Purpose:
Accurate prediction of overall survival (OS) in glioma patients is crucial for optimising treatment decisions. Despite advancements in imaging and machine learning, challenges persist due to tumour heterogeneity and confounding factors. This scoping review systematically assesses state-of-the-art image-based OS prediction models for glioma, focusing on tumour characteristics, imaging modalities, preprocessing techniques, and machine learning methods.

Material and Methods:
This scoping review was conducted following the Joanna Briggs Institute guidelines, comprising five key stages: identifying the research question, searching for relevant literature, selecting studies, charting the data, and collating, summarising, and reporting the results.

Results:
The initial search identified 3238 records, of which 70 articles were included in the final analysis. Most studies originated from China, the United States, and India, with datasets averaging approximately 450 cases. To enhance predictive accuracy, various techniques were utilised, including image segmentation, multimodal magnetic resonance imaging (MRI) protocols, and advanced feature extraction methods. Notably, T1-weighted contrast-enhanced MRI and grade-specific glioma analyses improved model performance. Although deep learning models generally outperformed traditional methods, they required large, balanced datasets. Hybrid models showed promising potential; however, their performance was inconsistent due to challenges such as limited image quality and issues with model interpretability.

Conclusions:
Increasing sample size alone does not guarantee improved accuracy in glioma prediction models, because data quality and feature selection are critical factors. Incorporating diverse imaging modalities can significantly enhance predictive performance. To ensure greater clinical reliability in decision-making, integrating clinical features with imaging data is essential.