Sparse Feature Learning For Neurodegenerative Disease Early Diagnosis

With the aging of the world,neurodegenerative diseases are spreading,seriously affecting people’s physical and mental health and quality of life.Common neurodegenerative diseases include Parkinson’s disease(PD)and Alzheimer’s disease(AD).The major pathological changes in these diseases are the loss of neuronal cells or the degeneration of neuronal cell structure and function,but the specific causes of these pathological changes have not yet been found.In China,neurodegenerative diseases are still in a low rate of diagnosis,high misdiagnosis rate,and low treatment rate.Therefore,early diagnosis is especially important for timely intervention.Computer aided diagnosis(CAD)utilize subtle brain structural changes provided by neuroimaging techniques,combined with data analysis,to improve the accuracy of diagnosis,and is an effective auxiliary tool for early diagnosis.However,the small sample size and large feature dimension of neuroimaging data are likely to cause overfitting,which is the main problem affecting the accuracy of computer-aided diagnosis.In this paper,an early diagnosis method for neurodegenerative diseases based on sparse learning is proposed for this problem,which mainly includes the following three parts:First,in order to solve the problem of small sample size and large feature dimension,this paper proposes a feature selection method based on sparse regularization.By setting the weight coefficient of most irrelevant features to zero,the disease-related features are screened out,and the multi-classifier is trained by effective features to realize the early diagnosis of neurodegenerative diseases.Second,in order to make full use of neuroimaging data,this paper proposes a feature selection method of Multi-template adaptive sparse learning(MTASL).The method is based on multiple brain segmentation templates with different regions of interest,extracting different feature sets for fusion,using adaptive methods to select adjustable sparsity,and using subspace learning to construct a least squares regression model based on linear discriminant analysis and locally preserved projections take into account both global and local constraint information of the feature space.Third,in order to diagnose more early stages of neurodegenerative diseases,this paper proposes a multi-task low-rank sparse learning(MLRSL)method,which uses the sparseness and low-rank of neuroimaging data for analysis.Construct multi-tasking framework based on data space,capture multi-task correlation through multi-task learning;preserve the most discriminative features through low-rank learning to reveal the intrinsic relationship between input data and output target,and provide multiclassification model the most effective features.In this paper,the early diagnosis methods of neurodegenerative diseases were studied.The feature selection methods based on sparse learning were used to screen out the disease-related features and multi-classification.The experimental results show that the proposed methods produce higher performance in the multi-classification task than state-of-art methods.Using the feature selection methods proposed in this paper,this paper analyzes the brain-related regions of interest related to the disease,and finds some potential brain regions related to the disease,which is helpful for the in-depth further study of neurodegenerative diseases.