Abstract
Machine learning is now widely applied to magnetic resonance imaging-based brain tumor analysis, with task-specific performance metrics: the Dice similarity coefficient (DSC) quantifies segmentation overlap, while the area under the curve (AUC) and accuracy measure classification performance. A PRISMA 2020 systematic review was conducted, searching Scopus, Web of Science, PubMed, and IEEE Xplore, and identified 213 eligible studies published between January 2018 and the final search date of March 16, 2026. Segmentation performance was synthesized by random-effects meta-analysis and classification performance was summarized descriptively. Classification, detection, and grading tasks dominated the literature, followed by segmentation; U-Net variants were the most frequent architectures, with transformer-based models growing fastest. The pooled mean DSC was 0.871 (95% CI: 0.852–0.889) with a 95% prediction interval of 0.7880.953, and heterogeneity was extreme ( I 2 = 99.9%), so the pooled mean should be read as a descriptive summary of the literature rather than a single underlying effect. The descriptive mean AUC across classification studies was 0.960. Methodological gaps were pervasive: only 18.8% of studies incorporated explainability, only 18.8% performed external validation, and none reported robustness metrics. Benchmark accuracy has outpaced clinical evidence: until models are validated on independent cohorts, evaluated for robustness under realistic acquisition shifts, and made interpretable to clinicians, reported performance is unlikely to translate into patient care.