Research On Data Fusion Based On Nonnegative Tensotr Factorization

Researches on biomedical imaging have been developing with the advent of new modalities, which provides new tools for the description of biological tissues. Reserch on fusion methods benefits the integration of multimodal image information and facilitates the interpretation of new modalities, which is useful for a wide range of applications.Multi-modal data fusion is a method which integrates different forms of data, and re-expresses them in a low-dimensional form. By using the method, the essential characteristics of targets can be described. Nowadays, utilizing data fusion techniques to achieve brain tumor recognition has been a tough issue in the area of electronic information and biomedical engineering. Non-negative tensor factorization(NTF) is an advanced blind source separation technique. Under the non-negative constraint, using tensor decomposition algorithm to achieve dimensionality reduction, feature extraction, fusion process is effective for patients with brain tumors MRI and MRSI signal.1. In this thesis, the basic idea of magnetic resonance imaging(MRI) and magnetic resonance spectroscopy imaging(MRSI) are introduced. Moreover, the characteristics and advantage of these two types of images are analyzed. In order to solve the problem that poor signal may affect the outcome when MRSI signals was extracted, this thesis adopts some methods to perform signal preprocessing.2. Focused on the current feature extraction algorithms in the MRSI image, several conventional algorithms are analyzed and their characteristics are discussed. Their accuracy of feature extraction is compared and analyzed using MRSI data of glioblastoma(GBM) of human brain.3. Researchs on nonnegative-tensor decomposition and its application are introduced. Non-negative tensor decomposition is applied to solve some defects which exist in the traditional feature extraction algorithms. Meanwhile, some basic concepts and operations of non-negative tensor decomposition are explained. Here two basic algorithms including CP decomposition and TUCKER decomposition are applied to solve the problem of feature extraction.4. Feature extraction of MRSI spectral data is achieved by non-negative tensor decomposition, and then the MRS spatial distribution is investigated. By using the weighted average method the integration of different modes of data is achieved. In terms of the complementary advantages of these two kinds of data, valuable results of brain tumor recognition is provided by leverage resolution and accuracy. Finally, the experiments on real patient data indicate that the proposed method has a high diagnostic value in clinical application.