Towards Trustworthy AI in Neuro-Oncology: A Systematic Review of Explainable AI Models for Brain Tumor Detection
Year: 2025
Type: review
Study design: Systematic Review
Abstract: Artificial intelligence has shown significant potential in brain tumor detection using MRI. Despite high accuracy, most AI models function as “black boxes,” limiting their trust and usability in clinical settings. Explainable AI (XAI) has emerged to address this by improving transparency and interpretability of AI decisions. This systematic review investigates the landscape of XAI models developed for brain tumor detection and provides a structured evaluation of their clinical trustworthiness. The key contribution of this review is a synthesis of current methods, datasets, evaluation strategies, and practical implications, with a focus on how trust is assessed and established. Following PRISMA guidelines, 21 peer-reviewed studies published between 2020 and 2024 were selected. Each study was analyzed in terms of its AI model, dataset characteristics, XAI technique (SHAP, LIME, Grad-CAM), trustworthiness evaluation method, and clinical relevance. The findings reveal a dominant reliance on retrospective datasets (BraTS, Figshare, and institutional MRI repositories) with significant variability in data volume and annotation quality. XAI methods were primarily used for pixel-level saliency, modality attribution, or case-level interpretation, often without standardized validation protocols. Only a few studies involved clinical experts in evaluation, and fewer still reported outcomes like clinician trust or decision confidence. Furthermore, visual interpretability was often qualitative and lacked measurable alignment with expert-defined tumor regions. Uncertainty quantification and cross-institutional generalizability were also rarely addressed. In conclusion, while XAI shows promise in enhancing the trustworthiness of AI-based neuro-oncological tools, future research must prioritize human-centered validation, prospective clinical testing, and standardized evaluation frameworks to enable meaningful clinical translation.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7160953415