In this project, we want to assess the capacity of MR spectroscopy (MRS) to provide prognostic information in brain gliomas in general, and in glioblastomas in particular. Our goal is to use machine learning methods to correlate the metabolic information provided by MRS with molecular biomarkers of favourable prognosis in glioma subtypes. We will retrospectively analyse a retrospective MRS dataset of more than 800 gliomas, explored at the coordinated partner during more than 20 years. To this end, we will apply supervised and unsupervised machine learning methods, to classify and stratify groups of patients according to two biomarkers, 1p/19q codeletion and MGMT promoter methylation. We will validate the classifiers developed with the spectra acquired during the project from prospective patients. Additionally, we will test if multivoxel (MV) MRS can help in the prognostic evaluation of glioblastomas, in particular the differentiation of true progression vs. pseudoprogression. This will be done applying a blind source separation methods (cNMF), and a tensor-based approach for the multiparametric MRS-MRI data, to generate spectral pattern maps (nosologic images) that will be superimposed to the segmented MR images, that will be correlated with the progression status. To validate our predictor, we will prospectively acquire MV MRS in the first MR follow-up study and correlate the findings with the MV acquired in the next MR follow-up, and the information on progression-free survival and overall survival.

espectroscopía de resonancia magnética
tumores cerebrales
glioma
glioblastoma
aprendizaje de máquina
reconocimiento de patrones