Radiomics and Machine Learning in Diagnostics of Glial Brain Tumors: a Systematic Review and Meta-Analysis

Glial tumors are the most common neuroepithelial neoplasms of the brain. Consequently, investigating robust, non-invasive techniques for subtyping these tumors - specifically through advanced multimodal neuroimaging and radiomics - is warranted. The present systematic review of scientific literature, including meta-analysis, was conducted to specify the major challenges of radiomics and machine learning in diagnostics of glial tumors based on the MRI data as well as to assess the quality of such non-invasive diagnostics. We analyzed 42 publications utilizing radiomics and machine learning to predict molecular biomarker status in glial tumors based on MRI data. The analysis covered mutations in the IDH, ATRX, BRAF, and H3K27M genes, as well as TERT promoter mutations, 1p/19q codeletion, MGMT promoter methylation, and proliferative activity (Ki-67 labeling index). The overall accuracy of these techniques was high and equaled 0.86 [0.83; 0.89]. At the same time, the studies demonstrated significant methodological heterogeneity, in particular, related to the lack of uniform standards to select the location, size, and shape of the area of interest for obtaining radiomic features. This greatly hinders reproduction of the experimental results in clinical practice. Therefore, standardization of radiomics procedures remains relevant for further research of glial tumors.