The study evaluated the performance of neuroimaging-based machine learning models in non-invasively predicting H3K27M mutations in diffuse midline gliomas (DMG). Four PRISMA-DTA databases were searched until May 2025 and included 16 studies with 2357 patients in the internal and 1792 in the external validation cohorts. MRI-based models showed a pooled sensitivity of 0.86 (95% CI: 0.79–0.91), specificity of 0.82 (95% CI: 0.75–0.87) and AUC of 0.91 (95% CI: 0.88–0.93). PET/CT-based models had lower performance with a sensitivity of 0.58 (95% CI: 0.44–0.71), specificity of 0.65 (95% CI: 0.46–0.81) and AUC of 0.61 (95% CI: 0.57–0.66). MRI models had significantly higher sensitivity (Z = 3.71; P < 0.01) and AUC (Z = 11.42; P < 0.01) than PET/CT. Deep learning models on MRI outperformed conventional algorithms (P = 0.01), and models with DNA sequencing as the reference standard had higher specificity than those with immunohistochemistry (P < 0.001). MRI-based ML shows high accuracy for H3K27M prediction in DMG.