**Keywords**: non-negative matrix factorization, magnetic resonance imaging, brain tumor
Treatment of brain tumors is complicated by their high degree of heterogeneity. Various stages of the disease can occur throughout the same lesion, and transitions between the pathological tissue regions (i.e. active tumor, necrosis and edema) are diffuse [@price2006improved]. Clinical practice could benefit from an accurate and reproducible method to differentiate brain tumor tissue based on medical imaging data.
We present a hierarchical variant of non-negative matrix factorization (hNMF) for characterizing brain tumors using multi-parametric magnetic resonance imaging (MRI) data [@sauwen2015hierarchical]. Non-negative matrix factorization (NMF) decomposes a non-negative input matrix *X* into 2 factor matrices *W* and *H*, thereby providing a parts-based representation of the input data. In the current context, the columns of *X* correspond to the image voxels and the rows represent the different MRI parameters. The columns of *W* represent tissue-specific signatures and the rows of *H* contain the relative abundances per tissue type over the different voxels.
**hNMF** is available as an *R* package on CRAN and compatible with the **NMF** package. Besides the standard NMF algorithms that come with the **NMF** package, an effcient NMF algorithm called hierarchical alternating least-squares NMF was implemented and used within the hNMF framework. hNMF can be used as a general matrix factorization technique, but in the context of this talk it will be shown that valid tissue signatures are obtained using hNMF. Tissue abundances can be mapped back to the imaging domain, providing tissue differentiation on a voxel-wise basis (see Figure 1).
