Tensor Based Approaches in Magnetic Resonance Spectroscopic Imaging and Multi-parametric MRI Data Analysis
Accurate characterisation and localization of pathologic tissue types play a key role in diagnosis and treatment planning of brain tumors. Neuroimaging techniques such as magnetic resonance imaging (MRI), magnetic resonance spectroscopic imaging (MRSI), perfusion-weighted imaging (PWI) and diffusion weighted imaging (DWI) are being used to characterize brain tumors and detect full tumor extent. Analysing these data is both time consuming and challenging for clinicians. Automated algorithms will aid clinicians to analyse the data faster and more accurately. Blind source separation is one such technique that is commonly used to extract useful information from the data. Most of these algorithms use matrix based approaches. Working with tensor tools such as tensor decompositions can be of great benefit compared to their
matrix counterpart. Tensors are applied in domains such as signal processing, biomedical engineering, statistics and machine learning. In this thesis, we aim to develop tensor based blind source separation algorithms for analysing the MRSI and multi-parametric MRI (MP-MRI) signals.
First, tensor based blind source separation methods are developed to remove artefacts. In this thesis, we focus on residual water suppression in the MRSI signal. To suppress the residual water, a Löwner/Hankel tensor is constructed from the MRSI signal. Canonical polyadiac decomposition (CPD)/Multilinear singular value decomposition is applied on the tensor to extract the water component, and to subsequently remove it from the original MRSI signal. The tensor based water suppression methods show significant improvement in performance for both simulated and in-vivo MRSI signals compared to the matrix-based approaches.
Second, tensor based blind source seperation is applied to differentiate various tissue types in glioma patients from MRSI/multi-parametric MRI signals. Such a tensor based tumor tissue type differentiation approach is developed which consists of building a xxT structured 3-D tensor from the MRSI spectra and then applying a non-negative CPD to extract tissue specific spectra and its corresponding distribution in the MRSI grid. An in-vivo study shows that our tensor based approach significantly outperforms the matrix-based approaches in identifying tumor and necrotic tissue type in glioma patients. This tensor based tissue characterization approach is further extended to multiparametric magnetic resonance imaging (MP-MRI) including conventional magnetic resonance imaging, perfusion-weighted imaging, diffusion-weighted
imaging and MRSI modalities to perform tumor segmentation. Third, we explore the applicability of tensor decompositions in supervised algorithms for voxel classification in MRSI and tumour tissue segmentation
in MP-MRI. A CNN based low-rank regularized classifier is developed to classify voxels in MRSI. Multilinear singular value decomposition (MLSVD) is used to apply regularization in the convolution layer. Low-rank regularization provides slight improvement in computational complexity without degrading the classification performance. For tumour tissue segmentation, a superpixel-wise two stage random forest algorithm is developed. The whole tumor is segmented in the first stage and in the second stage sub-compartments are segmented from the whole tumor. Multilinear singular value decomposition (MLSVD) is used
to extract some of the features as input to the random forest classifier. The proposed algorithm was analysed on the BRATS 2017 challenge dataset, which showed a very good performance in segmenting the whole tumor and average performance in segmenting sub-compartments. This shows that tensor based feature extraction is a viable option for tumor tissue segmentation in MP-MRI.