TISSUE TYPE DIFFERENTIATION FOR BRAIN GLIOMA USING NON-NEGATIVE MATRIX FACTORIZATION

摘要

The purpose of this paper is to introduce a hierarchical Non-negative Matrix Factorization (NMF) approach, customized for the problem of blindly separating brain glioma tumor tissue types using short-echo time proton magnetic resonance spectroscopic imaging (~1H MRSI) signals. The proposed algorithm consists of two levels of NMF, where two constituent spectra are computed in each level. The first level is able to correctly detect the spectral profile corresponding to the most predominant tissue type, i.e., normal tissue, while the second level is optimized in order to detect two 'abnormal' spectral profiles so that the 3 recovered spectral profiles are least correlated with each other. The two-level decomposition is followed by the reestimation of the overall spatial distribution of each tissue type via standard Non-negative Least Square (NNLS). This method is demonstrated on in vivo short-TE ~1H MRSI brain data of a glioblastoma multiforme patient and a grade II-III glioma patient. The results show the possibility of differentiating normal tissue, tumor tissue and necrotic tissue in the form of recovered tissue-specific spectra with accurate spatial distributions.