVM classification model,the accuracy rate is up to 98.44%,and the improved EN algorithm can significantly reduce the root mean square error of the classification model,which shows that the improved algorithm has better performance.