Authors:
Yuqian Li
1
;
Diana M. Sima
2
;
Sofie Van Cauter
3
;
Uwe Himmelreich
4
;
Yiming Pi
5
and
Sabine Van Huffel
2
Affiliations:
1
University of Electronic Science and Technology of China, Katholieke Universiteit Leuven and IBBT-K.U.Leuven, China
;
2
Katholieke Universiteit Leuven and IBBT-K.U.Leuven, Belgium
;
3
University Hospitals of Leuven and Katholieke Universiteit Leuven, Belgium
;
4
Katholieke Universiteit Leuven, Belgium
;
5
University of Electronic Science and Technology of China, China
Keyword(s):
Glioma, Magnetic Resonance Spectroscopic Imaging (MRSI), Non-negative Matrix Factorization (NMF), Blind Source Separation (BSS).
Related
Ontology
Subjects/Areas/Topics:
Applications and Services
;
Biomedical Engineering
;
Biomedical Signal Processing
;
Computer Vision, Visualization and Computer Graphics
;
Medical Image Detection, Acquisition, Analysis and Processing
Abstract:
Finding the brain tumor tissue-specific magnetic resonance spectra and their corresponding spatial distribution is a typical Blind Source Separation (BSS) problem. Non-negative Matrix Factorization (NMF), which only requires non-negativity constraints, has become popular because of its advantages compared to other BSS methods. A variety of algorithms based on traditional NMF have been recently proposed. This study focuses on the performance comparison of several NMF implementations, including some newly released methods, in brain glioma tissue differentiation using simulated magnetic resonance spectroscopic imaging (MRSI) signals. Experimental results demonstrate the possibility of finding typical tissue types and their distributions using NMF algorithms. The (accelerated) hierarchical alternating least squares algorithm was found to be the most accurate.