Abstract
Abstract Background The Glioma binary classification has become a significant field of research because of its clinical application in the diagnosis and the treatment planning. The recent advances in machine learning (ML) and deep learning (DL) allowed analyzing medical images automatically, specifically magnetic resonance imaging (MRI). The study is a state-of-the-art overview of ML and DL based Glioma binary classification algorithms published in 2020–2025, where primary focus is put on MRI-based methods, but CT and mixed MRI-CT are also considered. The research will examine the current approaches, outline critical issues, and outline gaps in research that restrict clinical translation. Methodology A comprehensive search across IEEE Xplore, PubMed, Web of Science, and Scopus yielded over 115 candidate articles, of which 82 studies fulfilled the predefined inclusion criteria following a structured screening process. The identified papers are systematically reviewed according to the imaging modality, preprocessing and data augmentation methods, learning architecture and evaluation metrics. Moreover, a detection-based framework based on YOLOv9 and YOLOv8 are also implemented as well as comparatively analyzed. Accuracy, area under the curve (AUC), sensitivity, specificity, and computational complexity are used to measure performance of these models. Result and conclusion According to this study, MRI is the best modality of Glioma detection as well as classification as it provides better soft-tissue contrast. Among the examined methods, the approaches based on classification as Deep transfer learnings are more stable and yield high diagnostic accuracy, whereas the detection methods like YOLOv9 have the potential of locating and classifying the tumors simultaneously with the needed adaptations that need to be made in the volumetric data. In the experimental component of this study, YOLOv9 consistently outperformed YOLOv8 on the held-out validation set, achieving higher precision (0.815 vs. 0.741), recall (0.766 vs. 0.678), mAP@50 (0.812 vs. 0.712) and mAP@50–95 (0.44 vs. 0.34), while both architectures maintained comparable, sub-11 ms per-image inference latency (~ 92 frames/second on the test set), confirming YOLOv9 as the more accurate detector without a meaningful speed penalty. The continuing issues are the heterogeneity of data, insufficient external validation, non-uniform evaluation procedures, and insufficient explainability. Our semantic and bibliometric mapping of the literature further reveals emerging research trajectories including transformer and attention-based multimodal fusion, explainable and uncertainty-aware AI, radio genomic (imaging–genomics) integration, and lightweight, real-time detection frameworks such as the YOLO-based models evaluated here that are expected to shape the next generation of clinically deployable Glioma classification systems. Future studies should prioritize standardized multi-center benchmarks, multimodal validation, and interpretable, regulatory-grade AI models to facilitate safe clinical adoption.