Unsupervised and semi-supervised Non-negative Matrix Factorization methods for brain tumor segmentation using multi-parametric MRI data

Gliomas represent about 80% of all malignant primary brain tumors. Despite recent advancements in glioma research, patient outcome remains poor. The 5 year survival rate of the most common and most malignant subtype, i.e. glioblastoma, is about 5%. Magnetic resonance imaging (MRI) has become the imaging modality of choice in the management of brain tumor patients. Conventional MRI (cMRI) provides excellent soft tissue contrast without exposing the patient to potentially harmful ionizing radiation. Over the past decade, advanced MRI modalities, such as perfusion-weighted imaging (PWI diffusion-weighted imaging (DWI) and magnetic resonance spectroscopic imaging (MRSI) have gained interest in the clinical field, and their added value regarding brain tumor diagnosis, treatment planning and follow-up has been recognized. Tumor segmentation involves the imaging-based delineation of a tumor and its subcompartments. In gliomas, segmentation plays an important role in treatment planning as well as during follow-up. Manual segmentation by a clinical expert is currently the gold standard, but it is a tedious and time-consuming task. Gliomas are known to be heterogeneous: several stages of the disease can occur throughout the same lesion and diffuse boundaries exist between active tumor, necrosis, edema and the surrounding healthy brain. Clinical practice would benefit from accurate and automated volumetric delineation of the tumor and its subcompartments. This thesis aims at developing methods for automated segmentation and characterization of brain tumors. The proposed methods are based on an unsupervised learning technique called Non-negative matrix factorization (NMF). NMF provides an additive parts-based representation of the input data, revealing the basic components which are present. Applied to the multi-parametric MR imaging data of a brain tumor patient, NMF is able to extract tissue-specific signatures as well as the relative proportions of the different tissue types in each voxel. Being an unsupervised method, NMF cannot benefit from an extensive training dataset to learn decision boundaries between tissue classes, but it is directly applicable to any multi-parametric MRI (MP-MRI) dataset of any individual patient. Two MP-MRI datasets were available for conducting the studies throughout this thesis. They were acquired at the university hospitals of Leuven (UZ Leuven) and Gent (UZ Gent), both including cMRI, PWI, DWI and MRSI, but with a different scanning protocol. Whereas only cMRI is usually considered for brain tumor segmentation, in this thesis all the available MP-MRI features are combined, and the added value of the individual MRI modalities will be explored. A hierarchical variant of NMF called hierarchical NMF (hNMF) is presented, which essentially consists of 2 levels of NMF. hNMF is able to provide valid tissue di?erentiation, and segmentation performance in terms of the Dice score is competitive. When considering reduced MP-MRI datasets by omitting one MRI modality at a time, statistically significantly higher Dice scores were found for the high-grade glioma patients when using the full MP-MRI dataset. Computational e?ciency of the hNMF algorithm is improved by considering advanced initialization methods. In particular, the successive projection algorithm (SPA) is proposed. Whereas SPA is commonly considered as a direct NMF source extraction tool, it was found in our work that better segmentation performance is achieved by using SPA as an initialization method. hNMF is also compared to other common unsupervised classifiers, including several single-level NMF and clustering methods. hNMF was found to outperform the other methods on the MP-MRI datasets of both hospitals. Thanks to its hierarchical structure, hNMF is better able to differentiate closely related tissue signatures, such as non-enhancing tumor and edema. To be competitive with state-of-the-art, unsupervised segmentation algorithms require the incorporation of prior knowledge. A semi-supervised NMF-based segmentation framework is presented, incorporating user-defined voxel selection to initialize the NMF pathological tissue sources. L1-regularization is included in the NMF objective function to promote spatial consistency and sparseness of the tissue abundance maps. A morphological post-processing procedure exploits the location of the selected voxels along with tissue adjacency constraints to further enhance the segmentation results. Careful voxel selection in the tumor compartments was found necessary to obtain robust segmentation results. The semi-automated NMF-based segmentation framework was applied to the publicly available Leaderboard dataset of the BRATS 2013 challenge, allowing direct comparison with state-of-the-art. Competitive segmentation results are obtained, especially for the active tumor region for which our method ranks first among all participants. This thesis demonstrates that unsupervised segmentation algorithms can compete with state-of-the-art supervised segmentation methods by incorporating adequate prior knowledge into the algorithm.