Applications of artificial intelligence in magnetic resonance imaging for glioma diagnosis: Advances, techniques, and limitations

Abstract

Introduction: Magnetic resonance imaging (MRI) is one of the main techniques used in the diagnosis of gliomas, but it has limitations in tumor differentiation and in predicting molecular markers. Artificial intelligence (AI) has emerged as a complementary tool, capable of increasing diagnostic accuracy and supporting clinical decisions. Method: A systematic review of articles published between 2015 and 2025 in the PubMed, MEDLINE, and SciELO databases was conducted. Methodological quality was assessed using the AMSTAR-2 tool. Results: DL and ML-based models showed promising performance, with accuracy exceeding 95% in some cases, especially convolutional neural networks (CNNs). Hybrid models integrating radiomic and clinical-molecular data showed better sensitivity and specificity in differentiating between low- and high-grade gliomas. However, limitations such as methodological heterogeneity, lack of standardization of imaging protocols, risk of overfitting, and lack of robust external validation still restrict large-scale clinical application. Discussion: AI has shown promise in automating complex image analyses, reducing subjective biases, and offering greater diagnostic accuracy. However, challenges persist regarding the standardization of protocols, the difficulty of compatibility between systems, and the transparency of algorithms, which are factors that hinder its clinical incorporation. Conclusion: The integration of AI in MRI represents a milestone in oncological neuroimaging, with great revolutionary potential in the diagnosis of gliomas. To safely include these techniques in clinical practice, multicenter studies, interpretable models, and policies that ensure ethical validation, reproducibility, and equitable accessibility are necessary.