Machine Learning in Magnetic Resonance Images of Glioblastoma: A Review

Background: The purpose of this work was to identify which Glioblastoma (GBM) problems can be handled by Magnetic Resonance Imaging (MRI) and Machine Learning (ML) techniques. Results, limitations, and trends through a review of the scientific literature in the last 5 years were performed. Google Scholar, PubMed, Elsevier databases, and forward and backward citations were used for searching articles applying ML techniques in GBM. The 50 most relevant papers fulfilling the selection criteria were deeply analyzed. The PRISMA statement was followed to structure our report. Methods: A partial taxonomy of the GBM problems tackled with ML methods was formulated with 15 subcategories grouped into four categories: extraction of characteristics from tumoral regions, differentiation, characterization, and problems based on genetics. Results: The dominant techniques in solving these problems are: Radiomics for feature extraction, Least Absolute Shrinkage and Selection Operator for feature selection, Support Vector Machines and Random Forest for classification, and Convolutional Neural Networks for characterization. A noticeable trend is that the application of Deep Learning on GBM problems is growing exponentially. The main limitations of ML methods are their interpretability and generalization. Conclusion: The diagnosis, treatment, and characterization of GBM have advanced with the aid of ML methods and MRI data, and this improvement is expected to continue. ML methods are effective in solving GBM-related problems with different precisions, Overall Survival being the hardest problem to solve with accuracies ranging from 57%-71%, and GBM differentiation the one with the highest accuracy ranging from 80%-97%.