Multi-Modality Feature Selection With Applications In Brain Disease Classification

Recently, with the development of medical imaging techniques, extracting whole brain structure and the function of the connection model of the human brain based on image data of these medical imaging techniques, and used it for the prediction and diagnosis of brain disease, has become a new focus of research. It has been a research trend to analyze the medical imaging data, to find the rule and further to forecast and to classify the unknown data effectively by using the techniques of machine learning and pattern recognition. We analyze and research for the multi-modality neuroimaging data based on multi-task feature selection method to perform the classification of brain disease. The main work and innovation points are as follows:Firstly, we propose a discriminative multi-task feature selection method to select the most discrim-inative features for multi-modality based classification of brain disease patients from healthy controls. Specifically, for each modality, we train a linear regression model using the corresponding modality of data, and further enforce the group-sparsity regularization on weights of those regression models for joint selection of common features across multiple modalities. Furthermore, we propose a discrimi-native regularization term based on the intra-class and inter-class Laplacian matrices to better use the discriminative information among subjects. After selecting the discriminative and common features across multiple modalities, we then use the multi-kernel SVM method for final classification of brain disease patients from healthy controls. The experiment results show that our proposed method not only improves the classification performance, but also has potential to discover the disease-related biomark-ers useful for diagnosis of disease, so it has large biomedical significance.In addition, we find that existed multi-modality feature selection method using traditional dis-tance, e.g., Euclidean distance, to measure the similarity between two samples, since the static nature of Euclidean distance, it ignores the global structure information between a target sample and all the other samples. Consequently, for considering the dynamic global information of samples adequately, we adopt effective distance take the place of traditional Euclidean distance as a similarity measure way for feature selection learning as well as use the selected more discriminative features to achieve the classification of brain disease patients from healthy controls. Specifically, we first use sparse represen-tation algorithm to obtain effective distance, then define effective distance-based Laplacian matrix, and further propose our effective distance-based multi-modality feature selection method for brain disease classification. Experiment results show that it is helpful for obtaining the global and local information of samples based on effective-distance feature selection algorithm, and can acquire a superior classification performance.