A paradigm shift toward explainable and federated intelligence for brain tumor MRI analysis

Deep learning techniques in brain tumor analysis using magnetic resonance imaging (MRI) scans have improved significantly, but their clinical deployment is limited by models with limited interpretability ('black-box' models) and data privacy limitations. This paper analyses recent literature and aims to explore the intersection of Explainable Artificial Intelligence (XAI) and Federated Learning (FL). Following the PRISMA protocol, a systematic search was conducted in Scopus on February 24, 2026, covering English-language journal articles and conference proceedings published from January 2020 to February 2026. The search initially identified 767 records, which were screened using predefined inclusion and exclusion criteria for MRI-based brain tumor studies addressing deep learning, explainability, privacy, or Federated Learning, resulting in 147 initially retained studies for the final screening set. One study was subsequently identified as formally retracted and excluded from the final synthesis, resulting in 146 included studies. The review classifies the literature into five primary research themes in brain tumor analysis: explainability, deep learning frameworks, privacy-preserving learning, advanced architectures (Transformers and Mamba), and multimodal learning. The literature suggests a paradigm shift in brain tumor analysis toward model interpretability using XAI. However, there is a significant knowledge gap in privacy-preserving collaborative learning models. Advanced architectures are proposed to solve brain tumor problems, but there are still limitations in model generalization, computational costs, and clinical reliability.