Artificial intelligence-driven advances in automatic segmentation of meningioma brain tumors: A systematic review

The accurate segmentation of meningiomas, the most common intracranial tumors in adults, in medical imaging data is an essential component of clinical workflows for diagnosis, treatment planning, and longitudinal monitoring. Manual segmentation is labor-intensive, subjective, and challenging for small, irregular, and atypical lesions. In recent years, deep learning has emerged as a transformative artificial intelligence (AI) tool that offers automated solutions for enhancing the efficiency, consistency, and scalability across diverse imaging settings. This review synthesizes findings from 34 peer-reviewed studies published between January 1, 2020, and October 31, 2025, identified using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, with a specific focus on AI-based methods for meningioma segmentation in magnetic resonance imaging (MRI) scans. We evaluate recent advances, including U-shaped convolutional neural network (U-Net) variants, attention-enhanced frameworks, and hybrid models. Additionally, we analyze the impact of dataset characteristics, imaging modalities, and pre-processing choices on performance. The findings indicate that architectural innovation, rather than reliance on imaging protocols or pre-processing, is the primary driver of performance gains, with top models achieving Dice scores of up to 0.980 on large datasets. While numerous high-performing models rely on large public repositories, such as Figshare and brain tumor segmentation (BraTS) challenge, studies still employ custom datasets for targeted clinical use. Contrast-enhanced T1-weighted imaging is the most commonly used and effective imaging modality for meningioma segmentation. Nonetheless, challenges remain, including the segmentation of small tumors, generalizability across clinical sites, and real-time deployment of computationally demanding models. These insights highlight the need for future research to develop optimized architectures that generalize well across multi-institutional datasets while aligning with the computational constraints of real-world clinical environments.