Artificial Intelligence in Neuro-Oncology: A Review of Radiomics and Machine Learning for Predicting Glioma Molecular Profiles

Modern glioma diagnosis depends on molecular alterations that cannot be determined reliably from conventional imaging alone. Radiomics and machine learning attempt to bridge this gap by transforming magnetic resonance images into quantitative signatures associated with isocitrate dehydrogenase mutation, 1p/19q codeletion, O6-methylguanine-DNA methyltransferase promoter methylation, ATRX loss, and other clinically relevant profiles. This narrative review summarizes the principal workflows, reported performance, methodological limitations, and translational priorities of artificial intelligence in glioma radiogenomics. Evidence is strongest for noninvasive prediction of isocitrate dehydrogenase status, while results for 1p/19q and O6-methylguanine-DNA methyltransferase are promising but less stable across institutions. Handcrafted radiomics offers interpretability and modest data requirements, whereas deep learning can learn hierarchical spatial patterns but is more vulnerable to hidden confounding, scanner variation, and overfitting. Clinical adoption will require harmonized imaging, reproducible segmentation, transparent reporting, prospective multicenter validation, calibration analysis, and workflows that preserve tissue-based diagnosis as the reference standard. Artificial intelligence is therefore best viewed as a preoperative decision support and sampling tool rather than a replacement for neuropathology.