Non-negative matrix factorization and self-organizing map for brain tumor segmentation

Over a decade, automatic segmentation of brain tumor in Magnetic Resonance Imaging (MRI) is a challenging task for researchers. Large amount of data is produced by MRI and the task of marking the tumor slice by slice is a very tedious and time consuming process for radiologists. Accurate and reliable segmentation methods are gaining more attention. This paper shows a novel framework based on Non-Negative Matrix Factorization (NMF) and Self-Organizing Map (SOM). The NMF helps to extract features from homogeneous regions. A feature vector is modeled using NMF and clustered using SOM. For segmentation SOM is trained using competitive unsupervised learning. The performance of a given framework is evaluated on BRATS 2012 dataset. It consists of MR scans from both high grade glioma (HGG) and low grade glioma (LGG) tumor patients. It uses FLAIR as well as T2 slice to segment tumor core along with whole tumor. The results showed that average dice coefficients are 0.79 for whole tumor and 0.72 for tumor core. Performance measure such as dice coefficient shows that the proposed framework gives better results as compare to other state of the art brain tumor segmentation methods.