Systematic review of MRI-based machine learning for brain tumour analysis

Background

Machine learning is increasingly used to analyse brain tumours on magnetic resonance imaging (MRI)¹. Segmentation is typically assessed with the Dice similarity coefficient (DSC), which quantifies overlap¹. Classification is judged by accuracy and AUC (area under the curve)¹.

Methods

Following PRISMA 2020, the authors searched PubMed, Scopus, IEEE Xplore and Web of Science¹. They included 213 studies published from January 2018 up to a final search on 16 March 2026¹. Segmentation results were pooled using random-effects meta-analysis, and classification results were summarised descriptively¹.

Findings

• Most studies addressed classification, detection or grading, with segmentation next¹.
• U-Net variants were the commonest architectures, while transformer-based models showed the quickest growth¹.
• Pooled mean DSC was 0.871 (95% CI: 0.852-0.889), and the 95% prediction interval was 0.788-0.953¹.
• Heterogeneity was extreme (I² = 99.9%). The authors therefore regard the pooled figure as descriptive of the published literature rather than one underlying effect¹.
• Averaged over the classification studies, AUC was 0.960¹.

Methodological gaps

Just 18.8% of studies included explainability, and 18.8% undertook external validation. No study reported robustness metrics¹.

Authors' conclusions

The authors argue that benchmark accuracy has run ahead of clinical evidence¹. They consider translation into patient care unlikely until three conditions are met¹:

• models are validated in independent cohorts
• robustness is tested under realistic shifts in image acquisition
• outputs are made interpretable for clinicians