Final Report Summary - BAYESIANMULTIMODALMR (Multimodal analysis of MR data using Bayesian methods) Nuclear magnetic resonance (MR) is widely used for non-invasive investigations of brain tumours, in particular in vivo diagnosis and grading, surgical planning and assessment of response to therapy. It is generally applied as MR imaging (MRI), however there are important examples (i.e. malignant gliomas) where current imaging techniques still lack sufficient diagnostic accuracy and in which MR spectroscopy (MRS) plays an important role. MR spectroscopic imaging (MRSI) has been successfully applied to monitoring the metabolic heterogeneity of human brain tumours.The rich information contained in MR signals makes them ideally suited to the application of pattern recognition techniques. In the current project, this problem is approached by using Non-negative Matrix Factorisation (NMF), a group of unsupervised techniques in which a data matrix X is approximately factorised into (usually) two matrices W (known as the source signals) and H (the mixing matrix). Different variants of NMF have previously been applied in the context of neuro-oncology to provide answers to a number of research questions. With this project, we aimed to address the following challenges in an interdisciplinary and multidisciplinary approach:- To use prior knowledge about the sample to be analysed, in order to more accurately identify the sources- Estimate the number of underlying tissue types present in a sample, especially when there is high heterogeneity- Track the response to a particular choice of therapyThe results obtained during this project can be divided into two categories: research and training. The results of the research are summarised below:a) We developed a method that uses the areas that arise from segmenting the MRI (i.e. T2W images) to guide the extraction of relevant source from the MRSI, hence using the anatomical information of the tumour provided by the MRI. Importantly, this new approach does not involve combining spectra from different subjects, thus focusing on intra-subject variation without contamination from inter-subject overlaps. It is also worth mentioning that the semi-supervised nature of the methodology allows to use only partial regions, so that areas of uncertainty can be left outside the initial segmentation. We tested the proposed methodology using MR data from 15 brain-tumour bearing mice and the results obtained are very encouraging, suggesting that this new approach could be successfully used for embedding MRI information into the source extraction in MRSI data.b) We produced a strategy to automatically determine the number of relevant sources when using one of the robust versions of Bayesian NMF tested. To determine the relevant sources, we took advantage of the ability of this Bayesian NMF method to favour sparse representations by controlling the hyperparameters of its priors. Then, by using a greedy strategy during the iterative process of matrices decomposition of NMF, we propose: i) to discard sources where the corresponding columns in the mixing matrix are zero (or a very small value) in all of their entries (which indicates that these sources are irrelevant or meaningless); and ii) where two sources are highly correlated (>0.98) between them (which suggests that both of them are representing the same kind of information), to discard one of these sources. We used simulated and real-world data to test and validate this study. The real data were extracted from an international multi-centre database compiled by the INTERPRET European project, and consisted of single-voxel MRS data from 188 humans, including astrocytoma grade II (22), glioblastomas (86), meningiomas (58) and normal brain controls (22). The simulated data used in this study was modelled from samples extracted from the INTERPRET database.c) We developed a method that is able to extract meaningful source signals that represent healthy tissue, brain tumour in response to therapy, and untreated brain tumour or without response. These sources are used later on for the generation of nosological images of the response to therapy. For the extraction of the sources we used a total of 508 MR spectral vectors from 14 mice, 8 of them treated and showing transient response to temozolomide, and 6 of them from the control group. We validated the obtained results with an independent test set (7 control and 17 treated mice) and against histopathology. For some mice several MRSI were acquired at different time points. The nosologic images obtained allowed convenient tracking of response to treatment and differentiated the intratumoural heterogeneity of such response, hinting the growth arrest and relapse, before changes in tumour volume were observed.Most of these results were disseminated in leading conferences of the relevant areas (methodology and application) and well-known international journals. This includes 3 journal papers (in PLoS ONE, NMR in Biomedicine, and BMC Bioinformatics), 3 conference papers (in IEEE SSCI and ESSAN), 6 conference abstracts (in ISMRM, ESMRMB and the Spanish Biophysical Society), and 1 book chapter (eMagRes); for a grand total of 13 peer-reviewed publications/communications.Regarding the training, the researcher of this training-through-research project has been involved in many activities with the primary purpose of strengthening her research career via the acquisition of a series of new skills, knowledge and perspectives, under the guidance and supervision of the host group. They include the dissemination of the results, further training for the preparation of a new project proposal, gaining teaching experience, mentoring a student, organising special sessions at renowned conferences, engaging in public talks, expanding her network of collaborators.
